maxmustermann/DeepLearningExamples
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Merge pull request #561 from NVIDIA/nvpstr/fed7ba

Browse source 
Nvpstr/fed7ba
This commit is contained in:
nv-kkudrynski 2020-06-12 14:48:24 +02:00 committed by GitHub
parent b709b3ae57 f11884b38a
commit 6c1d562eb9
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
152 changed files with 3314 additions and 4165 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
PyTorch/SpeechSynthesis/FastPitch/fastpitch/transformer.py
View file

@ -29,7 +29,7 @@ class PositionalEmbedding(nn.Module):
def forward(self, pos_seq, bsz=None):
sinusoid_inp = torch.ger(pos_seq, self.inv_freq)
pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=-1)
pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=1)
if bsz is not None:
return pos_emb[None, :, :].expand(bsz, -1, -1)
else:

2
PyTorch/SpeechSynthesis/FastPitch/fastpitch/transformer_jit.py
View file

@ -36,7 +36,7 @@ class PositionalEmbedding(nn.Module):
def forward(self, pos_seq, bsz: Optional[int] = None):
sinusoid_inp = torch.ger(pos_seq, self.inv_freq)
pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=-1)
pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=1)
if bsz is not None:
return pos_emb[None, :, :].expand(bsz, -1, -1)
else:

3
PyTorch/SpeechSynthesis/FastPitch/inference.py
View file

@ -296,6 +296,9 @@ def main():
'pitch_tgt': None,
'pitch_transform': build_pitch_transformation(args)}
if args.torchscript:
gen_kw.pop('pitch_transform')
all_utterances = 0
all_samples = 0
all_letters = 0

3
PyTorch/SpeechSynthesis/FastPitch/models.py
View file

@ -101,8 +101,7 @@ def get_model(model_name, model_config, device,
class FastPitch__forward_is_infer(_FastPitchJIT):
def forward(self, inputs, input_lengths, pace: float = 1.0,
dur_tgt: Optional[torch.Tensor] = None,
pitch_tgt: Optional[torch.Tensor] = None,
pitch_transform: Optional[bool] = None):
pitch_tgt: Optional[torch.Tensor] = None):
return self.infer(inputs, input_lengths, pace=pace,
dur_tgt=dur_tgt, pitch_tgt=pitch_tgt)
else:

1
PyTorch/SpeechSynthesis/FastPitch/scripts/inference_benchmark.sh
View file

@ -21,6 +21,7 @@ for PRECISION in fp16 fp32; do
--include-warmup \
--batch-size ${BSZ} \
--repeats 1000 \
--torchscript \
-i phrases/benchmark_8_128.tsv
done
done

12
README.md
View file

@ -15,9 +15,9 @@ These examples, along with our NVIDIA deep learning software stack, are provided
The examples are organized first by framework, such as TensorFlow, PyTorch, etc. and second by use case, such as computer vision, natural language processing, etc. We hope this structure enables you to quickly locate the example networks that best suit your needs. Here are the currently supported models:
### Computer Vision
- __ResNet-50__ [[MXNet](https://github.com/NVIDIA/DeepLearningExamples/tree/master/MxNet/Classification/RN50v1.5)] [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/RN50v1.5)]
- __ResNext__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)]
- __SE-ResNext__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)]
- __ResNet-50__ [[MXNet](https://github.com/NVIDIA/DeepLearningExamples/tree/master/MxNet/Classification/RN50v1.5)] [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/ConvNets)]
- __ResNext__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/ConvNets)]
- __SE-ResNext__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/ConvNets)]
- __SSD__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Detection/SSD)]
- __Mask R-CNN__ [[PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Segmentation/MaskRCNN)] [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow2/Segmentation/MaskRCNN)] [[TensorFlow 2](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow2/Segmentation/MaskRCNN)]
- __U-Net(industrial)__ [[TensorFlow](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Segmentation/UNet_Industrial)]
@ -65,7 +65,7 @@ The examples are organized first by framework, such as TensorFlow, PyTorch, etc.
## Feature Matrix
| Models | Framework | DALI | AMP | Multi-GPU | Multi-Node | TensorRT | ONNX | Triton | TF-TRT |
| ------------- | ------------- | ------------- | ------------- | ------------- | ------------- |------------- |------------- |------------- |------------- |
| [ResNet50 v1.5](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets/resnet50v1.5) |PyTorch | Yes | Yes | Yes | - | - | - | - | - |
| [ResNet-50 v1.5](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets/resnet50v1.5) |PyTorch | Yes | Yes | Yes | - | - | - | - | - |
| [ResNeXt101-32x4d](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets/resnext101-32x4d) |PyTorch | Yes | Yes | Yes | - | - | - | - | - |
| [SE-ResNeXt101-32x4d](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets/se-resnext101-32x4d) |PyTorch | Yes | Yes | Yes | - | - | - | - | - |
| [SSD300 v1.1](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD) |PyTorch | Yes | Yes | Yes | - | - | - | - | - |
@ -79,7 +79,9 @@ The examples are organized first by framework, such as TensorFlow, PyTorch, etc.
| [FastPitch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/SpeechSynthesis/FastPitch) | PyTorch | N/A | Yes | Yes | - | - | - | - | - |
| [GNMT v2](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Translation/GNMT) |PyTorch | N/A | Yes | Yes | - | - | - | - | - |
| [Transformer](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Translation/Transformer) |PyTorch | N/A | Yes | Yes | - | - | - | - | - |
| [ResNet-50 v1.5](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/RN50v1.5) |TensorFlow | Yes | Yes | Yes | - | - | - | - | - |
| [ResNet-50 v1.5](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/ConvNets/resnet50v1.5) |TensorFlow | Yes | Yes | Yes | - | - | - | - | - |
| [ResNeXt101-32x4d](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/ConvNets/resnext101-32x4d) |TensorFlow | Yes | Yes | Yes | - | - | - | - | - |
| [SE-ResNeXt101-32x4d](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TEnsorFlow/Classification/ConvNets/se-resnext101-32x4d) |TensorFlow | Yes | Yes | Yes | - | - | - | - | - |
| [SSD320 v1.2](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Detection/SSD) | TensorFlow | N/A | Yes | Yes | - | - | - | - | - |
| [BERT](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT) |TensorFlow | N/A | Yes | Yes | Yes | Yes | - | [Yes](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT/triton) | Yes |
| [BioBert](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT/biobert) | TensorFlow | N/A | Yes | Yes | - | - | - | - | - |

0
TensorFlow/Classification/RN50v1.5/.gitignore → TensorFlow/Classification/ConvNets/.gitignore vendored
View file

0
TensorFlow/Classification/RN50v1.5/dllogger/__init__.py → TensorFlow/Classification/ConvNets/.gitmodules vendored
View file

0
TensorFlow/Classification/RN50v1.5/.style.yapf → TensorFlow/Classification/ConvNets/.style.yapf
View file

8
TensorFlow/Classification/ConvNets/Dockerfile Normal file
View file

@ -0,0 +1,8 @@
ARG FROM_IMAGE_NAME=nvcr.io/nvidia/tensorflow:20.03-tf1-py3
FROM ${FROM_IMAGE_NAME}
ADD requirements.txt .
RUN pip install -r requirements.txt
ADD . /workspace/rn50v15_tf
WORKDIR /workspace/rn50v15_tf

0
TensorFlow/Classification/RN50v1.5/LICENSE → TensorFlow/Classification/ConvNets/LICENSE
View file

68
TensorFlow/Classification/ConvNets/README.md Normal file
View file

@ -0,0 +1,68 @@
# Resnet-family Convolutional Neural Networks for Image Classification in Tensorflow
In this repository you will find implementation of Resnet and its variations for image
classification
## Table Of Contents
* [Models](#models)
* [Validation accuracy results](#validation-accuracy-results)
* [Training performance results](#training-performance-results)
* [Training performance: NVIDIA DGX-1 (8x V100 16G)](#training-performance-nvidia-dgx-1-(8x-v100-16G))
* [Release notes](#release-notes)
* [Changelog](#changelog)
## Models
The following table provides links to where you can find additional information on each model:
| **Model** | **Link**|
|-----------|---------|
| resnet50 | [README](./resnet50v1.5/README.md) |
| resnext101-32x4d | [README](./resnext101-32x4d/README.md) |
| se-resnext101-32x4d | [README](./se-resnext101-32x4d/README.md) |
## Validation accuracy results
Our results were obtained by running the applicable training scripts in the tensorflow-20.03-tf1-py3 NGC container
on NVIDIA DGX-1 with (8x V100 16G) GPUs. The specific training script that was run is documented in the corresponding model's README.
The following table shows the validation accuracy results of the
three classification models side-by-side.
| **arch** | **AMP Top1** | **AMP Top5** | **FP32 Top1** | **FP32 Top5** |
|:-:|:-:|:-:|:-:|:-:|
| resnet50 | 78.35 | 94.21 | 78.34 | 94.21 |
| resnext101-32x4d | 80.21 | 95.00 | 80.21 | 94.99 |
| se-resnext101-32x4d | 80.87 | 95.35 | 80.84 | 95.37 |
## Training performance results
### Training performance: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the applicable
training scripts in the tensorflow-20.03-tf1-py3 NGC container
on NVIDIA DGX-1 with (8x V100 16G) GPUs.
Performance numbers (in images per second)
were averaged over an entire training epoch.
The specific training script that was run is documented
in the corresponding model's README.
The following table shows the training accuracy results of the
three classification models side-by-side.
| **arch** | **Mixed Precision** | **Mixed Prcesision XLA** | **FP32** | **Mixed Precision speedup** | **XLA Mixed Precision speedup**|
|:-:|:-:|:-:|:-:|:-:|:-:|
| resnet50 | 8277.91 img/s | 9485.21 img/s | 2785.81 img/s | 2.97x | 1.14x |
| resnext101-32x4d | 3151.81 img/s | 4231.42 img/s | 1055.82 img/s | 2.98x | 1.34x |
| se-resnext101-32x4d | 2168.40 img/s | 3297.39 img/s | 921.38 img/s | 2.35x | 1.52x |
## Release notes
### Changelog
June 2020
- ConvNets repo restructurization
- Initial release of ResNext and SE-Resnext

127
TensorFlow/Classification/ConvNets/main.py Executable file
View file

@ -0,0 +1,127 @@
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import warnings
warnings.simplefilter("ignore")
import tensorflow as tf
import horovod.tensorflow as hvd
import dllogger
from utils import hvd_utils
from runtime import Runner
from model.resnet import model_architectures
from utils.cmdline_helper import parse_cmdline
if __name__ == "__main__":
tf.logging.set_verbosity(tf.logging.ERROR)
FLAGS = parse_cmdline(model_architectures.keys())
hvd.init()
if hvd.rank() == 0:
log_path = os.path.join(FLAGS.results_dir, FLAGS.log_filename)
os.makedirs(FLAGS.results_dir, exist_ok=True)
dllogger.init(
backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE, filename=log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
]
)
else:
dllogger.init(backends=[])
dllogger.log(data=vars(FLAGS), step='PARAMETER')
runner = Runner(
# ========= Model HParams ========= #
n_classes=1001,
architecture=FLAGS.arch,
input_format='NHWC',
compute_format=FLAGS.data_format,
dtype=tf.float32 if FLAGS.precision == 'fp32' else tf.float16,
n_channels=3,
height=224,
width=224,
distort_colors=False,
log_dir=FLAGS.results_dir,
model_dir=FLAGS.model_dir if FLAGS.model_dir is not None else FLAGS.results_dir,
data_dir=FLAGS.data_dir,
data_idx_dir=FLAGS.data_idx_dir,
weight_init=FLAGS.weight_init,
use_xla=FLAGS.use_xla,
use_tf_amp=FLAGS.use_tf_amp,
use_dali=FLAGS.use_dali,
gpu_memory_fraction=FLAGS.gpu_memory_fraction,
gpu_id=FLAGS.gpu_id,
seed=FLAGS.seed
)
if FLAGS.mode in ["train", "train_and_evaluate", "training_benchmark"]:
runner.train(
iter_unit=FLAGS.iter_unit,
num_iter=FLAGS.num_iter,
run_iter=FLAGS.run_iter,
batch_size=FLAGS.batch_size,
warmup_steps=FLAGS.warmup_steps,
log_every_n_steps=FLAGS.display_every,
weight_decay=FLAGS.weight_decay,
lr_init=FLAGS.lr_init,
lr_warmup_epochs=FLAGS.lr_warmup_epochs,
momentum=FLAGS.momentum,
loss_scale=FLAGS.loss_scale,
label_smoothing=FLAGS.label_smoothing,
mixup=FLAGS.mixup,
use_static_loss_scaling=FLAGS.use_static_loss_scaling,
use_cosine_lr=FLAGS.use_cosine_lr,
is_benchmark=FLAGS.mode == 'training_benchmark',
)
if FLAGS.mode in ["train_and_evaluate", 'evaluate', 'inference_benchmark']:
if FLAGS.mode == 'inference_benchmark' and hvd_utils.is_using_hvd():
raise NotImplementedError("Only single GPU inference is implemented.")
elif not hvd_utils.is_using_hvd() or hvd.rank() == 0:
runner.evaluate(
iter_unit=FLAGS.iter_unit if FLAGS.mode != "train_and_evaluate" else "epoch",
num_iter=FLAGS.num_iter if FLAGS.mode != "train_and_evaluate" else 1,
warmup_steps=FLAGS.warmup_steps,
batch_size=FLAGS.batch_size,
log_every_n_steps=FLAGS.display_every,
is_benchmark=FLAGS.mode == 'inference_benchmark',
export_dir=FLAGS.export_dir
)
if FLAGS.mode == 'predict':
if FLAGS.to_predict is None:
raise ValueError("No data to predict on.")
if not os.path.isfile(FLAGS.to_predict):
raise ValueError("Only prediction on single images is supported!")
if hvd_utils.is_using_hvd():
raise NotImplementedError("Only single GPU inference is implemented.")
elif not hvd_utils.is_using_hvd() or hvd.rank() == 0:
runner.predict(FLAGS.to_predict)

2
TensorFlow/Classification/RN50v1.5/model/__init__.py → TensorFlow/Classification/ConvNets/model/__init__.py
View file

@ -14,4 +14,4 @@
from model import layers
from model import blocks
from model import resnet_v1_5
from model import resnet

0
TensorFlow/Classification/RN50v1.5/model/blocks/__init__.py → TensorFlow/Classification/ConvNets/model/blocks/__init__.py
View file

70
TensorFlow/Classification/RN50v1.5/model/blocks/conv2d_block.py → TensorFlow/Classification/ConvNets/model/blocks/conv2d_block.py
View file

@ -34,7 +34,8 @@ def conv2d_block(
data_format='NHWC',
conv2d_hparams=None,
batch_norm_hparams=None,
name='conv2d'
name='conv2d',
cardinality=1,
):
if not isinstance(conv2d_hparams, tf.contrib.training.HParams):
@ -44,8 +45,7 @@ def conv2d_block(
raise ValueError("The paramater `conv2d_hparams` is not of type `HParams`")
with tf.variable_scope(name):
if mode != 'SAME_RESNET':
if cardinality == 1:
net = layers.conv2d(
inputs,
n_channels=n_channels,
@ -56,58 +56,18 @@ def conv2d_block(
use_bias=not use_batch_norm,
trainable=is_training,
kernel_initializer=conv2d_hparams.kernel_initializer,
bias_initializer=conv2d_hparams.bias_initializer,
)
else: # Special padding mode for ResNet models
if strides == (1, 1):
net = layers.conv2d(
inputs,
n_channels=n_channels,
kernel_size=kernel_size,
strides=strides,
padding='SAME',
data_format=data_format,
use_bias=not use_batch_norm,
trainable=is_training,
kernel_initializer=conv2d_hparams.kernel_initializer,
bias_initializer=conv2d_hparams.bias_initializer,
)
else:
rate = 1 # Unused (for 'a trous' convolutions)
kernel_height_effective = kernel_size[0] + (kernel_size[0] - 1) * (rate - 1)
pad_h_beg = (kernel_height_effective - 1) // 2
pad_h_end = kernel_height_effective - 1 - pad_h_beg
kernel_width_effective = kernel_size[1] + (kernel_size[1] - 1) * (rate - 1)
pad_w_beg = (kernel_width_effective - 1) // 2
pad_w_end = kernel_width_effective - 1 - pad_w_beg
padding = [[0, 0], [pad_h_beg, pad_h_end], [pad_w_beg, pad_w_end], [0, 0]]
if data_format == 'NCHW':
padding = [padding[0], padding[3], padding[1], padding[2]]
padded_inputs = tf.pad(inputs, padding)
net = layers.conv2d(
padded_inputs, # inputs,
n_channels=n_channels,
kernel_size=kernel_size,
strides=strides,
padding='VALID',
data_format=data_format,
use_bias=not use_batch_norm,
trainable=is_training,
kernel_initializer=conv2d_hparams.kernel_initializer,
bias_initializer=conv2d_hparams.bias_initializer,
)
bias_initializer=conv2d_hparams.bias_initializer)
else:
group_filter = tf.get_variable(
name=name + 'group_filter',
shape=[3, 3, n_channels // cardinality, n_channels],
trainable=is_training,
dtype=tf.float32)
net = tf.nn.conv2d(inputs,
group_filter,
strides=strides,
padding='SAME',
data_format=data_format)
if use_batch_norm:
net = layers.batch_norm(
net,

37
TensorFlow/Classification/RN50v1.5/model/blocks/resnet_bottleneck_block.py → TensorFlow/Classification/ConvNets/model/blocks/resnet_bottleneck_block.py
View file

@ -30,11 +30,14 @@ def bottleneck_block(
depth,
depth_bottleneck,
stride,
cardinality=1,
training=True,
data_format='NCHW',
conv2d_hparams=None,
batch_norm_hparams=None,
block_name="bottleneck_block"
block_name="bottleneck_block",
use_se=False,
ratio=1
):
if data_format not in ['NHWC', 'NCHW']:
@ -44,7 +47,7 @@ def bottleneck_block(
raise ValueError("The paramater `conv2d_hparams` is not of type `HParams`")
if not isinstance(batch_norm_hparams, tf.contrib.training.HParams):
raise ValueError("The paramater `conv2d_hparams` is not of type `HParams`")
raise ValueError("The paramater `batch_norm_hparams` is not of type `HParams`")
in_shape = inputs.get_shape()
@ -54,23 +57,17 @@ def bottleneck_block(
with tf.variable_scope("shortcut"):
if depth == in_size:
if stride == 1:
shortcut = tf.identity(inputs)
else:
shortcut = model.layers.average_pooling2d(
inputs,
pool_size=(1, 1),
strides=(stride, stride),
padding='valid',
data_format='channels_first' if data_format == 'NCHW' else 'channels_last',
name="average_pooling2d",
)
name="average_pooling2d")
else:
shortcut = model.blocks.conv2d_block(
inputs,
n_channels=depth,
@ -85,8 +82,13 @@ def bottleneck_block(
batch_norm_hparams=batch_norm_hparams
)
#cardinality_to_bottleneck_width = { 1:64, 2:40, 4:24, 8:14, 32:4, 64:4 }
#cardinality_to_grouped_conv_width = { 1:64, 2:80, 4:96, 8:112, 32:128, 64:256 }
#per_group_ck = cardinality_to_bottleneck_width[cardinality] * depth_bottleneck / 64
bottleneck = model.blocks.conv2d_block(
inputs,
#n_channels=per_group_ck * cardinality if cardinality != 1 else depth_bottleneck,
n_channels=depth_bottleneck,
kernel_size=(1, 1),
strides=(1, 1),
@ -97,23 +99,22 @@ def bottleneck_block(
data_format=data_format,
conv2d_hparams=conv2d_hparams,
batch_norm_hparams=batch_norm_hparams,
name='bottleneck_1'
)
name='bottleneck_1')
bottleneck = model.blocks.conv2d_block(
bottleneck,
n_channels=depth_bottleneck,
kernel_size=(3, 3),
strides=(stride, stride),
mode='SAME_RESNET',
mode='SAME',
use_batch_norm=True,
activation='relu',
is_training=training,
data_format=data_format,
conv2d_hparams=conv2d_hparams,
batch_norm_hparams=batch_norm_hparams,
name='bottleneck_2'
)
name='bottleneck_2',
cardinality=cardinality)
bottleneck = model.blocks.conv2d_block(
bottleneck,
@ -130,4 +131,12 @@ def bottleneck_block(
name='bottleneck_3'
)
if use_se:
bottleneck = model.layers.squeeze_excitation_layer(
inputs=bottleneck,
ratio=ratio,
training=training,
data_format=data_format,
name='bottleneck_se_layer')
return model.layers.relu(shortcut + bottleneck, name='relu')

7
TensorFlow/Classification/RN50v1.5/model/layers/__init__.py → TensorFlow/Classification/ConvNets/model/layers/__init__.py
View file

@ -18,6 +18,7 @@
from model.layers.activation import relu
from model.layers.activation import softmax
from model.layers.activation import tanh
from model.layers.activation import sigmoid
from model.layers.conv2d import conv2d
@ -31,6 +32,7 @@ from model.layers.padding import pad
from model.layers.pooling import average_pooling2d
from model.layers.pooling import max_pooling2d
from model.layers.squeeze_excitation_layer import squeeze_excitation_layer
__all__ = [
@ -38,6 +40,7 @@ __all__ = [
'relu',
'softmax',
'tanh',
'sigmoid',
# conv layers
'conv2d',
@ -56,5 +59,7 @@ __all__ = [
# pooling layers
'average_pooling2d',
'max_pooling2d'
'max_pooling2d',
'squeeze_excitation_layer'
]

8
TensorFlow/Classification/RN50v1.5/model/layers/activation.py → TensorFlow/Classification/ConvNets/model/layers/activation.py
View file

@ -17,7 +17,7 @@
import tensorflow as tf
__all__ = ['relu', 'softmax', 'tanh']
__all__ = ['relu', 'softmax', 'tanh', 'sigmoid']
def relu(inputs, name='relu'):
@ -43,3 +43,9 @@ def tanh(inputs, name='tanh'):
net = tf.math.tanh(inputs, name=name)
return net
def sigmoid(inputs, name='sigmoid'):
net = tf.math.sigmoid(inputs, name=name)
return net

0
TensorFlow/Classification/RN50v1.5/model/layers/conv2d.py → TensorFlow/Classification/ConvNets/model/layers/conv2d.py
View file

0
TensorFlow/Classification/RN50v1.5/model/layers/dense.py → TensorFlow/Classification/ConvNets/model/layers/dense.py
View file

0
TensorFlow/Classification/RN50v1.5/model/layers/math_ops.py → TensorFlow/Classification/ConvNets/model/layers/math_ops.py
View file

0
TensorFlow/Classification/RN50v1.5/model/layers/normalization.py → TensorFlow/Classification/ConvNets/model/layers/normalization.py
View file

0
TensorFlow/Classification/RN50v1.5/model/layers/padding.py → TensorFlow/Classification/ConvNets/model/layers/padding.py
View file

0
TensorFlow/Classification/RN50v1.5/model/layers/pooling.py → TensorFlow/Classification/ConvNets/model/layers/pooling.py
View file

84
TensorFlow/Classification/ConvNets/model/layers/squeeze_excitation_layer.py Normal file
View file

@ -0,0 +1,84 @@
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import tensorflow as tf
from model import layers
from model import blocks
__all__ = ['squeeze_excitation_layer']
def squeeze_excitation_layer(
inputs,
ratio,
training=True,
data_format='NCHW',
kernel_initializer=tf.variance_scaling_initializer(),
bias_initializer=tf.zeros_initializer(),
name="squeeze_excitation_layer"
):
if data_format not in ['NHWC', 'NCHW']:
raise ValueError("Unknown data format: `%s` (accepted: ['NHWC', 'NCHW'])" % data_format)
in_shape = inputs.get_shape()
num_channels = in_shape[1] if data_format == "NCHW" else in_shape[-1]
with tf.variable_scope(name):
net = inputs
# squeeze
squeeze = layers.reduce_mean(
net,
keepdims=False,
data_format=data_format,
name='squeeze_spatial_mean'
)
# fc + relu
excitation = layers.dense(
inputs=squeeze,
units=num_channels // ratio,
use_bias=True,
trainable=training,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer
)
excitation = layers.relu(excitation)
# fc + sigmoid
excitation = layers.dense(
inputs=excitation,
units=num_channels,
use_bias=True,
trainable=training,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer
)
excitation = layers.sigmoid(excitation)
out_shape = [-1, num_channels, 1, 1] if data_format == "NCHW" else [-1, 1, 1, num_channels]
excitation = tf.reshape(excitation, out_shape)
net = net * excitation
return net

206
TensorFlow/Classification/RN50v1.5/model/resnet_v1_5.py → TensorFlow/Classification/ConvNets/model/resnet.py
View file

@ -33,7 +33,6 @@ from utils.data_utils import normalized_inputs
from utils.learning_rate import learning_rate_scheduler
from utils.optimizers import FixedLossScalerOptimizer
from dllogger.logger import LOGGER
__all__ = [
'ResnetModel',
@ -47,10 +46,17 @@ class ResnetModel(object):
self,
model_name,
n_classes,
layers_count,
layers_depth,
expansions,
compute_format='NCHW',
input_format='NHWC',
weight_init='fan_out',
dtype=tf.float32,
use_dali=False,
cardinality=1,
use_se=False,
se_ratio=1,
):
self.model_hparams = tf.contrib.training.HParams(
@ -58,9 +64,14 @@ class ResnetModel(object):
compute_format=compute_format,
input_format=input_format,
dtype=dtype,
layer_counts=(3, 4, 6, 3),
layers_count=layers_count,
layers_depth=layers_depth,
expansions=expansions,
model_name=model_name,
use_dali=use_dali
use_dali=use_dali,
cardinality=cardinality,
use_se=use_se,
se_ratio=se_ratio
)
self.batch_norm_hparams = tf.contrib.training.HParams(
@ -78,63 +89,44 @@ class ResnetModel(object):
self.conv2d_hparams = tf.contrib.training.HParams(
kernel_initializer=tf.variance_scaling_initializer(
scale=2.0, distribution='truncated_normal', mode='fan_out'
scale=2.0, distribution='truncated_normal', mode=weight_init
),
bias_initializer=tf.constant_initializer(0.0)
)
self.dense_hparams = tf.contrib.training.HParams(
kernel_initializer=tf.variance_scaling_initializer(
scale=2.0, distribution='truncated_normal', mode='fan_out'
scale=2.0, distribution='truncated_normal', mode=weight_init
),
bias_initializer=tf.constant_initializer(0.0)
)
if hvd.rank() == 0:
LOGGER.log("Model HParams:")
LOGGER.log("Name", model_name)
LOGGER.log("Number of classes", n_classes)
LOGGER.log("Compute_format", compute_format)
LOGGER.log("Input_format", input_format)
LOGGER.log("dtype", str(dtype))
print("Model HParams:")
print("Name", model_name)
print("Number of classes", n_classes)
print("Compute_format", compute_format)
print("Input_format", input_format)
print("dtype", str(dtype))
def __call__(self, features, labels, mode, params):
if mode == tf.estimator.ModeKeys.TRAIN:
mandatory_params = ["batch_size", "lr_init", "num_gpus", "steps_per_epoch",
"momentum", "weight_decay", "loss_scale", "label_smoothing"]
for p in mandatory_params:
if p not in params:
raise RuntimeError("Parameter {} is missing.".format(p))
if "batch_size" not in params.keys():
raise RuntimeError("Parameter `batch_size` is missing...")
if "lr_init" not in params.keys():
raise RuntimeError("Parameter `lr_init` is missing...")
if "num_gpus" not in params.keys():
raise RuntimeError("Parameter `num_gpus` is missing...")
if "steps_per_epoch" not in params.keys():
raise RuntimeError("Parameter `steps_per_epoch` is missing...")
if "momentum" not in params.keys():
raise RuntimeError("Parameter `momentum` is missing...")
if "weight_decay" not in params.keys():
raise RuntimeError("Parameter `weight_decay` is missing...")
if "loss_scale" not in params.keys():
raise RuntimeError("Parameter `loss_scale` is missing...")
if "label_smoothing" not in params.keys():
raise RuntimeError("Parameter `label_smoothing` is missing...")
if mode == tf.estimator.ModeKeys.TRAIN and not self.model_hparams.use_dali:
with tf.device('/cpu:0'):
# Stage inputs on the host
cpu_prefetch_op, (features, labels) = ResnetModel._stage([features, labels])
cpu_prefetch_op, (features, labels) = self._stage([features, labels])
with tf.device('/gpu:0'):
# Stage inputs to the device
gpu_prefetch_op, (features, labels) = ResnetModel._stage([features, labels])
gpu_prefetch_op, (features, labels) = self._stage([features, labels])
with tf.device("/gpu:0"):
@ -159,7 +151,7 @@ class ResnetModel(object):
off_value = eta/1001)
if mixup != 0:
LOGGER.log("Using mixup training with beta=", params['mixup'])
print("Using mixup training with beta=", params['mixup'])
beta_distribution = tf.distributions.Beta(params['mixup'], params['mixup'])
feature_coefficients = beta_distribution.sample(sample_shape=[params['batch_size'], 1, 1, 1])
@ -205,13 +197,13 @@ class ResnetModel(object):
tf.identity(probs, name="probs_ref")
tf.identity(y_preds, name="y_preds_ref")
if mode == tf.estimator.ModeKeys.TRAIN:
assert (len(tf.trainable_variables()) == 161)
else:
assert (len(tf.trainable_variables()) == 0)
#if mode == tf.estimator.ModeKeys.TRAIN:
#
# assert (len(tf.trainable_variables()) == 161)
#
#else:
#
# assert (len(tf.trainable_variables()) == 0)
if mode == tf.estimator.ModeKeys.PREDICT:
@ -303,11 +295,9 @@ class ResnetModel(object):
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=params["momentum"])
if params["apply_loss_scaling"]:
optimizer = FixedLossScalerOptimizer(optimizer, scale=params["loss_scale"])
if hvd_utils.is_using_hvd():
optimizer = hvd.DistributedOptimizer(optimizer)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
@ -321,15 +311,13 @@ class ResnetModel(object):
if self.model_hparams.use_dali:
train_ops = tf.group(backprop_op, update_ops, name='train_ops')
else:
train_ops = tf.group(backprop_op, cpu_prefetch_op, gpu_prefetch_op, update_ops, name='train_ops')
return tf.estimator.EstimatorSpec(mode=mode, loss=total_loss, train_op=train_ops)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = {
"top1_accuracy": (acc_top1, acc_top1_update_op),
"top5_accuracy": (acc_top5, acc_top5_update_op)
@ -363,22 +351,20 @@ class ResnetModel(object):
return put_op, get_tensors
def build_model(self, inputs, training=True, reuse=False):
with var_storage.model_variable_scope(
self.model_hparams.model_name,
reuse=reuse,
dtype=self.model_hparams.dtype
):
dtype=self.model_hparams.dtype):
with tf.variable_scope("input_reshape"):
if self.model_hparams.input_format == 'NHWC' and self.model_hparams.compute_format == 'NCHW':
# Reshape inputs: NHWC => NCHW
inputs = tf.transpose(inputs, [0, 3, 1, 2])
elif self.model_hparams.input_format == 'NCHW' and self.model_hparams.compute_format == 'NHWC':
# Reshape inputs: NCHW => NHWC
inputs = tf.transpose(inputs, [0, 2, 3, 1])
@ -390,7 +376,7 @@ class ResnetModel(object):
n_channels=64,
kernel_size=(7, 7),
strides=(2, 2),
mode='SAME_RESNET',
mode='SAME',
use_batch_norm=True,
activation='relu',
is_training=training,
@ -409,74 +395,28 @@ class ResnetModel(object):
name="max_pooling2d",
)
for block_id, _ in enumerate(range(self.model_hparams.layer_counts[0])):
model_bottlenecks = self.model_hparams.layers_depth
for block_id, block_bottleneck in enumerate(model_bottlenecks):
for layer_id in range(self.model_hparams.layers_count[block_id]):
stride = 2 if (layer_id == 0 and block_id != 0) else 1
net = blocks.bottleneck_block(
inputs=net,
depth=256,
depth_bottleneck=64,
stride=1,
training=training,
data_format=self.model_hparams.compute_format,
conv2d_hparams=self.conv2d_hparams,
batch_norm_hparams=self.batch_norm_hparams,
block_name="btlnck_block_1_%d" % (block_id + 1)
)
for block_id, i in enumerate(range(self.model_hparams.layer_counts[1])):
stride = 2 if i == 0 else 1
net = blocks.bottleneck_block(
inputs=net,
depth=512,
depth_bottleneck=128,
stride=stride,
training=training,
data_format=self.model_hparams.compute_format,
conv2d_hparams=self.conv2d_hparams,
batch_norm_hparams=self.batch_norm_hparams,
block_name="btlnck_block_2_%d" % (block_id + 1)
)
for block_id, i in enumerate(range(self.model_hparams.layer_counts[2])):
block_id += 1
stride = 2 if i == 0 else 1
net = blocks.bottleneck_block(
inputs=net,
depth=1024,
depth_bottleneck=256,
stride=stride,
training=training,
data_format=self.model_hparams.compute_format,
conv2d_hparams=self.conv2d_hparams,
batch_norm_hparams=self.batch_norm_hparams,
block_name="btlnck_block_3_%d" % (block_id + 1)
)
for block_id, i in enumerate(range(self.model_hparams.layer_counts[3])):
stride = 2 if i == 0 else 1
net = blocks.bottleneck_block(
inputs=net,
depth=2048,
depth_bottleneck=512,
stride=stride,
training=training,
data_format=self.model_hparams.compute_format,
conv2d_hparams=self.conv2d_hparams,
batch_norm_hparams=self.batch_norm_hparams,
block_name="btlnck_block_4_%d" % (block_id + 1)
)
net = blocks.bottleneck_block(
inputs=net,
depth=block_bottleneck * self.model_hparams.expansions,
depth_bottleneck=block_bottleneck,
cardinality=self.model_hparams.cardinality,
stride=stride,
training=training,
data_format=self.model_hparams.compute_format,
conv2d_hparams=self.conv2d_hparams,
batch_norm_hparams=self.batch_norm_hparams,
block_name="btlnck_block_%d_%d" % (block_id, layer_id),
use_se=self.model_hparams.use_se,
ratio=self.model_hparams.se_ratio)
with tf.variable_scope("output"):
net = layers.reduce_mean(
net, keepdims=False, data_format=self.model_hparams.compute_format, name='spatial_mean'
)
net, keepdims=False, data_format=self.model_hparams.compute_format, name='spatial_mean')
logits = layers.dense(
inputs=net,
@ -484,8 +424,7 @@ class ResnetModel(object):
use_bias=True,
trainable=training,
kernel_initializer=self.dense_hparams.kernel_initializer,
bias_initializer=self.dense_hparams.bias_initializer
)
bias_initializer=self.dense_hparams.bias_initializer)
if logits.dtype != tf.float32:
logits = tf.cast(logits, tf.float32)
@ -493,3 +432,28 @@ class ResnetModel(object):
probs = layers.softmax(logits, name="softmax", axis=1)
return probs, logits
model_architectures = {
'resnet50': {
'layers': [3, 4, 6, 3],
'widths': [64, 128, 256, 512],
'expansions': 4,
},
'resnext101-32x4d': {
'layers': [3, 4, 23, 3],
'widths': [128, 256, 512, 1024],
'expansions': 2,
'cardinality': 32,
},
'se-resnext101-32x4d' : {
'cardinality' : 32,
'layers' : [3, 4, 23, 3],
'widths' : [128, 256, 512, 1024],
'expansions' : 2,
'use_se': True,
'se_ratio': 16,
},
}

1
TensorFlow/Classification/ConvNets/requirements.txt Normal file
View file

@ -0,0 +1 @@
-e git://github.com/NVIDIA/dllogger#egg=dllogger

663
TensorFlow/Classification/ConvNets/resnet50v1.5/README.md Normal file
View file

@ -0,0 +1,663 @@
# ResNet-50 v1.5 for TensorFlow
This repository provides a script and recipe to train the ResNet-50 v1.5 model to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.
## Table Of Contents
* [Model overview](#model-overview)
* [Default configuration](#default-configuration)
* [Optimizer](#optimizer)
* [Data augmentation](#data-augmentation)
* [Feature support matrix](#feature-support-matrix)
* [Features](#features)
* [Mixed precision training](#mixed-precision-training)
* [Enabling mixed precision](#enabling-mixed-precision)
* [Setup](#setup)
* [Requirements](#requirements)
* [Quick Start Guide](#quick-start-guide)
* [Advanced](#advanced)
* [Scripts and sample code](#scripts-and-sample-code)
* [Parameters](#parameters)
* [The `main.py` script](#the-mainpy-script)
* [Inference process](#inference-process)
* [Performance](#performance)
* [Benchmarking](#benchmarking)
* [Training performance benchmark](#training-performance-benchmark)
* [Inference performance benchmark](#inference-performance-benchmark)
* [Results](#results)
* [Training accuracy results](#training-accuracy-results)
* [Training accuracy: NVIDIA DGX-1 (8x V100 16G)](#training-accuracy-nvidia-dgx-1-8x-v100-16g)
* [Training performance results](#training-performance-results)
* [Training performance: NVIDIA DGX-1 (8x V100 16G)](#training-performance-nvidia-dgx-1-8x-v100-16g)
* [Training performance: NVIDIA DGX-2 (16x V100 32G)](#training-performance-nvidia-dgx-2-16x-v100-32g)
* [Training time for 90 Epochs](#training-time-for-90-epochs)
* [Training time: NVIDIA DGX-1 (8x V100 16G)](#training-time-nvidia-dgx-1-8x-v100-16g)
* [Training time: NVIDIA DGX-2 (16x V100 32G)](#training-time-nvidia-dgx-2-16x-v100-32g)
* [Inference performance results](#inference-performance-results)
* [Inference performance: NVIDIA DGX-1 (1x V100 16G)](#inference-performance-nvidia-dgx-1-1x-v100-16g)
* [Inference performance: NVIDIA DGX-2 (1x V100 32G)](#inference-performance-nvidia-dgx-2-1x-v100-32g)
* [Inference performance: NVIDIA T4 (1x T4)](#inference-performance-nvidia-t4-1x-t4-16g)
* [Release notes](#release-notes)
* [Changelog](#changelog)
* [Known issues](#known-issues)
## Model overview
The ResNet50 v1.5 model is a modified version of the [original ResNet50 v1 model](https://arxiv.org/abs/1512.03385).
The difference between v1 and v1.5 is in the bottleneck blocks which requires
downsampling, for example, v1 has stride = 2 in the first 1x1 convolution, whereas v1.5 has stride = 2 in the 3x3 convolution.
This difference makes ResNet50 v1.5 slightly more accurate (~0.5% top1) than v1,
but comes with a small performance drawback (~5% imgs/sec).
The following performance optimizations were implemented in this model:
* JIT graph compilation with [XLA](https://www.tensorflow.org/xla)
* Multi-GPU training with [Horovod](https://github.com/horovod/horovod)
* NVIDIA Data Loading ([DALI](https://github.com/NVIDIA/DALI)).
This model is trained with mixed precision using Tensor Cores on NVIDIA Volta and Turing GPUs. Therefore, researchers can get results 3x faster than training without Tensor Cores, while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.
### Default configuration
The following sections highlight the default configuration for the ResNet50 model.
#### Optimizer
This model uses the SGD optimizer with the following hyperparameters:
* Momentum (0.875).
* Learning rate (LR) = 0.256 for 256 batch size, for other batch sizes we linearly scale the learning
rate.
* Learning rate schedule - we use cosine LR schedule.
* For bigger batch sizes (512 and up) we use linear warmup of the learning rate.
during the first 5 epochs according to [Training ImageNet in 1 hour](https://arxiv.org/abs/1706.02677).
* Weight decay: 3.0517578125e-05 (1/32768).
* We do not apply Weight decay on batch norm trainable parameters (gamma/bias).
* Label Smoothing: 0.1.
* We train for:
* 90 Epochs -> 90 epochs is a standard for ResNet50
* 250 Epochs -> best possible accuracy.
* For 250 epoch training we also use [MixUp regularization](https://arxiv.org/pdf/1710.09412.pdf).
#### Data Augmentation
This model uses the following data augmentation:
* For training:
* Normalization.
* Random resized crop to 224x224.
* Scale from 8% to 100%.
* Aspect ratio from 3/4 to 4/3.
* Random horizontal flip.
* For inference:
* Normalization.
* Scale to 256x256.
* Center crop to 224x224.
### Feature support matrix
The following features are supported by this model.
| Feature | ResNet-50 v1.5 Tensorflow |
|-----------------------|--------------------------
|Multi-GPU training with [Horovod](https://github.com/horovod/horovod) | Yes |
|[NVIDIA DALI](https://docs.nvidia.com/deeplearning/dali/release-notes/index.html) | Yes |
|Automatic mixed precision (AMP) | Yes |
#### Features
Multi-GPU training with Horovod - Our model uses Horovod to implement efficient multi-GPU training with NCCL.
For details, refer to the example sources in this repository or the [TensorFlow tutorial](https://github.com/horovod/horovod/#usage).
NVIDIA DALI - DALI is a library accelerating data preparation pipeline. To accelerate your input pipeline, you only need to define your data loader
with the DALI library. For details, refer to the example sources in this repository or the [DALI documentation](https://docs.nvidia.com/deeplearning/dali/index.html).
Automatic mixed precision (AMP) - Computation graph can be modified by TensorFlow on runtime to support mixed precision training.
Detailed explanation of mixed precision can be found in the next section.
### Mixed precision training
Mixed precision is the combined use of different numerical precisions in a computational method. [Mixed precision](https://arxiv.org/abs/1710.03740) training offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of [Tensor Cores](https://developer.nvidia.com/tensor-cores) in the Volta and Turing architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using [mixed precision training](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html) previously required two steps:
1. Porting the model to use the FP16 data type where appropriate.
2. Manually adding loss scaling to preserve small gradient values.
This can now be achieved using Automatic Mixed Precision (AMP) for TensorFlow to enable the full [mixed precision methodology](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html#tensorflow) in your existing TensorFlow model code. AMP enables mixed precision training on Volta and Turing GPUs automatically. The TensorFlow framework code makes all necessary model changes internally.
In TF-AMP, the computational graph is optimized to use as few casts as necessary and maximize the use of FP16, and the loss scaling is automatically applied inside of supported optimizers. AMP can be configured to work with the existing tf.contrib loss scaling manager by disabling the AMP scaling with a single environment variable to perform only the automatic mixed-precision optimization. It accomplishes this by automatically rewriting all computation graphs with the necessary operations to enable mixed precision training and automatic loss scaling.
For information about:
* How to train using mixed precision, see the [Mixed Precision Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed Precision](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html) documentation.
* How to access and enable AMP for TensorFlow, see [Using TF-AMP](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-user-guide/index.html#tfamp) from the TensorFlow User Guide.
* Techniques used for mixed precision training, see the [Mixed-Precision Training of Deep Neural Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) blog.
#### Enabling mixed precision
Mixed precision is enabled in TensorFlow by using the Automatic Mixed Precision (TF-AMP) extension which casts variables to half-precision upon retrieval, while storing variables in single-precision format. Furthermore, to preserve small gradient magnitudes in backpropagation, a [loss scaling](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html#lossscaling) step must be included when applying gradients. In TensorFlow, loss scaling can be applied statically by using simple multiplication of loss by a constant value or automatically, by TF-AMP. Automatic mixed precision makes all the adjustments internally in TensorFlow, providing two benefits over manual operations. First, programmers need not modify network model code, reducing development and maintenance effort. Second, using AMP maintains forward and backward compatibility with all the APIs for defining and running TensorFlow models.
To enable mixed precision, you can simply add the values to the environmental variables inside your training script:
- Enable TF-AMP graph rewrite:
```
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "1"
```
- Enable Automated Mixed Precision:
```
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
```
## Setup
The following section lists the requirements that you need to meet in order to use the ResNet50 v1.5 model.
### Requirements
This repository contains Dockerfile which extends the TensorFlow NGC container and encapsulates all dependencies. Aside from these dependencies, ensure you have the following software:
* [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)
* [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
* [NVIDIA Volta](https://www.nvidia.com/en-us/data-center/volta-gpu-architecture/) or [Turing](https://www.nvidia.com/en-us/geforce/turing/) based GPU
For more information about how to get started with NGC containers, see the
following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:
* [Getting Started Using NVIDIA GPU Cloud](https://docs.nvidia.com/ngc/ngc-getting-started-guide/index.html),
* [Accessing And Pulling From The NGC container registry](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#accessing_registry),
* [Running TensorFlow](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/running.html#running).
For those unable to use the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) to set up the required environment or create your own container, see the versioned [NVIDIA Container Support Matrix](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).
## Quick Start Guide
To train your model using mixed precision with Tensor Cores or FP32, perform the following steps using the default parameters of the ResNet-50 v1.5 model on the [ImageNet](http://www.image-net.org/) dataset. For the specifics concerning training and inference, see the [Advanced](#advanced) section.
1. Clone the repository.
```
git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/Classification/RN50v1.5
```
2. Download and preprocess the dataset.
The ResNet50 v1.5 script operates on ImageNet 1k, a widely popular image classification dataset from the ILSVRC challenge.
To download and preprocess the dataset, use the [Generate ImageNet for TensorFlow](https://github.com/tensorflow/models/blob/master/research/inception/inception/data/download_and_preprocess_imagenet.sh) script. The dataset will be downloaded to a directory specified as the first parameter of the script.
3. Build the ResNet-50 v1.5 TensorFlow NGC container.
```bash
docker build . -t nvidia_rn50
```
4. Start an interactive session in the NGC container to run training/inference.
After you build the container image, you can start an interactive CLI session with
```bash
nvidia-docker run --rm -it -v <path to imagenet>:/data/tfrecords --ipc=host nvidia_rn50
```
5. (Optional) Create index files to use DALI.
To allow proper sharding in a multi-GPU environment, DALI has to create index files for the dataset. To create index files, run inside the container:
```bash
bash ./utils/dali_index.sh /data/tfrecords <index file store location>
```
Index files can be created once and then reused. It is highly recommended to save them into a persistent location.
6. Start training.
To run training for a standard configuration (as described in [Default
configuration](#default-configuration), DGX1V, DGX2V, single GPU, FP16, FP32, 50, 90, and 250 epochs), run
one of the scripts int the `resnet50v1.5/training` directory. Ensure ImageNet is mounted in the
`/data/tfrecords` directory.
For example, to train on DGX-1 for 90 epochs using AMP, run:
`bash ./resnet50v1.5/training/AMP/DGX1_RN50_AMP_90E.sh`
Additionally, features like DALI data preprocessing or TensorFlow XLA can be enabled with
environmental variables when running those scripts:
`USE_XLA=1 USE_DALI=1 bash ./resnet50v1.5/training/AMP/DGX1_RN50_AMP_90E.sh`
To store results in a specific location, add a location as a first argument:
`bash ./resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh <location to store>`
7. Start validation/evaluation.
To evaluate the validation dataset located in `/data/tfrecords`, run `main.py` with
`--mode=evaluate`. For example:
`python main.py --mode=evaluate --data_dir=/data/tfrecords --batch_size <batch size> --model_dir
<model location> --result_dir <output location> [--use_xla] [--use_tf_amp]`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during evaluation.
## Advanced
The following sections provide greater details of the dataset, running training and inference, and the training results.
### Scripts and sample code
In the root directory, the most important files are:
- `main.py`: the script that controls the logic of training and validation of the ResNet-like models
- `Dockerfile`: Instructions for Docker to build a container with the basic set of dependencies to run ResNet like models for image classification
- `requirements.txt`: a set of extra Python requirements for running ResNet-like models
The `model/` directory contains the following modules used to define ResNet family models:
- `resnet.py`: the definition of ResNet, ResNext, and SE-ResNext model
- `blocks/conv2d_block.py`: the definition of 2D convolution block
- `blocks/resnet_bottleneck_block.py`: the definition of ResNet-like bottleneck block
- `layers/*.py`: definitions of specific layers used in the ResNet-like model
The `utils/` directory contains the following utility modules:
- `cmdline_helper.py`: helper module for command line processing
- `data_utils.py`: module defining input data pipelines
- `dali_utils.py`: helper module for DALI
- `hvd_utils.py`: helper module for Horovod
- `image_processing.py`: image processing and data augmentation functions
- `learning_rate.py`: definition of used learning rate schedule
- `optimizers.py`: definition of used custom optimizers
- `hooks/*.py`: definitions of specific hooks allowing logging of training and inference process
The `runtime/` directory contains the following module that define the mechanics of the training process:
- `runner.py`: module encapsulating the training, inference and evaluation
### Parameters
#### The `main.py` script
The script for training and evaluating the ResNet-50 v1.5 model has a variety of parameters that control these processes.
```
usage: main.py [-h]
[--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}]
[--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}]
[--data_dir DATA_DIR] [--data_idx_dir DATA_IDX_DIR]
[--export_dir EXPORT_DIR] [--to_predict TO_PREDICT]
[--batch_size BATCH_SIZE] [--num_iter NUM_ITER]
[--iter_unit {epoch,batch}] [--warmup_steps WARMUP_STEPS]
[--model_dir MODEL_DIR] [--results_dir RESULTS_DIR]
[--log_filename LOG_FILENAME] [--display_every DISPLAY_EVERY]
[--lr_init LR_INIT] [--lr_warmup_epochs LR_WARMUP_EPOCHS]
[--weight_decay WEIGHT_DECAY] [--weight_init {fan_in,fan_out}]
[--momentum MOMENTUM] [--loss_scale LOSS_SCALE]
[--label_smoothing LABEL_SMOOTHING] [--mixup MIXUP]
[--use_static_loss_scaling | --nouse_static_loss_scaling]
[--use_xla | --nouse_xla] [--use_dali | --nouse_dali]
[--use_tf_amp | --nouse_tf_amp]
[--use_cosine_lr | --nouse_cosine_lr] [--seed SEED]
[--gpu_memory_fraction GPU_MEMORY_FRACTION] [--gpu_id GPU_ID]
JoC-RN50v1.5-TF
optional arguments:
-h, --help Show this help message and exit
--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}
Architecture of model to run (default is resnet50)
--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}
The execution mode of the script.
--data_dir DATA_DIR Path to dataset in TFRecord format. Files should be
named 'train-*' and 'validation-*'.
--data_idx_dir DATA_IDX_DIR
Path to index files for DALI. Files should be named
'train-*' and 'validation-*'.
--export_dir EXPORT_DIR
Directory in which to write exported SavedModel.
--to_predict TO_PREDICT
Path to file or directory of files to run prediction
on.
--batch_size BATCH_SIZE
Size of each minibatch per GPU.
--num_iter NUM_ITER Number of iterations to run.
--iter_unit {epoch,batch}
Unit of iterations.
--warmup_steps WARMUP_STEPS
Number of steps considered as warmup and not taken
into account for performance measurements.
--model_dir MODEL_DIR
Directory in which to write the model. If undefined,
results directory will be used.
--results_dir RESULTS_DIR
Directory in which to write training logs, summaries
and checkpoints.
--log_filename LOG_FILENAME
Name of the JSON file to which write the training log
--display_every DISPLAY_EVERY
How often (in batches) to print out running
information.
--lr_init LR_INIT Initial value for the learning rate.
--lr_warmup_epochs LR_WARMUP_EPOCHS
Number of warmup epochs for the learning rate schedule.
--weight_decay WEIGHT_DECAY
Weight Decay scale factor.
--weight_init {fan_in,fan_out}
Model weight initialization method.
--momentum MOMENTUM SGD momentum value for the momentum optimizer.
--loss_scale LOSS_SCALE
Loss scale for FP16 training and fast math FP32.
--label_smoothing LABEL_SMOOTHING
The value of label smoothing.
--mixup MIXUP The alpha parameter for mixup (if 0 then mixup is not
applied).
--use_static_loss_scaling
Use static loss scaling in FP16 or FP32 AMP.
--nouse_static_loss_scaling
--use_xla Enable XLA (Accelerated Linear Algebra) computation
for improved performance.
--nouse_xla
--use_dali Enable DALI data input.
--nouse_dali
--use_tf_amp Enable AMP to speedup FP32
computation using Tensor Cores.
--nouse_tf_amp
--use_cosine_lr Use cosine learning rate schedule.
--nouse_cosine_lr
--seed SEED Random seed.
--gpu_memory_fraction GPU_MEMORY_FRACTION
Limit memory fraction used by the training script for DALI
--gpu_id GPU_ID Specify the ID of the target GPU on a multi-device platform.
Effective only for single-GPU mode.
```
### Inference process
To run inference on a single example with a checkpoint and a model script, use:
`python main.py --mode predict --model_dir <path to model> --to_predict <path to image> --results_dir <path to results>`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during inference.
## Performance
### Benchmarking
The following section shows how to run benchmarks measuring the model performance in training and inference modes.
#### Training performance benchmark
To benchmark the training performance on a specific batch size, run:
* For 1 GPU
* FP32
`python ./main.py --mode=training_benchmark --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --mode=training_benchmark --use_tf_amp --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* For multiple GPUs
* FP32
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --mode=training_benchmark --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --mode=training_benchmark --use_tf_amp --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
Each of these scripts runs 200 warm-up iterations and measures the first epoch.
To control warmup and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags. Features like XLA or DALI can be controlled
with `--use_xla` and `--use_dali` flags.
Suggested batch sizes for training are 256 for mixed precision training and 128 for single precision training.
#### Inference performance benchmark
To benchmark the inference performance on a specific batch size, run:
* FP32
`python ./main.py --mode=inference_benchmark --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --mode=inference_benchmark --use_tf_amp --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
By default, each of these scripts runs 20 warm-up iterations and measures the next 80 iterations.
To control warm-up and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags.
### Results
The following sections provide details on how we achieved our performance and accuracy in training and inference.
#### Training accuracy results
##### Training accuracy: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the `/resnet50v1.5/training/{PRECISION}/DGX1_RN50_{PRECISION}_{EPOCHS}E.sh`
training script in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
NGC container on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| Epochs | Batch Size / GPU | Accuracy - FP32 | Accuracy - mixed precision |
|--------|------------------|-----------------|----------------------------|
| 50 | 128 (FP32) / 256 (AMP) | 76.06 | 75.96 |
| 90 | 128 (FP32) / 256 (AMP) | 77.08 | 77.01 |
| 250 | 128 (FP32) / 256 (AMP) | 78.34 | 78.35 |
**Example training loss plot**
![TrainingLoss](./imgs/train_loss.png)
#### Training performance results
##### Training performance: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (8x V100 16G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|------------------------|-----------------|-----------|-------------------|
| 1 | 128 (FP32) / 256 (AMP) | 369.62 img/s | 1099.55 img/s | 2.97x | 1.00x | 1.00x |
| 8 | 128 (FP32) / 256 (AMP) | 2785.81 img/s | 8277.91 img/s | 2.97x | 7.54x | 7.53x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|------------|---------------|---------------------|-----------|
| 1 | 256 | 1099.55 img/s |1217.90 img/s | 1.11x |
| 8 | 256 | 8277.91 img/s |9485.21 img/s | 1.14x |
##### Training performance: NVIDIA DGX-2 (16x V100 32G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (16x V100 32G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|-------------------------|-------|--------|--------|
| 1 | 128 (FP32) / 256 (AMP) | 385.26 img/s | 1020.80 img/s | 2.64x | 1.00x | 1.00x |
| 16 | 128 (FP32) / 256 (AMP) | 5347.83 img/s | 16290.42 img/s | 3.04x | 13.88x | 15.95x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|-----|----------|---------------------|-----------|
| 1 | 256 | 1020.80 img/s | 1157.97 img/s |1.13x |
| 16 | 256 | 16290.42 img/s | 18304.71 img/s |1.12x |
#### Training Time for 90 Epochs
##### Training time: NVIDIA DGX-1 (8x V100 16G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-1-8x-v100-16g)
on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|---|--------|---------|---------|
| 1 | ~26h | ~29h | ~86h |
| 8 | ~3.5h | ~4h | ~11.5h |
##### Training time: NVIDIA DGX-2 (16x V100 32G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-2-16x-v100-32g)
on NVIDIA DGX-2 with (16x V100 32G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|----|-------|--------|-------|
| 1 | ~27h | ~31h | ~83h |
| 16 | ~1.7h | ~2h | ~6h |
#### Inference performance results
##### Inference performance: NVIDIA DGX-1 (1x V100 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (1x V100 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |173.35 img/s |5.79 ms |5.90 ms |5.95 ms |6.04 ms |
|2 |303.65 img/s |6.61 ms |6.80 ms |6.87 ms |7.01 ms |
|4 |562.35 img/s |7.12 ms |7.32 ms |7.42 ms |7.69 ms |
|8 |783.24 img/s |10.22 ms |10.37 ms |10.44 ms |10.60 ms |
|16 |1003.10 img/s |15.99 ms |16.07 ms |16.12 ms |16.29 ms |
|32 |1140.12 img/s |28.19 ms |28.27 ms |28.38 ms |28.54 ms |
|64 |1252.06 img/s |51.12 ms |51.82 ms |52.75 ms |53.45 ms |
|128 |1324.91 img/s |96.61 ms |97.02 ms |97.25 ms |99.08 ms |
|256 |1348.52 img/s |189.85 ms |191.16 ms |191.77 ms |192.47 ms|
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |237.35 img/s |4.25 ms |4.39 ms |4.54 ms |5.30 ms |
|2 |464.94 img/s |4.32 ms |4.63 ms |4.83 ms |5.52 ms |
|4 |942.44 img/s |4.26 ms |4.55 ms |4.74 ms |5.45 ms |
|8 |1454.93 img/s |5.57 ms |5.73 ms |5.91 ms |6.51 ms |
|16 |2003.75 img/s |8.13 ms |8.19 ms |8.29 ms |8.50 ms |
|32 |2356.17 img/s |13.69 ms |13.82 ms |13.92 ms |14.26 ms |
|64 |2706.11 img/s |23.86 ms |23.82 ms |23.89 ms |24.10 ms |
|128 |2770.61 img/s |47.04 ms |49.36 ms |62.43 ms |90.05 ms |
|256 |2742.14 img/s |94.67 ms |108.02 ms |119.34 ms |145.55 ms|
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |162.95 img/s |6.16 ms |6.28 ms |6.34 ms |6.50 ms |
|2 |335.63 img/s |5.96 ms |6.10 ms |6.14 ms |6.25 ms |
|4 |637.72 img/s |6.30 ms |6.53 ms |7.17 ms |8.10 ms |
|8 |1153.92 img/s |7.03 ms |7.97 ms |8.22 ms |9.00 ms |
|16 |1906.52 img/s |8.64 ms |9.51 ms |9.88 ms |10.47 ms |
|32 |2492.78 img/s |12.84 ms |13.06 ms |13.13 ms |13.24 ms |
|64 |2910.05 img/s |22.66 ms |21.82 ms |24.71 ms |48.61 ms |
|128 |2964.31 img/s |45.25 ms |59.30 ms |71.42 ms |98.72 ms |
|256 |2898.12 img/s |90.53 ms |106.12 ms |118.12 ms |150.78 ms|
##### Inference performance: NVIDIA DGX-2 (1x V100 32G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (1x V100 32G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |187.41 img/s |5.374 ms |5.61 ms |5.70 ms |6.33 ms |
|2 |339.52 img/s |5.901 ms |6.16 ms |6.29 ms |6.53 ms |
|4 |577.50 img/s |6.940 ms |7.07 ms |7.24 ms |7.99 ms |
|8 |821.15 img/s |9.751 ms |9.99 ms |10.15 ms |10.80 ms|
|16 |1055.64 img/s |15.209 ms |15.26 ms |15.30 ms |16.14 ms|
|32 |1195.74 img/s |26.772 ms |26.93 ms |26.98 ms |27.80 ms|
|64 |1313.83 img/s |48.796 ms |48.99 ms |49.72 ms |51.83 ms|
|128 |1372.58 img/s |93.262 ms |93.90 ms |94.97 ms |96.57 ms|
|256 |1414.99 img/s |180.923 ms |181.65 ms |181.92 ms |183.37 ms|
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |289.89 img/s |3.50 ms |3.81 ms |3.90 ms |4.19 ms |
|2 |606.27 img/s |3.38 ms |3.56 ms |3.76 ms |4.25 ms |
|4 |982.92 img/s |4.09 ms |4.42 ms |4.53 ms |4.81 ms |
|8 |1553.34 img/s |5.22 ms |5.31 ms |5.50 ms |6.74 ms |
|16 |2091.27 img/s |7.82 ms |7.77 ms |7.82 ms |8.77 ms |
|32 |2457.61 img/s |13.14 ms |13.15 ms |13.21 ms |13.37 ms|
|64 |2746.11 img/s |23.31 ms |23.50 ms |23.56 ms |24.31 ms|
|128 |2937.20 img/s |43.58 ms |43.76 ms |43.82 ms |44.37 ms|
|256 |3009.83 img/s |85.06 ms |86.23 ms |87.37 ms |88.67 ms|
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |240.66 img/s |4.22 ms |4.59 ms |4.69 ms |4.84 ms |
|2 |428.60 img/s |4.70 ms |5.11 ms |5.44 ms |6.01 ms |
|4 |945.38 img/s |4.26 ms |4.35 ms |4.42 ms |4.74 ms |
|8 |1518.66 img/s |5.33 ms |5.50 ms |5.63 ms |5.88 ms |
|16 |2091.66 img/s |7.83 ms |7.74 ms |7.79 ms |8.88 ms |
|32 |2604.17 img/s |12.40 ms |12.45 ms |12.51 ms |12.61 ms|
|64 |3101.15 img/s |20.64 ms |20.93 ms |21.00 ms |21.17 ms|
|128 |3408.72 img/s |37.55 ms |37.93 ms |38.05 ms |38.53 ms|
|256 |3633.85 img/s |70.85 ms |70.93 ms |71.12 ms |71.45 ms|
##### Inference performance: NVIDIA T4 (1x T4 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA T4 with (1x T4 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |136.44 img/s |7.34 ms |7.43 ms |7.47 ms |7.54 ms |
|2 |215.38 img/s |9.29 ms |9.42 ms |9.46 ms |9.59 ms |
|4 |289.29 img/s |13.83 ms |14.08 ms |14.16 ms |14.40 ms |
|8 |341.77 img/s |23.41 ms |23.79 ms |23.86 ms |24.11 ms |
|16 |394.36 img/s |40.58 ms |40.87 ms |40.98 ms |41.41 ms |
|32 |414.66 img/s |77.18 ms |78.05 ms |78.29 ms |78.67 ms |
|64 |424.42 img/s |150.82 ms |152.99 ms |153.44 ms |154.34 ms |
|128 |429.83 img/s |297.82 ms |301.09 ms |301.60 ms |302.51 ms |
|256 |425.72 img/s |601.37 ms |605.74 ms |606.47 ms |608.74 ms |
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |211.04 img/s |4.77 ms |5.05 ms |5.08 ms |5.15 ms |
|2 |381.23 img/s |5.27 ms |5.40 ms |5.45 ms |5.52 ms |
|4 |593.13 img/s |6.75 ms |6.89 ms |6.956 ms |7.02 ms |
|8 |791.12 img/s |10.16 ms |10.35 ms |10.43 ms |10.68 ms |
|16 |914.26 img/s |17.55 ms |17.80 ms |17,89 ms |18.19 ms |
|32 |972.36 img/s |32.92 ms |33.33 ms |33.46 ms |33.61 ms |
|64 |991.39 img/s |64.56 ms |65.62 ms |65.92 ms |66.35 ms |
|128 |995.81 img/s |128.55 ms |130.03 ms |130.37 ms |131.08 ms |
|256 |993.39 img/s |257.71 ms |259.26 ms |259.62 ms |260.36 ms |
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |167.01 img/s |6.01 ms |6.12 ms |6.14 ms |6.18 ms |
|2 |333.67 img/s |6.03 ms |6.11 ms |6.15 ms |6.23 ms |
|4 |605.94 img/s |6.63 ms |6.79 ms |6.86 ms |7.02 ms |
|8 |802.13 img/s |9.98 ms |10.14 ms |10.22 ms |10.36 ms |
|16 |986.85 img/s |16.27 ms |16.36 ms |16.42 ms |16.52 ms |
|32 |1090.38 img/s |29.35 ms |29.68 ms |29.79 ms |30.07 ms |
|64 |1131.56 img/s |56.63 ms |57.22 ms |57.41 ms |57.76 ms |
|128 |1167.62 img/s |109.77 ms |111.06 ms |111.27 ms |111.85 ms|
|256 |1193.74 img/s |214.46 ms |216.28 ms |216.86 ms |217.80 ms|
## Release notes
### Changelog
1. March, 2019
* Initial release
2. May, 2019
* Added DALI support
* Added scripts for DGX-2
* Added benchmark results for DGX-2 and XLA-enabled DGX-1 and DGX-2.
3. July, 2019
* Added Cosine learning rate schedule
3. August, 2019
* Added mixup regularization
* Added T4 benchmarks
* Improved inference capabilities
* Added SavedModel export
4. January, 2020
* Removed manual checks for dataset paths to facilitate cloud storage solutions
* Move to a new logging solution
* Bump base docker image version
5. March, 2020
* Code cleanup and refactor
* Improved training process
6. June, 2020
* Added ResNext and SE-ResNext architectures
### Known issues
There are no known issues with this model.

BIN
TensorFlow/Classification/ConvNets/resnet50v1.5/imgs/train_loss.png Normal file
View file

Binary file not shown.
Side by side

After

Width:  |  Height:  |  Size: 59 KiB

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX1_RN50_AMP_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 250 256 amp --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX1_RN50_AMP_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 50 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX1_RN50_AMP_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 90 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX2_RN50_AMP_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 250 256 amp --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX2_RN50_AMP_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 50 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/DGX2_RN50_AMP_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 90 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/GPU1_RN50_AMP_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 250 256 amp --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/GPU1_RN50_AMP_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 50 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/AMP/GPU1_RN50_AMP_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 90 256 amp

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX1_RN50_FP16_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 250 256 fp16 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX1_RN50_FP16_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 50 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX1_RN50_FP16_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 90 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX2_RN50_FP16_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 250 256 fp16 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX2_RN50_FP16_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 50 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/DGX2_RN50_FP16_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 90 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/GPU1_RN50_FP16_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 250 256 fp16 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/GPU1_RN50_FP16_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 50 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP16/GPU1_RN50_FP16_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 90 256 fp16

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX1_RN50_FP32_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 250 128 fp32 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX1_RN50_FP32_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 50 128 fp32

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX1_RN50_FP32_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
8 90 128 fp32

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX2_RN50_FP32_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 250 128 fp32 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX2_RN50_FP32_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 50 128 fp32

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/DGX2_RN50_FP32_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
16 90 128 fp32

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/GPU1_RN50_FP32_250E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 250 128 fp32 --mixup=0.2

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/GPU1_RN50_FP32_50E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 50 128 fp32

5
TensorFlow/Classification/ConvNets/resnet50v1.5/training/FP32/GPU1_RN50_FP32_90E.sh Normal file
View file

@ -0,0 +1,5 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
bash ${WORKSPACE}/resnet50v1.5/training/GENERIC.sh ${WORKSPACE} ${DATA_DIR} \
1 90 128 fp32

39
TensorFlow/Classification/ConvNets/resnet50v1.5/training/GENERIC.sh Normal file
View file

@ -0,0 +1,39 @@
WORKSPACE=${1:-"/workspace/rn50v15_tf"}
DATA_DIR=${2:-"/data"}
GPU_COUNT=${3:-8}
ITER_COUNT=${4:-50}
BATCH_SIZE=${5:-128}
PRECISION=${6:-"fp32"}
OTHER=${@:7}
if [[ ! -z "${BIND_TO_SOCKET}" ]]; then
BIND_TO_SOCKET="--bind-to socket"
fi
if [[ ! -z "${USE_DALI}" ]]; then
USE_DALI="--use_dali --data_idx_dir=${DATA_DIR}/dali_idx"
fi
if [[ ! -z "${USE_XLA}" ]]; then
USE_XLA="--use_xla"
fi
CMD=""
case $PRECISION in
"fp32") CMD+="--precision=fp32";;
"fp16") CMD+="--precision=fp16 --use_static_loss_scaling --loss_scale=128";;
"amp") CMD+="--precision=fp32 --use_tf_amp --use_static_loss_scaling --loss_scale=128";;
esac
CMD="--arch=resnet50 --mode=train_and_evaluate --iter_unit=epoch --num_iter=${ITER_COUNT} \
--batch_size=${BATCH_SIZE} --warmup_steps=100 --use_cosine --label_smoothing 0.1 \
--lr_init=0.256 --lr_warmup_epochs=8 --momentum=0.875 --weight_decay=3.0517578125e-05 \
${CMD} --data_dir=${DATA_DIR}/tfrecords ${USE_DALI} ${USE_XLA} \
--results_dir=${WORKSPACE}/results --weight_init=fan_in ${OTHER}"
if [[ ${GPU_COUNT} -eq 1 ]]; then
python3 main.py ${CMD}
else
mpiexec --allow-run-as-root ${BIND_TO_SOCKET} -np ${GPU_COUNT} python3 main.py ${CMD}
fi

654
TensorFlow/Classification/ConvNets/resnext101-32x4d/README.md Normal file
View file

@ -0,0 +1,654 @@
# ResNext101-32x4d for TensorFlow
This repository provides a script and recipe to train the ResNext101-32x4d model to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.
## Table Of Contents
* [Model overview](#model-overview)
* [Model architecture](#model-architecture)
* [Default configuration](#default-configuration)
* [Optimizer](#optimizer)
* [Data augmentation](#data-augmentation)
* [Feature support matrix](#feature-support-matrix)
* [Features](#features)
* [Mixed precision training](#mixed-precision-training)
* [Enabling mixed precision](#enabling-mixed-precision)
* [Setup](#setup)
* [Requirements](#requirements)
* [Quick Start Guide](#quick-start-guide)
* [Advanced](#advanced)
* [Scripts and sample code](#scripts-and-sample-code)
* [Parameters](#parameters)
* [The `main.py` script](#the-mainpy-script)
* [Inference process](#inference-process)
* [Performance](#performance)
* [Benchmarking](#benchmarking)
* [Training performance benchmark](#training-performance-benchmark)
* [Inference performance benchmark](#inference-performance-benchmark)
* [Results](#results)
* [Training accuracy results](#training-accuracy-results)
* [Training accuracy: NVIDIA DGX-1 (8x V100 16G)](#training-accuracy-nvidia-dgx-1-8x-v100-16g)
* [Training performance results](#training-performance-results)
* [Training performance: NVIDIA DGX-1 (8x V100 16G)](#training-performance-nvidia-dgx-1-8x-v100-16g)
* [Training performance: NVIDIA DGX-2 (16x V100 32G)](#training-performance-nvidia-dgx-2-16x-v100-32g)
* [Training time for 90 Epochs](#training-time-for-90-epochs)
* [Training time: NVIDIA DGX-1 (8x V100 16G)](#training-time-nvidia-dgx-1-8x-v100-16g)
* [Training time: NVIDIA DGX-2 (16x V100 32G)](#training-time-nvidia-dgx-2-16x-v100-32g)
* [Inference performance results](#inference-performance-results)
* [Inference performance: NVIDIA DGX-1 (1x V100 16G)](#inference-performance-nvidia-dgx-1-1x-v100-16g)
* [Inference performance: NVIDIA DGX-2 (1x V100 32G)](#inference-performance-nvidia-dgx-2-1x-v100-32g)
* [Inference performance: NVIDIA T4 (1x T4)](#inference-performance-nvidia-t4-1x-t4-16g)
* [Release notes](#release-notes)
* [Changelog](#changelog)
* [Known issues](#known-issues)
## Model overview
The ResNeXt101-32x4d is a model introduced in the [Aggregated Residual Transformations for Deep Neural Networks](https://arxiv.org/pdf/1611.05431.pdf) paper.
It is based on a regular ResNet model, substituting 3x3 convolutions inside the bottleneck block for 3x3 grouped convolutions.
The following performance optimizations were implemented in this model:
* JIT graph compilation with [XLA](https://www.tensorflow.org/xla)
* Multi-GPU training with [Horovod](https://github.com/horovod/horovod)
* Automated mixed precision [AMP](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html)
This model is trained with mixed precision using Tensor Cores on NVIDIA Volta and Turing GPUs. Therefore, researchers can get results 3x faster than training without Tensor Cores, while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.
### Model architecture
![ResNextArch](./imgs/ResNeXtArch.png)
_Image source: [Aggregated Residual Transformations for Deep Neural Networks](https://arxiv.org/pdf/1611.05431.pdf)_
Image shows difference between ResNet bottleneck block and ResNeXt bottleneck block.
ResNeXt bottleneck block splits single convolution into multiple smaller, parallel convolutions.
ResNeXt101-32x4d model's cardinality equals 32 and bottleneck width equals 4. This means instead of single convolution with 64 filters
32 parallel convolutions with only 4 filters are used.
### Default configuration
The following sections highlight the default configuration for the ResNext101-32x4d model.
#### Optimizer
This model uses the SGD optimizer with the following hyperparameters:
* Momentum (0.875).
* Learning rate (LR) = 0.256 for 256 batch size, for other batch sizes we linearly scale the learning
rate.
* Learning rate schedule - we use cosine LR schedule.
* For bigger batch sizes (512 and up) we use linear warmup of the learning rate.
during the first 5 epochs according to [Training ImageNet in 1 hour](https://arxiv.org/abs/1706.02677).
* Weight decay: 6.103515625e-05 (1/16384).
* We do not apply Weight decay on batch norm trainable parameters (gamma/bias).
* Label Smoothing: 0.1.
* We train for:
* 90 Epochs -> 90 epochs is a standard for ImageNet networks.
* 250 Epochs -> best possible accuracy.
* For 250 epoch training we also use [MixUp regularization](https://arxiv.org/pdf/1710.09412.pdf).
#### Data Augmentation
This model uses the following data augmentation:
* For training:
* Normalization.
* Random resized crop to 224x224.
* Scale from 8% to 100%.
* Aspect ratio from 3/4 to 4/3.
* Random horizontal flip.
* For inference:
* Normalization.
* Scale to 256x256.
* Center crop to 224x224.
### Feature support matrix
The following features are supported by this model.
| Feature | ResNext101-32x4d Tensorflow |
|-----------------------|--------------------------
|Multi-GPU training with [Horovod](https://github.com/horovod/horovod) | Yes |
|[NVIDIA DALI](https://docs.nvidia.com/deeplearning/dali/release-notes/index.html) | Yes |
|Automatic mixed precision (AMP) | Yes |
#### Features
Multi-GPU training with Horovod - Our model uses Horovod to implement efficient multi-GPU training with NCCL.
For details, refer to the example sources in this repository or the [TensorFlow tutorial](https://github.com/horovod/horovod/#usage).
NVIDIA DALI - DALI is a library accelerating data preparation pipeline. To accelerate your input pipeline, you only need to define your data loader
with the DALI library. For details, refer to the example sources in this repository or the [DALI documentation](https://docs.nvidia.com/deeplearning/dali/index.html).
Automatic mixed precision (AMP) - Computation graph can be modified by TensorFlow on runtime to support mixed precision training.
Detailed explanation of mixed precision can be found in the next section.
### Mixed precision training
Mixed precision is the combined use of different numerical precisions in a computational method. [Mixed precision](https://arxiv.org/abs/1710.03740) training offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of [Tensor Cores](https://developer.nvidia.com/tensor-cores) in the Volta and Turing architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using [mixed precision training](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html) previously required two steps:
1. Porting the model to use the FP16 data type where appropriate.
2. Manually adding loss scaling to preserve small gradient values.
This can now be achieved using Automatic Mixed Precision (AMP) for TensorFlow to enable the full [mixed precision methodology](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html#tensorflow) in your existing TensorFlow model code. AMP enables mixed precision training on Volta and Turing GPUs automatically. The TensorFlow framework code makes all necessary model changes internally.
In TF-AMP, the computational graph is optimized to use as few casts as necessary and maximize the use of FP16, and the loss scaling is automatically applied inside of supported optimizers. AMP can be configured to work with the existing tf.contrib loss scaling manager by disabling the AMP scaling with a single environment variable to perform only the automatic mixed-precision optimization. It accomplishes this by automatically rewriting all computation graphs with the necessary operations to enable mixed precision training and automatic loss scaling.
For information about:
* How to train using mixed precision, see the [Mixed Precision Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed Precision](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html) documentation.
* How to access and enable AMP for TensorFlow, see [Using TF-AMP](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-user-guide/index.html#tfamp) from the TensorFlow User Guide.
* Techniques used for mixed precision training, see the [Mixed-Precision Training of Deep Neural Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) blog.
#### Enabling mixed precision
Mixed precision is enabled in TensorFlow by using the Automatic Mixed Precision (TF-AMP) extension which casts variables to half-precision upon retrieval, while storing variables in single-precision format. Furthermore, to preserve small gradient magnitudes in backpropagation, a [loss scaling](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html#lossscaling) step must be included when applying gradients. In TensorFlow, loss scaling can be applied statically by using simple multiplication of loss by a constant value or automatically, by TF-AMP. Automatic mixed precision makes all the adjustments internally in TensorFlow, providing two benefits over manual operations. First, programmers need not modify network model code, reducing development and maintenance effort. Second, using AMP maintains forward and backward compatibility with all the APIs for defining and running TensorFlow models.
To enable mixed precision, you can simply add the values to the environmental variables inside your training script:
- Enable TF-AMP graph rewrite:
```
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "1"
```
- Enable Automated Mixed Precision:
```
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
```
## Setup
The following section lists the requirements that you need to meet in order to use the ResNext101-32x4d model.
### Requirements
This repository contains Dockerfile which extends the TensorFlow NGC container and encapsulates all dependencies. Aside from these dependencies, ensure you have the following software:
* [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)
* [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
* [NVIDIA Volta](https://www.nvidia.com/en-us/data-center/volta-gpu-architecture/) or [Turing](https://www.nvidia.com/en-us/geforce/turing/) based GPU
For more information about how to get started with NGC containers, see the
following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:
* [Getting Started Using NVIDIA GPU Cloud](https://docs.nvidia.com/ngc/ngc-getting-started-guide/index.html),
* [Accessing And Pulling From The NGC container registry](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#accessing_registry),
* [Running TensorFlow](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/running.html#running).
For those unable to use the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) to set up the required environment or create your own container, see the versioned [NVIDIA Container Support Matrix](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).
## Quick Start Guide
To train your model using mixed precision with Tensor Cores or FP32, perform the following steps using the default parameters of the ResNext101-32x4d model on the [ImageNet](http://www.image-net.org/) dataset. For the specifics concerning training and inference, see the [Advanced](#advanced) section.
1. Clone the repository.
```
git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/Classification/RN50v1.5
```
2. Download and preprocess the dataset.
The ResNext101-32x4d script operates on ImageNet 1k, a widely popular image classification dataset from the ILSVRC challenge.
To download and preprocess the dataset, use the [Generate ImageNet for TensorFlow](https://github.com/tensorflow/models/blob/master/research/inception/inception/data/download_and_preprocess_imagenet.sh) script. The dataset will be downloaded to a directory specified as the first parameter of the script.
3. Build the ResNext101-32x4d TensorFlow NGC container.
```bash
docker build . -t nvidia_rn50
```
4. Start an interactive session in the NGC container to run training/inference.
After you build the container image, you can start an interactive CLI session with
```bash
nvidia-docker run --rm -it -v <path to imagenet>:/data/tfrecords --ipc=host nvidia_rn50
```
5. (Optional) Create index files to use DALI.
To allow proper sharding in a multi-GPU environment, DALI has to create index files for the dataset. To create index files, run inside the container:
```bash
bash ./utils/dali_index.sh /data/tfrecords <index file store location>
```
Index files can be created once and then reused. It is highly recommended to save them into a persistent location.
6. Start training.
To run training for a standard configuration (as described in [Default
configuration](#default-configuration), DGX1V, DGX2V, single GPU, FP16, FP32, 90, and 250 epochs), run
one of the scripts in the `resnext101-32x4d/training` directory. Ensure ImageNet is mounted in the
`/data/tfrecords` directory.
For example, to train on DGX-1 for 90 epochs using AMP, run:
`bash ./resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh`
Additionally, features like DALI data preprocessing or TensorFlow XLA can be enabled with
environmental variables when running those scripts:
`USE_XLA=1 USE_DALI=1 bash ./resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh`
To store results in a specific location, add a location as a first argument:
`bash ./resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh <location to store>`
7. Start validation/evaluation.
To evaluate the validation dataset located in `/data/tfrecords`, run `main.py` with
`--mode=evaluate`. For example:
`python main.py --arch=resnext101-32x4d --mode=evaluate --data_dir=/data/tfrecords --batch_size <batch size> --model_dir
<model location> --result_dir <output location> [--use_xla] [--use_tf_amp]`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during evaluation.
## Advanced
The following sections provide greater details of the dataset, running training and inference, and the training results.
### Scripts and sample code
In the root directory, the most important files are:
- `main.py`: the script that controls the logic of training and validation of the ResNet-like models
- `Dockerfile`: Instructions for Docker to build a container with the basic set of dependencies to run ResNet like models for image classification
- `requirements.txt`: a set of extra Python requirements for running ResNet-like models
The `model/` directory contains the following modules used to define ResNet family models:
- `resnet.py`: the definition of ResNet, ResNext, and SE-ResNext model
- `blocks/conv2d_block.py`: the definition of 2D convolution block
- `blocks/resnet_bottleneck_block.py`: the definition of ResNet-like bottleneck block
- `layers/*.py`: definitions of specific layers used in the ResNet-like model
The `utils/` directory contains the following utility modules:
- `cmdline_helper.py`: helper module for command line processing
- `data_utils.py`: module defining input data pipelines
- `dali_utils.py`: helper module for DALI
- `hvd_utils.py`: helper module for Horovod
- `image_processing.py`: image processing and data augmentation functions
- `learning_rate.py`: definition of used learning rate schedule
- `optimizers.py`: definition of used custom optimizers
- `hooks/*.py`: definitions of specific hooks allowing logging of training and inference process
The `runtime/` directory contains the following module that define the mechanics of the training process:
- `runner.py`: module encapsulating the training, inference and evaluation
### Parameters
#### The `main.py` script
The script for training and evaluating the ResNext101-32x4d model has a variety of parameters that control these processes.
```
usage: main.py [-h]
[--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}]
[--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}]
[--data_dir DATA_DIR] [--data_idx_dir DATA_IDX_DIR]
[--export_dir EXPORT_DIR] [--to_predict TO_PREDICT]
[--batch_size BATCH_SIZE] [--num_iter NUM_ITER]
[--iter_unit {epoch,batch}] [--warmup_steps WARMUP_STEPS]
[--model_dir MODEL_DIR] [--results_dir RESULTS_DIR]
[--log_filename LOG_FILENAME] [--display_every DISPLAY_EVERY]
[--lr_init LR_INIT] [--lr_warmup_epochs LR_WARMUP_EPOCHS]
[--weight_decay WEIGHT_DECAY] [--weight_init {fan_in,fan_out}]
[--momentum MOMENTUM] [--loss_scale LOSS_SCALE]
[--label_smoothing LABEL_SMOOTHING] [--mixup MIXUP]
[--use_static_loss_scaling | --nouse_static_loss_scaling]
[--use_xla | --nouse_xla] [--use_dali | --nouse_dali]
[--use_tf_amp | --nouse_tf_amp]
[--use_cosine_lr | --nouse_cosine_lr] [--seed SEED]
[--gpu_memory_fraction GPU_MEMORY_FRACTION] [--gpu_id GPU_ID]
JoC-RN50v1.5-TF
optional arguments:
-h, --help Show this help message and exit
--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}
Architecture of model to run (to run Resnext-32x4d set
--arch=rensext101-32x4d)
--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}
The execution mode of the script.
--data_dir DATA_DIR Path to dataset in TFRecord format. Files should be
named 'train-*' and 'validation-*'.
--data_idx_dir DATA_IDX_DIR
Path to index files for DALI. Files should be named
'train-*' and 'validation-*'.
--export_dir EXPORT_DIR
Directory in which to write exported SavedModel.
--to_predict TO_PREDICT
Path to file or directory of files to run prediction
on.
--batch_size BATCH_SIZE
Size of each minibatch per GPU.
--num_iter NUM_ITER Number of iterations to run.
--iter_unit {epoch,batch}
Unit of iterations.
--warmup_steps WARMUP_STEPS
Number of steps considered as warmup and not taken
into account for performance measurements.
--model_dir MODEL_DIR
Directory in which to write the model. If undefined,
results directory will be used.
--results_dir RESULTS_DIR
Directory in which to write training logs, summaries
and checkpoints.
--log_filename LOG_FILENAME
Name of the JSON file to which write the training log
--display_every DISPLAY_EVERY
How often (in batches) to print out running
information.
--lr_init LR_INIT Initial value for the learning rate.
--lr_warmup_epochs LR_WARMUP_EPOCHS
Number of warmup epochs for the learning rate schedule.
--weight_decay WEIGHT_DECAY
Weight Decay scale factor.
--weight_init {fan_in,fan_out}
Model weight initialization method.
--momentum MOMENTUM SGD momentum value for the momentum optimizer.
--loss_scale LOSS_SCALE
Loss scale for FP16 training and fast math FP32.
--label_smoothing LABEL_SMOOTHING
The value of label smoothing.
--mixup MIXUP The alpha parameter for mixup (if 0 then mixup is not
applied).
--use_static_loss_scaling
Use static loss scaling in FP16 or FP32 AMP.
--nouse_static_loss_scaling
--use_xla Enable XLA (Accelerated Linear Algebra) computation
for improved performance.
--nouse_xla
--use_dali Enable DALI data input.
--nouse_dali
--use_tf_amp Enable AMP to speedup FP32
computation using Tensor Cores.
--nouse_tf_amp
--use_cosine_lr Use cosine learning rate schedule.
--nouse_cosine_lr
--seed SEED Random seed.
--gpu_memory_fraction GPU_MEMORY_FRACTION
Limit memory fraction used by the training script for DALI
--gpu_id GPU_ID Specify the ID of the target GPU on a multi-device platform.
Effective only for single-GPU mode.
```
### Inference process
To run inference on a single example with a checkpoint and a model script, use:
`python main.py --arch=resnext101-32x4d --mode predict --model_dir <path to model> --to_predict <path to image> --results_dir <path to results>`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during inference.
## Performance
### Benchmarking
The following section shows how to run benchmarks measuring the model performance in training and inference modes.
#### Training performance benchmark
To benchmark the training performance on a specific batch size, run:
* For 1 GPU
* FP32
`python ./main.py --arch=resnext101-32x4d --mode=training_benchmark --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --arch=resnext101-32x4d --mode=training_benchmark --use_tf_amp --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* For multiple GPUs
* FP32
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --arch=resnext101-32x4d --mode=training_benchmark --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --arch=resnext101-32x4d --mode=training_benchmark --use_tf_amp --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
Each of these scripts runs 200 warm-up iterations and measures the first epoch.
To control warmup and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags. Features like XLA or DALI can be controlled
with `--use_xla` and `--use_dali` flags.
Suggested batch sizes for training are 128 for mixed precision training and 64 for single precision training per single V100 16 GB.
#### Inference performance benchmark
To benchmark the inference performance on a specific batch size, run:
* FP32
`python ./main.py --arch=resnext101-32x4d --mode=inference_benchmark --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --arch=resnext101-32x4d --mode=inference_benchmark --use_tf_amp --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
By default, each of these scripts runs 20 warm-up iterations and measures the next 80 iterations.
To control warm-up and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags.
The benchmark can be automated with the `inference_benchmark.sh` script provided in `resnext101-32x4d`, by simply running:
`bash ./resnext101-32x4d/inference_benchmark.sh <data dir> <data idx dir>`
The `<data dir>` parameter refers to the input data directory (by default `/data/tfrecords` inside the container).
By default, the benchmark tests the following configurations: **FP32**, **AMP**, **AMP + XLA** with different batch sizes.
When the optional directory with the DALI index files `<data idx dir>` is specified, the benchmark executes an additional **DALI + AMP + XLA** configuration.
### Results
The following sections provide details on how we achieved our performance and accuracy in training and inference.
#### Training accuracy results
##### Training accuracy: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the `/resnext101-32x4d/training/{PRECISION}/DGX1_RNxt101-32x4d_{PRECISION}_{EPOCHS}E.sh`
training script in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
NGC container on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| Epochs | Batch Size / GPU | Accuracy - FP32 | Accuracy - mixed precision |
|--------|------------------|-----------------|----------------------------|
| 90 | 64 (FP32) / 128 (AMP) | 79.34 | 79.31 |
| 250 | 64 (FP32) / 128 (AMP) | 80.21 | 80.21 |
**Example training loss plot**
![TrainingLoss](./imgs/train_loss.png)
#### Training performance results
##### Training performance: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (8x V100 16G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|------------------------|-----------------|-----------|-------------------|
| 1 | 64 (FP32) / 128 (AMP) | 142.10 img/s | 423.19 img/s | 2.97x | 1.00x | 1.00x |
| 8 | 64 (FP32) / 128 (AMP) | 1055.82 img/s | 3151.81 img/s | 2.98x | 7.43x | 7.44x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|------------|---------------|---------------------|-----------|
| 1 | 128 | 423.19 img/s | 588.49 img/s | 1.39x |
| 8 | 128 | 3151.81 img/s | 4231.42 img/s | 1.34x |
##### Training performance: NVIDIA DGX-2 (16x V100 32G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (16x V100 32G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|-------------------------|-------|--------|--------|
| 1 | 64 (FP32) / 128 (AMP) | 148.19 img/s | 403.13 img/s | 2.72x | 1.00x | 1.00x |
| 16 | 64 (FP32) / 128 (AMP) | 1961.31 img/s | 5601.13 img/s | 2.86x | 13.23x | 13.89x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|-----|----------|---------------------|-----------|
| 1 | 128 | 403.13 img/s | 555.33 img/s |1.13x |
| 16 | 128 | 5601.13 img/s | 7617.25 img/s |1.36x |
#### Training Time for 90 Epochs
##### Training time: NVIDIA DGX-1 (8x V100 16G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-1-8x-v100-16g)
on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|---|--------|---------|---------|
| 1 | ~54h | ~75h | ~225h |
| 8 | ~7.5h | ~10h | ~30h |
##### Training time: NVIDIA DGX-2 (16x V100 32G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-2-16x-v100-32g)
on NVIDIA DGX-2 with (16x V100 32G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|----|-------|--------|-------|
| 1 | ~57h | ~79h | ~216h |
| 16 | ~4.2h | ~6h | ~16h |
#### Inference performance results
##### Inference performance: NVIDIA DGX-1 (1x V100 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (1x V100 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |61.19 img/s |16.36 ms |16.66 ms |16.87 ms |17.31 ms |
|2 |120.52 img/s |16.60 ms |16.91 ms |17.00 ms |17.60 ms |
|4 |179.63 img/s |22.26 ms |22.44 ms |22.50 ms |22.73 ms |
|8 |287.94 img/s |27.78 ms |27.97 ms |28.08 ms |28.30 ms |
|16 |403.04 img/s |39.72 ms |39.93 ms |40.01 ms |40.29 ms |
|32 |463.61 img/s |69.03 ms |69.68 ms |70.99 ms |71.48 ms |
|64 |530.00 img/s |120.75 ms |121.12 ms |121.38 ms |123.17 ms |
|128 |570.60 img/s |224.32 ms |224.84 ms |224.98 ms |225.72 ms |
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |164.92 img/s |6.10 ms |6.17 ms |6.26 ms |7.73 ms |
|2 |326.59 img/s |6.14 ms |6.32 ms |6.39 ms |6.62 ms |
|4 |607.20 img/s |6.60 ms |6.77 ms |6.88 ms |8.08 ms |
|8 |892.31 img/s |8.97 ms |9.13 ms |9.49 ms |9.86 ms |
|16 |1259.92 img/s |12.82 ms |13.31 ms |13.44 ms |13.58 ms |
|32 |1508.73 img/s |31.30 ms |21.70 ms |21.86 ms |22.02 ms |
|64 |1618.77 img/s |39.55 ms |40.71 ms |41.33 ms |41.94 ms |
|128 |1730.40 img/s |73.98 ms |74.27 ms |76.01 ms |76.74 ms |
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |97.24 img/s |10.31 ms |10.48 ms |10.57 ms |10.81 ms |
|2 |191.68 img/s |10.44 ms |10.74 ms |10.84 ms |11.42 ms |
|4 |381.19 img/s |10.50 ms |10.85 ms |10.98 ms |11.74 ms |
|8 |744.11 img/s |10.77 ms |11.42 ms |11.85 ms |12.44 ms |
|16 |1174.29 img/s |13.83 ms |13.87 ms |14.29 ms |15.53 ms |
|32 |1439.07 img/s |22.33 ms |22.67 ms |22.84 ms |23.06 ms |
|64 |1712.76 img/s |37.37 ms |37.91 ms |38.09 ms |38.74 ms |
|128 |1883.71 img/s |67.95 ms |68.48 ms |68.63 ms |68.86 ms |
##### Inference performance: NVIDIA DGX-2 (1x V100 32G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (1x V100 32G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |62.02 img/s |16.22 ms |17.62 ms |17.92 ms |19.21 ms |
|2 |97.98 img/s |20.54 ms |20.59 ms |20.72 ms |23.21 ms |
|4 |168.16 img/s |23.79 ms |24.12 ms |24.24 ms |26.94 ms |
|8 |269.89 img/s |29.66 ms |30.01 ms |30.35 ms |34.05 ms|
|16 |379.81 img/s |42.14 ms |42.47 ms |42.85 ms |47.63 ms|
|32 |466.04 img/s |68.67 ms |68.99 ms |69.26 ms |74.87 ms|
|64 |547.64 img/s |117.01 ms |117.59 ms |118.37 ms |122.83 ms|
|128 |603.44 img/s |212.21 ms |212.92 ms |214.09 ms |217.06 ms|
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |198.53 img/s |5.14 ms |5.23 ms |5.41 ms |5.54 ms |
|2 |343.00 img/s |6.14 ms |6.08 ms |6.26 ms |7.72 ms |
|4 |592.25 img/s |6.77 ms |7.06 ms |7.18 ms |8.70 ms |
|8 |918.45 img/s |8.72 ms |8.90 ms |9.09 ms |9.77 ms |
|16 |1306.53 img/s |12.60 ms |12.65 ms |12.91 ms |17.06 ms |
|32 |1483.83 img/s |21.56 ms |21.61 ms |21.84 ms |27.05 ms|
|64 |1668.63 img/s |38.39 ms |38.50 ms |40.15 ms |43.15 ms|
|128 |1748.25 img/s |73.35 ms |75.23 ms |78.82 ms |80.17 ms|
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |143.65 img/s |6.97 ms |7.15 ms |7.24 ms |7.95 ms |
|2 |282.21 img/s |7.09 ms |7.32 ms |7.56 ms |7.97 ms |
|4 |511.55 img/s |7.85 ms |8.42 ms |8.62 ms |9.02 ms |
|8 |870.60 img/s |9.23 ms |9.46 ms |9.54 ms |9.88 ms |
|16 |1179.93 img/s |13.62 ms |14.04 ms |14.19 ms |14.51 ms|
|32 |1512.36 img/s |21.19 ms |21.70 ms |21.80 ms |22.04 ms|
|64 |1805.38 img/s |35.56 ms |36.33 ms |36.48 ms |36.94 ms|
|128 |1947.49 img/s |65.88 ms |66.50 ms |66.72 ms |67.17 ms|
##### Inference performance: NVIDIA T4 (1x T4 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA T4 with (1x T4 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |54.23 img/s |18.48 ms |19.62 ms |19.78 ms |20.13 ms |
|2 |94.22 img/s |21.24 ms |21.58 ms |21.71 ms |21.97 ms |
|4 |127.71 img/s |31.33 ms |31.90 ms |32.10 ms |32.50 ms |
|8 |151.88 img/s |52.67 ms |53.45 ms |53.80 ms |54.12 ms |
|16 |163.01 img/s |98.16 ms |99.52 ms |99.94 ms |100.49 ms |
|32 |176.13 img/s |181.71 ms |183.91 ms |184.54 ms |185.60 ms |
|64 |183.40 img/s |349.00 ms |352.65 ms |353.55 ms |355.03 ms |
|128 |182.77 img/s |700.35 ms |707.89 ms |708.80 ms |710.28 ms |
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |157.87 img/s |6.36 ms |6.47 ms |6.52 ms |6.64 ms |
|2 |274.10 img/s |7.29 ms |7.41 ms |7.45 ms |7.51 ms |
|4 |395.41 img/s |10.12 ms |10.35 ms |10.41 ms |10.53 ms |
|8 |479.83 img/s |16.68 ms |16.92 ms |17.01 ms |17.15 ms |
|16 |525.83 img/s |30.47 ms |30.80 ms |30.89 ms |31.27 ms |
|32 |536.31 img/s |59.67 ms |60.35 ms |60.51 ms |60.96 ms |
|64 |541.26 img/s |118.25 ms |119.51 ms |119.77 ms |120.38 ms |
|128 |538.20 img/s |237.84 ms |240.41 ms |240.82 ms |241.72 ms |
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |104.10 img/s |9.63 ms |9.75 ms |9.78 ms |9.86 ms |
|2 |220.23 img/s |9.08 ms |9.22 ms |9.26 ms |9.35 ms |
|4 |361.55 img/s |11.06 ms |11.19 ms |11.29 ms |11.68 ms |
|8 |452.95 img/s |17.66 ms |17.92 ms |18.00 ms |18.12 ms |
|16 |522.64 img/s |30.65 ms |30.92 ms |31.04 ms |31.36 ms |
|32 |542.06 img/s |59.03 ms |59.63 ms |59.77 ms |60.25 ms |
|64 |536.14 img/s |119.37 ms |120.31 ms |120.68 ms |121.39 ms |
|128 |548.43 img/s |233.50 ms |234.83 ms |235.31 ms |236.29 ms|
## Release notes
### Changelog
June 2020
- Initial release
### Known issues
There are no known issues with this model.

BIN
TensorFlow/Classification/ConvNets/resnext101-32x4d/imgs/ResNeXtArch.png Normal file
View file

Binary file not shown.
Side by side

After

Width:  |  Height:  |  Size: 78 KiB

BIN
TensorFlow/Classification/ConvNets/resnext101-32x4d/imgs/train_loss.png Normal file
View file

Binary file not shown.
Side by side

After

Width:  |  Height:  |  Size: 50 KiB

33
TensorFlow/Classification/ConvNets/resnext101-32x4d/inference_benchmark.sh Normal file
View file

@ -0,0 +1,33 @@
#!/bin/bash
DATA_DIR=${1:-"/data/tfrecords"}
DALI_DIR=${2}
BATCH_SIZE_TO_TEST="1 2 4 8 16 32 64 128"
INFERENCE_BENCHMARK=$(mktemp /tmp/inference-benchmark.XXXXXX)
function test_configuration() {
echo "Testing configuration: $1" | tee -a $INFERENCE_BENCHMARK
for BATCH in $BATCH_SIZE_TO_TEST; do
python ./main.py --arch=resnext101-32x4d --mode=inference_benchmark --warmup_steps 50 --num_iter 400 --iter_unit batch \
--batch_size $BATCH --data_dir=$DATA_DIR --results_dir=/tmp/results $2 | tail -n2 | head -n1 | sed \
's/^DLL \([0-9]*-\)*[0-9]* \([0-9]*:\)*[0-9]*.[0-9]* - ()/Results for BS='$BATCH'/' | tee -a $INFERENCE_BENCHMARK
if [ ! $? -eq 0 ]; then
echo "Failed test on batch size $BATCH_SIZE"
exit 1
fi
done
}
test_configuration "FP32 nodali noxla"
test_configuration "FP16 nodali noxla" "--use_tf_amp"
test_configuration "FP16 nodali xla" "--use_tf_amp --use_xla"
if [ ! -z $DALI_DIR ]; then
test_configuration "FP16 dali xla" "--use_tf_amp --use_xla --use_dali --data_idx_dir ${DALI_DIR}"
fi
cat $INFERENCE_BENCHMARK
rm $INFERENCE_BENCHMARK

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 90 128 amp

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/DGX2_RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/DGX2_RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 90 128 amp

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/GPU1_RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/AMP/GPU1_RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 90 128 amp

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/DGX1_RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/DGX1_RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 90 64 fp32

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/DGX2_RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/DGX2_RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 90 64 fp32

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/GPU1_RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/FP32/GPU1_RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 90 64 fp32

39
TensorFlow/Classification/ConvNets/resnext101-32x4d/training/GENERIC.sh Normal file
View file

@ -0,0 +1,39 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
DATA_DIR=${2:-"/data"}
GPU_COUNT=${3:-8}
ITER_COUNT=${4:-50}
BATCH_SIZE=${5:-128}
PRECISION=${6:-"fp32"}
OTHER=${@:7}
if [[ ! -z "${BIND_TO_SOCKET}" ]]; then
BIND_TO_SOCKET="--bind-to socket"
fi
if [[ ! -z "${USE_DALI}" ]]; then
USE_DALI="--use_dali --data_idx_dir=${DATA_DIR}/dali_idx"
fi
if [[ ! -z "${USE_XLA}" ]]; then
USE_XLA="--use_xla"
fi
CMD=""
case $PRECISION in
"fp32") CMD+="--precision=fp32";;
"fp16") CMD+="--precision=fp16 --use_static_loss_scaling --loss_scale=128";;
"amp") CMD+="--precision=fp32 --use_tf_amp --use_static_loss_scaling --loss_scale=128";;
esac
CMD="--arch=resnext101-32x4d --mode=train_and_evaluate --iter_unit=epoch --num_iter=${ITER_COUNT} \
--batch_size=${BATCH_SIZE} --warmup_steps=100 --use_cosine --label_smoothing 0.1 \
--lr_init=0.256 --lr_warmup_epochs=8 --momentum=0.875 --weight_decay=6.103515625e-05 \
${CMD} --data_dir=${DATA_DIR}/tfrecords ${USE_DALI} ${USE_XLA} \
--results_dir=${RESULT_DIR} --weight_init=fan_in ${OTHER}"
if [[ ${GPU_COUNT} -eq 1 ]]; then
python3 main.py ${CMD}
else
mpiexec --allow-run-as-root ${BIND_TO_SOCKET} -np ${GPU_COUNT} python3 main.py ${CMD}
fi

0
TensorFlow/Classification/RN50v1.5/runtime/__init__.py → TensorFlow/Classification/ConvNets/runtime/__init__.py
View file

304
TensorFlow/Classification/RN50v1.5/runtime/runner.py → TensorFlow/Classification/ConvNets/runtime/runner.py
View file

@ -19,10 +19,11 @@ import multiprocessing
import warnings
import tensorflow as tf
import numpy as np
import horovod.tensorflow as hvd
from model import resnet_v1_5
from model import resnet
from utils import hooks
from utils import data_utils
@ -30,21 +31,23 @@ from utils import hvd_utils
from runtime import runner_utils
from dllogger.logger import LOGGER
import dllogger
__all__ = [
'Runner',
]
class Runner(object):
def __init__(
self,
# ========= Model HParams ========= #
n_classes=1001,
input_format='NHWC', # NCHW or NHWC
architecture='resnet50',
input_format='NHWC', # NCHW or NHWC
compute_format='NCHW', # NCHW or NHWC
dtype=tf.float32, # tf.float32 or tf.float16
dtype=tf.float32, # tf.float32 or tf.float16
n_channels=3,
height=224,
width=224,
@ -53,6 +56,7 @@ class Runner(object):
log_dir=None,
data_dir=None,
data_idx_dir=None,
weight_init="fan_out",
# ======= Optimization HParams ======== #
use_xla=False,
@ -60,7 +64,7 @@ class Runner(object):
use_dali=False,
gpu_memory_fraction=1.0,
gpu_id=0,
# ======== Debug Flags ======== #
debug_verbosity=0,
seed=None
@ -78,10 +82,6 @@ class Runner(object):
if n_channels not in [1, 3]:
raise ValueError("Unsupported number of channels: %d (allowed: 1 (grayscale) and 3 (color))" % n_channels)
if data_dir is not None and not os.path.exists(data_dir):
raise ValueError("The `data_dir` received does not exists: %s" % data_dir)
hvd.init()
tf_seed = 2 * (seed + hvd.rank()) if seed is not None else None
# ============================================
@ -106,21 +106,20 @@ class Runner(object):
os.environ['TF_SYNC_ON_FINISH'] = '0'
os.environ['TF_AUTOTUNE_THRESHOLD'] = '2'
os.environ['TF_DISABLE_NVTX_RANGES'] = '1'
os.environ["TF_XLA_FLAGS"] = (os.environ.get("TF_XLA_FLAGS", "") + " --tf_xla_enable_lazy_compilation=false")
# ============================================
# TF-AMP Setup - Do not remove
# ============================================
if dtype == tf.float16:
if use_tf_amp:
raise RuntimeError("TF AMP can not be activated for FP16 precision")
elif use_tf_amp:
if hvd.rank() == 0:
LOGGER.log("TF AMP is activated - Experimental Feature")
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "1"
else:
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "0"
# =================================================
@ -135,12 +134,8 @@ class Runner(object):
distort_colors=distort_colors,
seed=tf_seed
)
if use_dali:
num_preprocessing_threads=4
else:
num_preprocessing_threads=10
num_preprocessing_threads = 10 if not use_dali else 4
run_config_performance = tf.contrib.training.HParams(
num_preprocessing_threads=num_preprocessing_threads,
use_tf_amp=use_tf_amp,
@ -160,20 +155,30 @@ class Runner(object):
self.run_hparams = Runner._build_hparams(model_hparams, run_config_additional, run_config_performance)
self._model = resnet_v1_5.ResnetModel(
model_name="resnet50_v1.5",
model_name = architecture
architecture = resnet.model_architectures[architecture]
self._model = resnet.ResnetModel(
model_name=model_name,
n_classes=model_hparams.n_classes,
layers_count=architecture["layers"],
layers_depth=architecture["widths"],
expansions=architecture["expansions"],
input_format=model_hparams.input_format,
compute_format=model_hparams.compute_format,
dtype=model_hparams.dtype,
use_dali=use_dali
weight_init=weight_init,
use_dali=use_dali,
cardinality=architecture['cardinality'] if 'cardinality' in architecture else 1,
use_se=architecture['use_se'] if 'use_se' in architecture else False,
se_ratio=architecture['se_ratio'] if 'se_ratio' in architecture else 1
)
if self.run_hparams.seed is not None:
if hvd.rank() == 0:
LOGGER.log("Deterministic Run - Seed: %d" % seed)
tf.set_random_seed(self.run_hparams.seed)
self.training_logging_hook = None
self.eval_logging_hook = None
@staticmethod
def _build_hparams(*args):
@ -209,34 +214,36 @@ class Runner(object):
def _get_session_config(mode, use_xla, use_dali, gpu_memory_fraction, gpu_id=0):
if mode not in ["train", 'validation', 'benchmark', 'inference']:
raise ValueError("Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode)
raise ValueError(
"Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode
)
# Limit available GPU memory (tune the size)
if use_dali:
LOGGER.log("DALI is activated, GPU memory fraction used for training is limited to", gpu_memory_fraction)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction)
config = tf.ConfigProto(gpu_options=gpu_options)
config.gpu_options.allow_growth = False
else:
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
config.log_device_placement = False
config.gpu_options.visible_device_list = str(gpu_id)
if hvd_utils.is_using_hvd():
config.gpu_options.visible_device_list = str(hvd.local_rank())
if use_xla:
LOGGER.log("XLA is activated - Experimental Feature")
if use_xla:
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
config.gpu_options.force_gpu_compatible = True # Force pinned memory
# Bug 2939231 - disable bn+relu fusion
from tensorflow.core.protobuf import rewriter_config_pb2
config.graph_options.rewrite_options.remapping = (rewriter_config_pb2.RewriterConfig.OFF)
if mode == 'train':
config.intra_op_parallelism_threads = 1 # Avoid pool of Eigen threads
@ -251,7 +258,9 @@ class Runner(object):
def _get_run_config(mode, model_dir, use_xla, use_dali, gpu_memory_fraction, gpu_id=0, seed=None):
if mode not in ["train", 'validation', 'benchmark', 'inference']:
raise ValueError("Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode)
raise ValueError(
"Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode
)
if seed is not None:
if hvd_utils.is_using_hvd():
@ -267,7 +276,9 @@ class Runner(object):
save_summary_steps=100 if mode in ['train', 'validation'] else 1e9, # disabled in benchmark mode
save_checkpoints_steps=None,
save_checkpoints_secs=None,
session_config=Runner._get_session_config(mode=mode, use_xla=use_xla, use_dali=use_dali, gpu_memory_fraction=gpu_memory_fraction, gpu_id=gpu_id),
session_config=Runner._get_session_config(
mode=mode, use_xla=use_xla, use_dali=use_dali, gpu_memory_fraction=gpu_memory_fraction, gpu_id=gpu_id
),
keep_checkpoint_max=5,
keep_checkpoint_every_n_hours=1e6, # disabled
log_step_count_steps=1e9,
@ -291,7 +302,9 @@ class Runner(object):
def _get_estimator(self, mode, run_params, use_xla, use_dali, gpu_memory_fraction, gpu_id=0):
if mode not in ["train", 'validation', 'benchmark', 'inference']:
raise ValueError("Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode)
raise ValueError(
"Unknown mode received: %s (allowed: 'train', 'validation', 'benchmark', 'inference')" % mode
)
run_config = Runner._get_run_config(
mode=mode,
@ -301,20 +314,17 @@ class Runner(object):
gpu_memory_fraction=gpu_memory_fraction,
gpu_id=gpu_id,
seed=self.run_hparams.seed
)
return tf.estimator.Estimator(
model_fn=self._model,
model_dir=self.run_hparams.model_dir,
config=run_config,
params=run_params
model_fn=self._model, model_dir=self.run_hparams.model_dir, config=run_config, params=run_params
)
def train(
self,
iter_unit,
num_iter,
run_iter,
batch_size,
warmup_steps=50,
weight_decay=1e-4,
@ -340,7 +350,6 @@ class Runner(object):
if use_static_loss_scaling:
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_LOSS_SCALING"] = "0"
else:
LOGGER.log("TF Loss Auto Scaling is activated")
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_LOSS_SCALING"] = "1"
else:
use_static_loss_scaling = False # Make sure it hasn't been set to True on FP32 training
@ -349,7 +358,7 @@ class Runner(object):
global_batch_size = batch_size * num_gpus
if self.run_hparams.data_dir is not None:
filenames,num_samples, num_steps, num_epochs, num_decay_steps = runner_utils.parse_tfrecords_dataset(
filenames, num_samples, num_steps, num_epochs, num_decay_steps = runner_utils.parse_tfrecords_dataset(
data_dir=self.run_hparams.data_dir,
mode="train",
iter_unit=iter_unit,
@ -366,51 +375,43 @@ class Runner(object):
num_decay_steps = num_steps
num_samples = num_steps * batch_size
if run_iter == -1:
run_iter = num_steps
else:
run_iter = steps_per_epoch * run_iter if iter_unit == "epoch" else run_iter
if self.run_hparams.data_idx_dir is not None:
idx_filenames = runner_utils.parse_dali_idx_dataset(
data_idx_dir=self.run_hparams.data_idx_dir,
mode="train"
data_idx_dir=self.run_hparams.data_idx_dir, mode="train"
)
training_hooks = []
if hvd.rank() == 0:
LOGGER.log('Starting Model Training...')
LOGGER.log("Training Epochs", num_epochs)
LOGGER.log("Total Steps", num_steps)
LOGGER.log("Steps per Epoch", steps_per_epoch)
LOGGER.log("Decay Steps", num_decay_steps)
LOGGER.log("Weight Decay Factor", weight_decay)
LOGGER.log("Init Learning Rate", lr_init)
LOGGER.log("Momentum", momentum)
LOGGER.log("Num GPUs", num_gpus)
LOGGER.log("Per-GPU Batch Size", batch_size)
print('Starting Model Training...')
print("Training Epochs", num_epochs)
print("Total Steps", num_steps)
print("Steps per Epoch", steps_per_epoch)
print("Decay Steps", num_decay_steps)
print("Weight Decay Factor", weight_decay)
print("Init Learning Rate", lr_init)
print("Momentum", momentum)
print("Num GPUs", num_gpus)
print("Per-GPU Batch Size", batch_size)
if is_benchmark:
benchmark_logging_hook = hooks.BenchmarkLoggingHook(
log_file_path=os.path.join(self.run_hparams.log_dir, "training_benchmark.json"),
global_batch_size=global_batch_size,
log_every=log_every_n_steps,
warmup_steps=warmup_steps
self.training_logging_hook = hooks.BenchmarkLoggingHook(
global_batch_size=global_batch_size, warmup_steps=warmup_steps
)
training_hooks.append(benchmark_logging_hook)
else:
training_logging_hook = hooks.TrainingLoggingHook(
log_file_path=os.path.join(self.run_hparams.log_dir, "training.json"),
self.training_logging_hook = hooks.TrainingLoggingHook(
global_batch_size=global_batch_size,
num_steps=num_steps,
num_samples=num_samples,
num_epochs=num_epochs,
log_every=log_every_n_steps
steps_per_epoch=steps_per_epoch
)
training_hooks.append(training_logging_hook)
training_hooks.append(self.training_logging_hook)
if hvd_utils.is_using_hvd():
bcast_hook = hvd.BroadcastGlobalVariablesHook(0)
@ -418,7 +419,6 @@ class Runner(object):
training_hooks.append(hooks.PrefillStagingAreasHook())
estimator_params = {
'batch_size': batch_size,
'steps_per_epoch': steps_per_epoch,
@ -445,11 +445,11 @@ class Runner(object):
)
def training_data_fn():
if self.run_hparams.use_dali and self.run_hparams.data_idx_dir is not None:
if hvd.rank() == 0:
LOGGER.log("Using DALI input... ")
print("Using DALI input... ")
return data_utils.get_dali_input_fn(
filenames=filenames,
idx_filenames=idx_filenames,
@ -461,7 +461,7 @@ class Runner(object):
num_threads=self.run_hparams.num_preprocessing_threads,
deterministic=False if self.run_hparams.seed is None else True
)
elif self.run_hparams.data_dir is not None:
return data_utils.get_tfrecords_input_fn(
@ -477,7 +477,7 @@ class Runner(object):
else:
if hvd.rank() == 0:
LOGGER.log("Using Synthetic Data ...")
print("Using Synthetic Data ...")
return data_utils.get_synth_input_fn(
batch_size=batch_size,
height=self.run_hparams.height,
@ -488,18 +488,19 @@ class Runner(object):
dtype=self.run_hparams.dtype,
)
try:
image_classifier.train(
input_fn=training_data_fn,
steps=num_steps,
steps=run_iter,
hooks=training_hooks,
)
except KeyboardInterrupt:
print("Keyboard interrupt")
if hvd.rank() == 0:
LOGGER.log('Ending Model Training ...')
print('Ending Model Training ...')
train_throughput = self.training_logging_hook.mean_throughput.value()
dllogger.log(data={'train_throughput': train_throughput}, step=tuple())
def evaluate(
self,
@ -522,13 +523,13 @@ class Runner(object):
raise RuntimeError('Multi-GPU inference is not supported')
estimator_params = {}
image_classifier = self._get_estimator(
mode='validation',
run_params=estimator_params,
use_xla=self.run_hparams.use_xla,
use_dali=self.run_hparams.use_dali,
gpu_memory_fraction=self.run_hparams.gpu_memory_fraction,
gpu_memory_fraction=self.run_hparams.gpu_memory_fraction,
gpu_id=self.run_hparams.gpu_id
)
@ -540,44 +541,35 @@ class Runner(object):
num_iter=num_iter,
global_batch_size=batch_size,
)
else:
num_epochs = 1
num_decay_steps = -1
num_steps = num_iter
if self.run_hparams.data_idx_dir is not None:
idx_filenames = runner_utils.parse_dali_idx_dataset(
data_idx_dir=self.run_hparams.data_idx_dir,
mode="validation"
data_idx_dir=self.run_hparams.data_idx_dir, mode="validation"
)
eval_hooks = []
if hvd.rank() == 0:
if is_benchmark:
benchmark_logging_hook = hooks.BenchmarkLoggingHook(
log_file_path=os.path.join(self.run_hparams.log_dir, "eval_benchmark.json"),
global_batch_size=batch_size,
log_every=log_every_n_steps,
warmup_steps=warmup_steps
)
eval_hooks.append(benchmark_logging_hook)
LOGGER.log('Starting Model Evaluation...')
LOGGER.log("Evaluation Epochs", num_epochs)
LOGGER.log("Evaluation Steps", num_steps)
LOGGER.log("Decay Steps", num_decay_steps)
LOGGER.log("Global Batch Size", batch_size)
eval_hooks = []
if hvd.rank() == 0:
self.eval_logging_hook = hooks.BenchmarkLoggingHook(global_batch_size=batch_size, warmup_steps=warmup_steps)
eval_hooks.append(self.eval_logging_hook)
print('Starting Model Evaluation...')
print("Evaluation Epochs", num_epochs)
print("Evaluation Steps", num_steps)
print("Decay Steps", num_decay_steps)
print("Global Batch Size", batch_size)
def evaluation_data_fn():
if self.run_hparams.use_dali and self.run_hparams.data_idx_dir is not None:
if hvd.rank() == 0:
LOGGER.log("Using DALI input... ")
print("Using DALI input... ")
return data_utils.get_dali_input_fn(
filenames=filenames,
idx_filenames=idx_filenames,
@ -589,22 +581,21 @@ class Runner(object):
num_threads=self.run_hparams.num_preprocessing_threads,
deterministic=False if self.run_hparams.seed is None else True
)
elif self.run_hparams.data_dir is not None:
return data_utils.get_tfrecords_input_fn(
filenames=filenames,
batch_size=batch_size,
height=self.run_hparams.height,
width=self.run_hparams.width,
training=False,
distort_color=self.run_hparams.distort_colors,
num_threads=self.run_hparams.num_preprocessing_threads,
deterministic=False if self.run_hparams.seed is None else True
)
filenames=filenames,
batch_size=batch_size,
height=self.run_hparams.height,
width=self.run_hparams.width,
training=False,
distort_color=self.run_hparams.distort_colors,
num_threads=self.run_hparams.num_preprocessing_threads,
deterministic=False if self.run_hparams.seed is None else True
)
else:
LOGGER.log("Using Synthetic Data ...\n")
print("Using Synthetic Data ...\n")
return data_utils.get_synth_input_fn(
batch_size=batch_size,
height=self.run_hparams.height,
@ -621,35 +612,50 @@ class Runner(object):
steps=num_steps,
hooks=eval_hooks,
)
LOGGER.log('Top-1 Accuracy: %.3f' % float(eval_results['top1_accuracy'] * 100))
LOGGER.log('Top-5 Accuracy: %.3f' % float(eval_results['top5_accuracy'] * 100))
#def get_serving_input_receiver_fn(batch_size, height, width, num_channels, data_format, dtype=tf.float32):
eval_throughput = self.eval_logging_hook.mean_throughput.value()
eval_latencies = np.array(self.eval_logging_hook.latencies) * 1000
eval_latencies_q = np.quantile(eval_latencies, q=[0.9, 0.95, 0.99])
eval_latencies_mean = np.mean(eval_latencies)
dllogger.log(
data={
'top1_accuracy': float(eval_results['top1_accuracy']),
'top5_accuracy': float(eval_results['top5_accuracy']),
'eval_throughput': eval_throughput,
'eval_latency_avg': eval_latencies_mean,
'eval_latency_p90': eval_latencies_q[0],
'eval_latency_p95': eval_latencies_q[1],
'eval_latency_p99': eval_latencies_q[2],
},
step=tuple()
)
if export_dir is not None:
LOGGER.log('Exporting to', export_dir)
dllogger.log(data={'export_dir': export_dir}, step=tuple())
input_receiver_fn = data_utils.get_serving_input_receiver_fn(
batch_size=None,
batch_size=None,
height=self.run_hparams.height,
width=self.run_hparams.width,
num_channels=self.run_hparams.n_channels,
data_format=self.run_hparams.input_format,
dtype=self.run_hparams.dtype)
dtype=self.run_hparams.dtype
)
image_classifier.export_savedmodel(export_dir, input_receiver_fn)
except KeyboardInterrupt:
print("Keyboard interrupt")
LOGGER.log('Ending Model Evaluation ...')
print('Model evaluation finished')
def predict(self, to_predict):
estimator_params = {}
if to_predict is not None:
filenames = runner_utils.parse_inference_input(to_predict)
image_classifier = self._get_estimator(
mode='inference',
run_params=estimator_params,
@ -657,28 +663,26 @@ class Runner(object):
use_dali=self.run_hparams.use_dali,
gpu_memory_fraction=self.run_hparams.gpu_memory_fraction
)
inference_hooks = []
def inference_data_fn():
return data_utils.get_inference_input_fn(
filenames=filenames,
height=self.run_hparams.height,
width=self.run_hparams.width,
num_threads=self.run_hparams.num_preprocessing_threads
)
filenames=filenames,
height=self.run_hparams.height,
width=self.run_hparams.width,
num_threads=self.run_hparams.num_preprocessing_threads
)
try:
inference_results = image_classifier.predict(
input_fn=inference_data_fn,
predict_keys=None,
hooks=inference_hooks,
yield_single_examples=True
input_fn=inference_data_fn, predict_keys=None, hooks=inference_hooks, yield_single_examples=True
)
for result in inference_results:
LOGGER.log(result['classes'], str(result['probabilities'][result['classes']]))
print(result['classes'], str(result['probabilities'][result['classes']]))
except KeyboardInterrupt:
print("Keyboard interrupt")
LOGGER.log('Ending Inference ...')
print('Ending Inference ...')

3
TensorFlow/Classification/RN50v1.5/runtime/runner_utils.py → TensorFlow/Classification/ConvNets/runtime/runner_utils.py
View file

@ -46,9 +46,6 @@ def list_filenames_in_dataset(data_dir, mode, count=True):
if mode not in ["train", 'validation']:
raise ValueError("Unknown mode received: %s" % mode)
if not os.path.exists(data_dir):
raise FileNotFoundError("The data directory: `%s` can't be found" % data_dir)
filename_pattern = os.path.join(data_dir, '%s-*' % mode)
file_list = sorted(tf.gfile.Glob(filename_pattern))

647
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/README.md Normal file
View file

@ -0,0 +1,647 @@
# SE-ResNext101-32x4d for TensorFlow
This repository provides a script and recipe to train the SE-ResNext101-32x4d model to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.
## Table Of Contents
* [Model overview](#model-overview)
* [Model architecture](#model-architecture)
* [Default configuration](#default-configuration)
* [Optimizer](#optimizer)
* [Data augmentation](#data-augmentation)
* [Feature support matrix](#feature-support-matrix)
* [Features](#features)
* [Mixed precision training](#mixed-precision-training)
* [Enabling mixed precision](#enabling-mixed-precision)
* [Setup](#setup)
* [Requirements](#requirements)
* [Quick Start Guide](#quick-start-guide)
* [Advanced](#advanced)
* [Scripts and sample code](#scripts-and-sample-code)
* [Parameters](#parameters)
* [The `main.py` script](#the-mainpy-script)
* [Inference process](#inference-process)
* [Performance](#performance)
* [Benchmarking](#benchmarking)
* [Training performance benchmark](#training-performance-benchmark)
* [Inference performance benchmark](#inference-performance-benchmark)
* [Results](#results)
* [Training accuracy results](#training-accuracy-results)
* [Training accuracy: NVIDIA DGX-1 (8x V100 16G)](#training-accuracy-nvidia-dgx-1-8x-v100-16g)
* [Training performance results](#training-performance-results)
* [Training performance: NVIDIA DGX-1 (8x V100 16G)](#training-performance-nvidia-dgx-1-8x-v100-16g)
* [Training performance: NVIDIA DGX-2 (16x V100 32G)](#training-performance-nvidia-dgx-2-16x-v100-32g)
* [Training time for 90 Epochs](#training-time-for-90-epochs)
* [Training time: NVIDIA DGX-1 (8x V100 16G)](#training-time-nvidia-dgx-1-8x-v100-16g)
* [Training time: NVIDIA DGX-2 (16x V100 32G)](#training-time-nvidia-dgx-2-16x-v100-32g)
* [Inference performance results](#inference-performance-results)
* [Inference performance: NVIDIA DGX-1 (1x V100 16G)](#inference-performance-nvidia-dgx-1-1x-v100-16g)
* [Inference performance: NVIDIA DGX-2 (1x V100 32G)](#inference-performance-nvidia-dgx-2-1x-v100-32g)
* [Inference performance: NVIDIA T4 (1x T4 16G)](#inference-performance-nvidia-t4-1x-t4-16g)
* [Release notes](#release-notes)
* [Changelog](#changelog)
* [Known issues](#known-issues)
## Model overview
The SE-ResNeXt101-32x4d is a [ResNeXt101-32x4d](https://arxiv.org/pdf/1611.05431.pdf)
model with added Squeeze-and-Excitation module introduced in the [Squeeze-and-Excitation Networks](https://arxiv.org/pdf/1709.01507.pdf) paper.
The following performance optimizations were implemented in this model:
* JIT graph compilation with [XLA](https://www.tensorflow.org/xla)
* Multi-GPU training with [Horovod](https://github.com/horovod/horovod)
* Automated mixed precision [AMP](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html)
This model is trained with mixed precision using Tensor Cores on NVIDIA Volta and Turing GPUs. Therefore, researchers can get results 2x faster than training without Tensor Cores, while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.
### Model architecture
Here is a diagram of the Squeeze and Excitation module architecture for ResNet-type models:
![SEArch](./imgs/SEArch.png)
_Image source: [Squeeze-and-Excitation Networks](https://arxiv.org/pdf/1709.01507.pdf)_
This image shows the architecture of the SE block and where it is placed in the ResNet bottleneck block.
### Default configuration
The following sections highlight the default configuration for the SE-ResNext101-32x4d model.
#### Optimizer
This model uses the SGD optimizer with the following hyperparameters:
* Momentum (0.875).
* Learning rate (LR) = 0.256 for 256 batch size, for other batch sizes we linearly scale the learning
rate.
* Learning rate schedule - we use cosine LR schedule.
* For bigger batch sizes (512 and up) we use linear warmup of the learning rate.
during the first 5 epochs according to [Training ImageNet in 1 hour](https://arxiv.org/abs/1706.02677).
* Weight decay: 6.103515625e-05 (1/16384).
* We do not apply Weight decay on batch norm trainable parameters (gamma/bias).
* Label Smoothing: 0.1.
* We train for:
* 90 Epochs -> 90 epochs is a standard for ImageNet networks.
* 250 Epochs -> best possible accuracy.
* For 250 epoch training we also use [MixUp regularization](https://arxiv.org/pdf/1710.09412.pdf).
#### Data Augmentation
This model uses the following data augmentation:
* For training:
* Normalization.
* Random resized crop to 224x224.
* Scale from 8% to 100%.
* Aspect ratio from 3/4 to 4/3.
* Random horizontal flip.
* For inference:
* Normalization.
* Scale to 256x256.
* Center crop to 224x224.
### Feature support matrix
The following features are supported by this model.
| Feature | SE-ResNext101-32x4d Tensorflow |
|-----------------------|--------------------------
|Multi-GPU training with [Horovod](https://github.com/horovod/horovod) | Yes |
|[NVIDIA DALI](https://docs.nvidia.com/deeplearning/dali/release-notes/index.html) | Yes |
|Automatic mixed precision (AMP) | Yes |
#### Features
Multi-GPU training with Horovod - Our model uses Horovod to implement efficient multi-GPU training with NCCL.
For details, refer to the example sources in this repository or the [TensorFlow tutorial](https://github.com/horovod/horovod/#usage).
NVIDIA DALI - DALI is a library accelerating data preparation pipeline. To accelerate your input pipeline, you only need to define your data loader
with the DALI library. For details, refer to the example sources in this repository or the [DALI documentation](https://docs.nvidia.com/deeplearning/dali/index.html).
Automatic mixed precision (AMP) - Computation graph can be modified by TensorFlow on runtime to support mixed precision training.
Detailed explanation of mixed precision can be found in the next section.
### Mixed precision training
Mixed precision is the combined use of different numerical precisions in a computational method. [Mixed precision](https://arxiv.org/abs/1710.03740) training offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of [Tensor Cores](https://developer.nvidia.com/tensor-cores) in the Volta and Turing architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using [mixed precision training](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html) previously required two steps:
1. Porting the model to use the FP16 data type where appropriate.
2. Manually adding loss scaling to preserve small gradient values.
This can now be achieved using Automatic Mixed Precision (AMP) for TensorFlow to enable the full [mixed precision methodology](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html#tensorflow) in your existing TensorFlow model code. AMP enables mixed precision training on Volta and Turing GPUs automatically. The TensorFlow framework code makes all necessary model changes internally.
In TF-AMP, the computational graph is optimized to use as few casts as necessary and maximize the use of FP16, and the loss scaling is automatically applied inside of supported optimizers. AMP can be configured to work with the existing tf.contrib loss scaling manager by disabling the AMP scaling with a single environment variable to perform only the automatic mixed-precision optimization. It accomplishes this by automatically rewriting all computation graphs with the necessary operations to enable mixed precision training and automatic loss scaling.
For information about:
* How to train using mixed precision, see the [Mixed Precision Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed Precision](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html) documentation.
* How to access and enable AMP for TensorFlow, see [Using TF-AMP](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-user-guide/index.html#tfamp) from the TensorFlow User Guide.
* Techniques used for mixed precision training, see the [Mixed-Precision Training of Deep Neural Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) blog.
#### Enabling mixed precision
Mixed precision is enabled in TensorFlow by using the Automatic Mixed Precision (TF-AMP) extension which casts variables to half-precision upon retrieval, while storing variables in single-precision format. Furthermore, to preserve small gradient magnitudes in backpropagation, a [loss scaling](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html#lossscaling) step must be included when applying gradients. In TensorFlow, loss scaling can be applied statically by using simple multiplication of loss by a constant value or automatically, by TF-AMP. Automatic mixed precision makes all the adjustments internally in TensorFlow, providing two benefits over manual operations. First, programmers need not modify network model code, reducing development and maintenance effort. Second, using AMP maintains forward and backward compatibility with all the APIs for defining and running TensorFlow models.
To enable mixed precision, you can simply add the values to the environmental variables inside your training script:
- Enable TF-AMP graph rewrite:
```
os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "1"
```
- Enable Automated Mixed Precision:
```
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
```
## Setup
The following section lists the requirements that you need to meet in order to use the SE-ResNext101-32x4d model.
### Requirements
This repository contains Dockerfile which extends the TensorFlow NGC container and encapsulates all dependencies. Aside from these dependencies, ensure you have the following software:
* [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)
* [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
* [NVIDIA Volta](https://www.nvidia.com/en-us/data-center/volta-gpu-architecture/) or [Turing](https://www.nvidia.com/en-us/geforce/turing/) based GPU
For more information about how to get started with NGC containers, see the
following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:
* [Getting Started Using NVIDIA GPU Cloud](https://docs.nvidia.com/ngc/ngc-getting-started-guide/index.html),
* [Accessing And Pulling From The NGC container registry](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#accessing_registry),
* [Running TensorFlow](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/running.html#running).
For those unable to use the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) to set up the required environment or create your own container, see the versioned [NVIDIA Container Support Matrix](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).
## Quick Start Guide
To train your model using mixed precision with Tensor Cores or FP32, perform the following steps using the default parameters of the SE-ResNext101-32x4d model on the [ImageNet](http://www.image-net.org/) dataset. For the specifics concerning training and inference, see the [Advanced](#advanced) section.
1. Clone the repository.
```
git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/Classification/RN50v1.5
```
2. Download and preprocess the dataset.
The SE-ResNext101-32x4d script operates on ImageNet 1k, a widely popular image classification dataset from the ILSVRC challenge.
To download and preprocess the dataset, use the [Generate ImageNet for TensorFlow](https://github.com/tensorflow/models/blob/master/research/inception/inception/data/download_and_preprocess_imagenet.sh) script. The dataset will be downloaded to a directory specified as the first parameter of the script.
3. Build the SE-ResNext101-32x4d TensorFlow NGC container.
```bash
docker build . -t nvidia_rn50
```
4. Start an interactive session in the NGC container to run training/inference.
After you build the container image, you can start an interactive CLI session with
```bash
nvidia-docker run --rm -it -v <path to imagenet>:/data/tfrecords --ipc=host nvidia_rn50
```
5. (Optional) Create index files to use DALI.
To allow proper sharding in a multi-GPU environment, DALI has to create index files for the dataset. To create index files, run inside the container:
```bash
bash ./utils/dali_index.sh /data/tfrecords <index file store location>
```
Index files can be created once and then reused. It is highly recommended to save them into a persistent location.
6. Start training.
To run training for a standard configuration (as described in [Default
configuration](#default-configuration), DGX1V, DGX2V, single GPU, FP16, FP32, 90, and 250 epochs), run
one of the scripts in the `se-resnext101-32x4d/training` directory. Ensure ImageNet is mounted in the
`/data/tfrecords` directory.
For example, to train on DGX-1 for 90 epochs using AMP, run:
`bash ./se-resnext101-32x4d/training/AMP/DGX1_SE-RNxt101-32x4d_AMP_90E.sh`
Additionally, features like DALI data preprocessing or TensorFlow XLA can be enabled with
environmental variables when running those scripts:
`USE_XLA=1 USE_DALI=1 bash ./se-resnext101-32x4d/training/AMP/DGX1_SE-RNxt101-32x4d_AMP_90E.sh`
To store results in a specific location, add a location as a first argument:
`bash ./resnext101-32x4d/training/AMP/DGX1_RNxt101-32x4d_AMP_90E.sh <location to store>`
7. Start validation/evaluation.
To evaluate the validation dataset located in `/data/tfrecords`, run `main.py` with
`--mode=evaluate`. For example:
`python main.py --arch=se-resnext101-32x4d --mode=evaluate --data_dir=/data/tfrecords --batch_size <batch size> --model_dir
<model location> --result_dir <output location> [--use_xla] [--use_tf_amp]`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during evaluation.
## Advanced
The following sections provide greater details of the dataset, running training and inference, and the training results.
### Scripts and sample code
In the root directory, the most important files are:
- `main.py`: the script that controls the logic of training and validation of the ResNet-like models
- `Dockerfile`: Instructions for Docker to build a container with the basic set of dependencies to run ResNet like models for image classification
- `requirements.txt`: a set of extra Python requirements for running ResNet-like models
The `model/` directory contains the following modules used to define ResNet family models:
- `resnet.py`: the definition of ResNet, ResNext, and SE-ResNext model
- `blocks/conv2d_block.py`: the definition of 2D convolution block
- `blocks/resnet_bottleneck_block.py`: the definition of ResNet-like bottleneck block
- `layers/*.py`: definitions of specific layers used in the ResNet-like model
The `utils/` directory contains the following utility modules:
- `cmdline_helper.py`: helper module for command line processing
- `data_utils.py`: module defining input data pipelines
- `dali_utils.py`: helper module for DALI
- `hvd_utils.py`: helper module for Horovod
- `image_processing.py`: image processing and data augmentation functions
- `learning_rate.py`: definition of used learning rate schedule
- `optimizers.py`: definition of used custom optimizers
- `hooks/*.py`: definitions of specific hooks allowing logging of training and inference process
The `runtime/` directory contains the following module that define the mechanics of the training process:
- `runner.py`: module encapsulating the training, inference and evaluation
### Parameters
#### The `main.py` script
The script for training and evaluating the ResNext101-32x4d model has a variety of parameters that control these processes.
```
usage: main.py [-h]
[--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}]
[--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}]
[--data_dir DATA_DIR] [--data_idx_dir DATA_IDX_DIR]
[--export_dir EXPORT_DIR] [--to_predict TO_PREDICT]
[--batch_size BATCH_SIZE] [--num_iter NUM_ITER]
[--iter_unit {epoch,batch}] [--warmup_steps WARMUP_STEPS]
[--model_dir MODEL_DIR] [--results_dir RESULTS_DIR]
[--log_filename LOG_FILENAME] [--display_every DISPLAY_EVERY]
[--lr_init LR_INIT] [--lr_warmup_epochs LR_WARMUP_EPOCHS]
[--weight_decay WEIGHT_DECAY] [--weight_init {fan_in,fan_out}]
[--momentum MOMENTUM] [--loss_scale LOSS_SCALE]
[--label_smoothing LABEL_SMOOTHING] [--mixup MIXUP]
[--use_static_loss_scaling | --nouse_static_loss_scaling]
[--use_xla | --nouse_xla] [--use_dali | --nouse_dali]
[--use_tf_amp | --nouse_tf_amp]
[--use_cosine_lr | --nouse_cosine_lr] [--seed SEED]
[--gpu_memory_fraction GPU_MEMORY_FRACTION] [--gpu_id GPU_ID]
JoC-RN50v1.5-TF
optional arguments:
-h, --help Show this help message and exit
--arch {resnet50,resnext101-32x4d,se-resnext101-32x4d}
Architecture of model to run (to run se-resnext-32x4d set
--arch=se-rensext101-32x4d)
--mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}
The execution mode of the script.
--data_dir DATA_DIR Path to dataset in TFRecord format. Files should be
named 'train-*' and 'validation-*'.
--data_idx_dir DATA_IDX_DIR
Path to index files for DALI. Files should be named
'train-*' and 'validation-*'.
--export_dir EXPORT_DIR
Directory in which to write exported SavedModel.
--to_predict TO_PREDICT
Path to file or directory of files to run prediction
on.
--batch_size BATCH_SIZE
Size of each minibatch per GPU.
--num_iter NUM_ITER Number of iterations to run.
--iter_unit {epoch,batch}
Unit of iterations.
--warmup_steps WARMUP_STEPS
Number of steps considered as warmup and not taken
into account for performance measurements.
--model_dir MODEL_DIR
Directory in which to write the model. If undefined,
results directory will be used.
--results_dir RESULTS_DIR
Directory in which to write training logs, summaries
and checkpoints.
--log_filename LOG_FILENAME
Name of the JSON file to which write the training log
--display_every DISPLAY_EVERY
How often (in batches) to print out running
information.
--lr_init LR_INIT Initial value for the learning rate.
--lr_warmup_epochs LR_WARMUP_EPOCHS
Number of warmup epochs for the learning rate schedule.
--weight_decay WEIGHT_DECAY
Weight Decay scale factor.
--weight_init {fan_in,fan_out}
Model weight initialization method.
--momentum MOMENTUM SGD momentum value for the momentum optimizer.
--loss_scale LOSS_SCALE
Loss scale for FP16 training and fast math FP32.
--label_smoothing LABEL_SMOOTHING
The value of label smoothing.
--mixup MIXUP The alpha parameter for mixup (if 0 then mixup is not
applied).
--use_static_loss_scaling
Use static loss scaling in FP16 or FP32 AMP.
--nouse_static_loss_scaling
--use_xla Enable XLA (Accelerated Linear Algebra) computation
for improved performance.
--nouse_xla
--use_dali Enable DALI data input.
--nouse_dali
--use_tf_amp Enable AMP to speedup FP32
computation using Tensor Cores.
--nouse_tf_amp
--use_cosine_lr Use cosine learning rate schedule.
--nouse_cosine_lr
--seed SEED Random seed.
--gpu_memory_fraction GPU_MEMORY_FRACTION
Limit memory fraction used by the training script for DALI
--gpu_id GPU_ID Specify the ID of the target GPU on a multi-device platform.
Effective only for single-GPU mode.
```
### Inference process
To run inference on a single example with a checkpoint and a model script, use:
`python main.py --arch=se-resnext101-32x4d --mode predict --model_dir <path to model> --to_predict <path to image> --results_dir <path to results>`
The optional `--use_xla` and `--use_tf_amp` flags control XLA and AMP during inference.
## Performance
### Benchmarking
The following section shows how to run benchmarks measuring the model performance in training and inference modes.
#### Training performance benchmark
To benchmark the training performance on a specific batch size, run:
* For 1 GPU
* FP32
`python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --use_tf_amp --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* For multiple GPUs
* FP32
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --use_tf_amp --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
Each of these scripts runs 200 warm-up iterations and measures the first epoch.
To control warmup and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags. Features like XLA or DALI can be controlled
with `--use_xla` and `--use_dali` flags.
Suggested batch sizes for training are 96 for mixed precision training and 64 for single precision training per single V100 16 GB.
#### Inference performance benchmark
To benchmark the inference performance on a specific batch size, run:
* FP32
`python ./main.py --arch=se-resnext101-32x4d --mode=inference_benchmark --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
* FP16
`python ./main.py --arch=se-resnext101-32x4d --mode=inference_benchmark --use_tf_amp --warmup_steps 20 --num_iter 100 --iter_unit batch --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>`
By default, each of these scripts runs 20 warm-up iterations and measures the next 80 iterations.
To control warm-up and benchmark length, use the `--warmup_steps`, `--num_iter` and `--iter_unit` flags.
The benchmark can be automated with the `inference_benchmark.sh` script provided in `se-resnext101-32x4d`, by simply running:
`bash ./se-resnext101-32x4d/inference_benchmark.sh <data dir> <data idx dir>`
The `<data dir>` parameter refers to the input data directory (by default `/data/tfrecords` inside the container).
By default, the benchmark tests the following configurations: **FP32**, **AMP**, **AMP + XLA** with different batch sizes.
When the optional directory with the DALI index files `<data idx dir>` is specified, the benchmark executes an additional **DALI + AMP + XLA** configuration.
### Results
The following sections provide details on how we achieved our performance and accuracy in training and inference.
#### Training accuracy results
##### Training accuracy: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the `/resnext101-32x4d/training/{PRECISION}/DGX1_RNxt101-32x4d_{PRECISION}_{EPOCHS}E.sh`
training script in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
NGC container on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| Epochs | Batch Size / GPU | Accuracy - FP32 | Accuracy - mixed precision |
|--------|------------------|-----------------|----------------------------|
| 90 | 64 (FP32) / 96 (AMP) | 79.69 | 79.81 |
| 250 | 64 (FP32) / 96 (AMP) | 80.87 | 80.84 |
**Example training loss plot**
![TrainingLoss](./imgs/train_loss.png)
#### Training performance results
##### Training performance: NVIDIA DGX-1 (8x V100 16G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (8x V100 16G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|-----------------------|---------------|-----------|-------|
| 1 | 64 (FP32) / 96 (AMP) | 126.44 img/s | 285.99 img/s | 2.26x | 1.00x | 1.00x |
| 8 | 64 (FP32) / 96 (AMP) | 921.38 img/s | 2168.40 img/s | 2.35x | 7.28x | 7.58x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|------------|---------------|---------------------|-----------|
| 1 | 128 | 285.99 img/s |453.39 img/s |1.58x |
| 8 | 128 | 2168.40 img/s |3297.39 img/s |1.52x |
##### Training performance: NVIDIA DGX-2 (16x V100 32G)
Our results were obtained by running the steps from [Training performance benchmark](#training-performance-benchmark) in the
[TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (16x V100 32G) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.
| GPUs | Batch Size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 - mixed precision) | Weak scaling - FP32 | Weak scaling - mixed precision |
|----|---------------|---------------|-------------------------|-------|--------|--------|
| 1 | 64 (FP32) / 96 (AMP) | 128.08 img/s | 309.34 img/s | 2.39x | 1.00x | 1.00x |
| 16 | 64 (FP32) / 96 (AMP) | 1759.44 img/s| 4210.85 img/s | 2.51x | 13.73x | 13.61x |
**XLA Enabled**
| GPUs | Batch Size / GPU | Throughput - mixed precision | Throughput - mixed precision + XLA | Throughput speedup (mixed precision - XLA) |
|----|-----|----------|---------------------|-----------|
| 1 | 96 | 309.34 img/s |520.10 img/s |1.68x |
| 16 | 96 | 4210.85 img/s |6835.66 img/s |1.62x |
#### Training Time for 90 Epochs
##### Training time: NVIDIA DGX-1 (8x V100 16G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-1-8x-v100-16g)
on NVIDIA DGX-1 with (8x V100 16G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|---|--------|---------|---------|
| 1 | ~71h | ~112h | ~253h |
| 8 | ~10h | ~15h | ~35h |
##### Training time: NVIDIA DGX-2 (16x V100 32G)
Our results were estimated based on the [training performance results](#training-performance-nvidia-dgx-2-16x-v100-32g)
on NVIDIA DGX-2 with (16x V100 32G) GPUs.
| GPUs | Time to train - mixed precision + XLA | Time to train - mixed precision | Time to train - FP32 |
|----|-------|--------|-------|
| 1 | ~61h | ~103h | ~247h |
| 16 | ~4.7h | ~7.5h | ~19h |
#### Inference performance results
##### Inference performance: NVIDIA DGX-1 (1x V100 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-1 with (1x V100 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |54.71 img/s |18.32 ms |18.73 ms |18.89 ms |19.95 ms |
|2 |98.36 img/s |20.37 ms |20.69 ms |20.75 ms |21.03 ms |
|4 |150.60 img/s |26.56 ms |26.83 ms |26.95 ms |27.46 ms |
|8 |235.17 img/s |34.02 ms |34.40 ms |34.57 ms |35.37 ms |
|16 |330.33 img/s |48.43 ms |48.91 ms |49.22 ms |49.79 ms |
|32 |393.96 img/s |81.22 ms |81.72 ms |81.99 ms |82.49 ms |
|64 |446.13 img/s |143.54 ms |144.37 ms |144.74 ms |145.93 ms |
|128 |490.61 img/s |260.89 ms |261.56 ms |261.76 ms |262.71 ms |
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |92.56 img/s |10.88 ms |11.04 ms |11.66 ms |14.34 ms |
|2 |180.61 img/s |11.11 ms |11.35 ms |11.47 ms |11.79 ms |
|4 |354.41 img/s |11.35 ms |11.90 ms |12.13 ms |13.28 ms |
|8 |547.79 img/s |14.63 ms |15.53 ms |15.99 ms |16.38 ms |
|16 |772.41 img/s |20.80 ms |21.76 ms |22.02 ms |23.02 ms |
|32 |965.89 img/s |33.15 ms |33.82 ms |34.24 ms |35.15 ms |
|64 |1086.99 img/s |59.01 ms |59.42 ms |59.56 ms |60.35 ms |
|128 |1162.59 img/s |110.36 ms |110.41 ms |110.64 ms |111.18 ms |
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |84.98 img/s |11.81 ms |12.00 ms |12.08 ms |12.85 ms |
|2 |150.35 img/s |13.37 ms |14.15 ms |14.56 ms |15.11 ms |
|4 |288.89 img/s |13.90 ms |14.37 ms |14.56 ms |16.19 ms |
|8 |526.94 img/s |15.19 ms |15.61 ms |15.85 ms |17.91 ms |
|16 |818.86 img/s |19.63 ms |19.85 ms |19.97 ms |20.70 ms |
|32 |1134.72 img/s |28.20 ms |28.60 ms |28.82 ms |30.03 ms |
|64 |1359.55 img/s |47.22 ms |47.51 ms |47.84 ms |48.96 ms |
|128 |1515.12 img/s |84.49 ms |85.51 ms |85.82 ms |86.89 ms |
##### Inference performance: NVIDIA DGX-2 (1x V100 32G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA DGX-2 with (1x V100 32G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |59.41 img/s |16.86 ms |17.02 ms |17.10 ms |17.47 ms |
|2 |92.74 img/s |21.59 ms |21.88 ms |22.01 ms |22.65 ms |
|4 |141.53 img/s |28.26 ms |28.45 ms |28.55 ms |28.77 ms |
|8 |228.80 img/s |34.96 ms |35.18 ms |35.38 ms |35.72 ms |
|16 |324.11 img/s |49.36 ms |49.61 ms |49.76 ms |50.17 ms |
|32 |397.66 img/s |80.47 ms |80.69 ms |80.82 ms |82.15 ms |
|64 |468.28 img/s |136.67 ms |137.03 ms |137.18 ms |138.11 ms|
|128 |514.25 img/s |248.91 ms |250.42 ms |251.89 ms |253.55 ms|
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |103.55 img/s |9.70 ms |9.94 ms |10.03 ms |10.26 ms |
|2 |227.06 img/s |8.86 ms |9.07 ms |9.19 ms |9.78 ms |
|4 |380.22 img/s |10.53 ms |10.97 ms |11.11 ms |11.64 ms |
|8 |579.59 img/s |13.82 ms |14.07 ms |14.29 ms |14.99 ms |
|16 |833.08 img/s |19.20 ms |19.33 ms |19.37 ms |19.68 ms |
|32 |990.96 img/s |32.30 ms |32.53 ms |32.70 ms |33.45 ms |
|64 |1114.78 img/s |57.41 ms |57.61 ms |57.86 ms |58.78 ms |
|128 |1203.04 img/s |106.40 ms |106.54 ms |106.62 ms |107.93 ms|
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |92.52 img/s |10.85 ms |10.95 ms |11.12 ms |11.99 ms |
|2 |177.54 img/s |11.32 ms |11.59 ms |11.66 ms |13.16 ms |
|4 |322.44 img/s |12.44 ms |12.63 ms |12.71 ms |12.98 ms |
|8 |548.33 img/s |14.59 ms |14.90 ms |15.58 ms |15.95 ms |
|16 |818.34 img/s |19.55 ms |19.75 ms |19.91 ms |20.30 ms |
|32 |1178.12 img/s |27.16 ms |27.41 ms |27.53 ms |28.24 ms|
|64 |1397.25 img/s |45.96 ms |46.03 ms |46.19 ms |47.01 ms|
|128 |1613.78 img/s |79.32 ms |80.02 ms |80.46 ms |81.71 ms|
##### Inference performance: NVIDIA T4 (1x T4 16G)
Our results were obtained by running the `inference_benchmark.sh` inferencing benchmarking script
in the [TensorFlow 20.03-tf1-py3 NGC container](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container
on NVIDIA T4 with (1x T4 16G) GPU.
**FP32 Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |48.92 img/s |20.47 ms |20.75 ms |20.88 ms |21.23 ms |
|2 |82.59 img/s |24.24 ms |24.42 ms |24.50 ms |24.64 ms |
|4 |111.08 img/s |36.03 ms |36.37 ms |36.45 ms |36.67 ms |
|8 |131.84 img/s |60.68 ms |61.36 ms |61.62 ms |62.21 ms |
|16 |144.27 img/s |110.90 ms |112.04 ms |112.29 ms |112.80 ms |
|32 |156.59 img/s |204.35 ms |206.12 ms |206.78 ms |208.08 ms |
|64 |162.58 img/s |393.66 ms |396.45 ms |396.94 ms |397.52 ms |
|128 |162.41 img/s |788.13 ms |790.86 ms |791.47 ms |792.43 ms |
**Mixed Precision Inference Latency**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |79.57 img/s |12.74 ms |13.31 ms |13.37 ms |13.85 ms |
|2 |190.25 img/s |10.55 ms |10.63 ms |10.67 ms |10.77 ms |
|4 |263.72 img/s |15.16 ms |15.37 ms |15.44 ms |15.58 ms |
|8 |312.07 img/s |25.64 ms |26.14 ms |26.25 ms |26.98 ms |
|16 |347.43 img/s |46.05 ms |46.60 ms |46.82 ms |47.09 ms |
|32 |360.20 img/s |88.84 ms |89.44 ms |89.60 ms |90.20 ms |
|64 |367.23 img/s |174.28 ms |175.81 ms |176.13 ms |176.74 ms |
|128 |362.43 img/s |353.17 ms |354.91 ms |355.52 ms |356.07 ms |
**Mixed Precision Inference Latency + XLA**
|**Batch Size**|**Avg throughput**|**Avg latency**|**90% Latency**|**95% Latency**|**99% Latency**|
|--------------|------------------|---------------|---------------|---------------|---------------|
|1 |89.30 img/s |11.24 ms |11.31 ms |11.38 ms |11.75 ms |
|2 |152.93 img/s |13.11 ms |13.20 ms |13.25 ms |13.58 ms |
|4 |254.87 img/s |15.69 ms |15.84 ms |15.87 ms |15.95 ms |
|8 |356.48 img/s |22.44 ms |22.79 ms |22.86 ms |23.08 ms |
|16 |442.24 img/s |36.18 ms |36.63 ms |36.76 ms |36.76 ms |
|32 |471.28 img/s |67.90 ms |68.62 ms |68.80 ms |69.14 ms |
|64 |483.18 img/s |132.46 ms |133.72 ms |134.08 ms |134.88 ms |
|128 |501.38 img/s |255.31 ms |258.46 ms |259.19 ms |260.17 ms|
## Release notes
### Changelog
June 2020
- Initial release
### Known issues
There are no known issues with this model.

BIN
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/imgs/SEArch.png Normal file
View file

Binary file not shown.
Side by side

After

Width:  |  Height:  |  Size: 76 KiB

BIN
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/imgs/train_loss.png Normal file
View file

Binary file not shown.
Side by side

After

Width:  |  Height:  |  Size: 52 KiB

33
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/inference_benchmark.sh Normal file
View file

@ -0,0 +1,33 @@
#!/bin/bash
DATA_DIR=${1:-"/data/tfrecords"}
DALI_DIR=${2}
BATCH_SIZE_TO_TEST="1 2 4 8 16 32 64 128"
INFERENCE_BENCHMARK=$(mktemp /tmp/inference-benchmark.XXXXXX)
function test_configuration() {
echo "Testing configuration: $1" | tee -a $INFERENCE_BENCHMARK
for BATCH in $BATCH_SIZE_TO_TEST; do
python ./main.py --arch=se-resnext101-32x4d --mode=inference_benchmark --warmup_steps 50 --num_iter 400 --iter_unit batch \
--batch_size $BATCH --data_dir=$DATA_DIR --results_dir=/tmp/results $2 | tail -n2 | head -n1 | sed \
's/^DLL \([0-9]*-\)*[0-9]* \([0-9]*:\)*[0-9]*.[0-9]* - ()/Results for BS='$BATCH'/' | tee -a $INFERENCE_BENCHMARK
if [ ! $? -eq 0 ]; then
echo "Failed test on batch size $BATCH_SIZE"
exit 1
fi
done
}
test_configuration "FP32 nodali noxla"
test_configuration "FP16 nodali noxla" "--use_tf_amp"
test_configuration "FP16 nodali xla" "--use_tf_amp --use_xla"
if [ ! -z $DALI_DIR ]; then
test_configuration "FP16 dali xla" "--use_tf_amp --use_xla --use_dali --data_idx_dir ${DALI_DIR}"
fi
cat $INFERENCE_BENCHMARK
rm $INFERENCE_BENCHMARK

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/DGX1_SE-RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/DGX1_SE-RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 90 128 amp

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/DGX2_SE-RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/DGX2_SE-RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 90 128 amp

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/GPU1_SE-RNxt101-32x4d_AMP_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 250 128 amp --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/AMP/GPU1_SE-RNxt101-32x4d_AMP_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 90 128 amp

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/DGX1_SE-RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/DGX1_SE-RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
8 90 64 fp32

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/DGX2_SE-RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/DGX2_SE-RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
16 90 64 fp32

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/GPU1_SE-RNxt101-32x4d_FP32_250E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 250 64 fp32 --mixup=0.2

6
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/FP32/GPU1_SE-RNxt101-32x4d_FP32_90E.sh Normal file
View file

@ -0,0 +1,6 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
WORKSPACE=${2:-"/workspace/rn50v15_tf"}
DATA_DIR=${3:-"/data"}
bash ${WORKSPACE}/se-resnext101-32x4d/training/GENERIC.sh ${RESULT_DIR} ${DATA_DIR} \
1 90 64 fp32

39
TensorFlow/Classification/ConvNets/se-resnext101-32x4d/training/GENERIC.sh Normal file
View file

@ -0,0 +1,39 @@
RESULT_DIR=${1:-"/workspace/rn50v15_tf/results"}
DATA_DIR=${2:-"/data"}
GPU_COUNT=${3:-8}
ITER_COUNT=${4:-50}
BATCH_SIZE=${5:-128}
PRECISION=${6:-"fp32"}
OTHER=${@:7}
if [[ ! -z "${BIND_TO_SOCKET}" ]]; then
BIND_TO_SOCKET="--bind-to socket"
fi
if [[ ! -z "${USE_DALI}" ]]; then
USE_DALI="--use_dali --data_idx_dir=${DATA_DIR}/dali_idx"
fi
if [[ ! -z "${USE_XLA}" ]]; then
USE_XLA="--use_xla"
fi
CMD=""
case $PRECISION in
"fp32") CMD+="--precision=fp32";;
"fp16") CMD+="--precision=fp16 --use_static_loss_scaling --loss_scale=128";;
"amp") CMD+="--precision=fp32 --use_tf_amp --use_static_loss_scaling --loss_scale=128";;
esac
CMD="--arch=se-resnext101-32x4d --mode=train_and_evaluate --iter_unit=epoch --num_iter=${ITER_COUNT} \
--batch_size=${BATCH_SIZE} --warmup_steps=100 --use_cosine --label_smoothing 0.1 \
--lr_init=0.256 --lr_warmup_epochs=8 --momentum=0.875 --weight_decay=6.103515625e-05 \
${CMD} --data_dir=${DATA_DIR}/tfrecords ${USE_DALI} ${USE_XLA} \
--results_dir=${RESULT_DIR} --weight_init=fan_in ${OTHER}"
if [[ ${GPU_COUNT} -eq 1 ]]; then
python3 main.py ${CMD}
else
mpiexec --allow-run-as-root ${BIND_TO_SOCKET} -np ${GPU_COUNT} python3 main.py ${CMD}
fi

2
TensorFlow/Classification/RN50v1.5/utils/__init__.py → TensorFlow/Classification/ConvNets/utils/__init__.py
View file

@ -27,4 +27,4 @@ from utils import image_processing
from utils import learning_rate
from utils import dali_utils
from utils import dali_utils

146
TensorFlow/Classification/RN50v1.5/utils/cmdline_helper.py → TensorFlow/Classification/ConvNets/utils/cmdline_helper.py
View file

@ -30,10 +30,19 @@ def _add_bool_argument(parser, name=None, default=False, required=False, help=No
feature_parser.set_defaults(name=default)
def parse_cmdline():
def parse_cmdline(available_arch):
p = argparse.ArgumentParser(description="JoC-RN50v1.5-TF")
p.add_argument(
'--arch',
choices=available_arch,
type=str,
default='resnet50',
required=False,
help="""Architecture of model to run"""
)
p.add_argument(
'--mode',
choices=['train', 'train_and_evaluate', 'evaluate', 'predict', 'training_benchmark', 'inference_benchmark'],
@ -58,43 +67,35 @@ def parse_cmdline():
type=str,
help="Path to index files for DALI. Files should be named 'train-*' and 'validation-*'."
)
p.add_argument(
'--export_dir',
required=False,
default=None,
type=str,
help="Directory in which to write exported SavedModel."
'--export_dir', required=False, default=None, type=str, help="Directory in which to write exported SavedModel."
)
p.add_argument(
p.add_argument(
'--to_predict',
required=False,
default=None,
type=str,
help="Path to file or directory of files to run prediction on."
)
p.add_argument(
'--batch_size',
type=int,
required=False,
help="""Size of each minibatch per GPU."""
)
p.add_argument('--batch_size', type=int, required=True, help="""Size of each minibatch per GPU.""")
p.add_argument('--num_iter', type=int, required=False, default=1, help="""Number of iterations to run.""")
p.add_argument(
'--num_iter',
type=int,
required=False,
default=1,
help="""Number of iterations to run."""
'--run_iter',
type=int,
required=False,
default=-1,
help="""Number of training iterations to run on single run."""
)
p.add_argument(
'--iter_unit',
choices=['epoch', 'batch'],
type=str,
required=False,
required=False,
default='epoch',
help="""Unit of iterations."""
)
@ -109,14 +110,9 @@ def parse_cmdline():
# Tensor format used for the computation.
p.add_argument(
'--data_format',
choices=['NHWC', 'NCHW'],
type=str,
default='NCHW',
required=False,
help=argparse.SUPPRESS
'--data_format', choices=['NHWC', 'NCHW'], type=str, default='NCHW', required=False, help=argparse.SUPPRESS
)
p.add_argument(
'--model_dir',
type=str,
@ -125,30 +121,34 @@ def parse_cmdline():
help="""Directory in which to write model. If undefined, results dir will be used."""
)
p.add_argument(
'--results_dir',
type=str,
required=True,
required=False,
default='.',
help="""Directory in which to write training logs, summaries and checkpoints."""
)
p.add_argument(
'--display_every',
'--log_filename',
type=str,
required=False,
default='log.json',
help="Name of the JSON file to which write the training log"
)
p.add_argument(
'--display_every',
default=10,
type=int,
required=False,
type=int,
required=False,
help="""How often (in batches) to print out running information."""
)
p.add_argument(
'--lr_init',
default=0.1,
type=float,
required=False,
help="""Initial value for the learning rate."""
'--lr_init', default=0.1, type=float, required=False, help="""Initial value for the learning rate."""
)
p.add_argument(
'--lr_warmup_epochs',
default=5,
@ -157,12 +157,15 @@ def parse_cmdline():
help="""Number of warmup epochs for learning rate schedule."""
)
p.add_argument('--weight_decay', default=1e-4, type=float, required=False, help="""Weight Decay scale factor.""")
p.add_argument(
'--weight_decay',
default=1e-4,
type=float,
required=False,
help="""Weight Decay scale factor."""
'--weight_init',
default='fan_out',
choices=['fan_in', 'fan_out'],
type=str,
required=False,
help="""Model weight initialization method."""
)
p.add_argument(
@ -172,15 +175,10 @@ def parse_cmdline():
required=False,
help="""SGD momentum value for the Momentum optimizer."""
)
#Select fp32 or non-AMP fp16 precision arithmetic.
p.add_argument(
'--precision',
choices=['fp32', 'fp16'],
type=str,
default='fp32',
required=False,
help=argparse.SUPPRESS
'--precision', choices=['fp32', 'fp16'], type=str, default='fp32', required=False, help=argparse.SUPPRESS
)
p.add_argument(
@ -190,15 +188,11 @@ def parse_cmdline():
required=False,
help="""Loss scale for FP16 Training and Fast Math FP32."""
)
p.add_argument(
'--label_smoothing',
type=float,
default=0.0,
required=False,
help="""The value of label smoothing."""
'--label_smoothing', type=float, default=0.0, required=False, help="""The value of label smoothing."""
)
p.add_argument(
'--mixup',
type=float,
@ -206,7 +200,7 @@ def parse_cmdline():
required=False,
help="""The alpha parameter for mixup (if 0 then mixup is not applied)."""
)
_add_bool_argument(
parser=p,
name="use_static_loss_scaling",
@ -223,13 +217,7 @@ def parse_cmdline():
help="Enable XLA (Accelerated Linear Algebra) computation for improved performance."
)
_add_bool_argument(
parser=p,
name="use_dali",
default=False,
required=False,
help="Enable DALI data input."
)
_add_bool_argument(parser=p, name="use_dali", default=False, required=False, help="Enable DALI data input.")
_add_bool_argument(
parser=p,
@ -238,37 +226,27 @@ def parse_cmdline():
required=False,
help="Enable Automatic Mixed Precision to speedup FP32 computation using tensor cores."
)
_add_bool_argument(
parser=p,
name="use_cosine_lr",
default=False,
required=False,
help="Use cosine learning rate schedule."
parser=p, name="use_cosine_lr", default=False, required=False, help="Use cosine learning rate schedule."
)
p.add_argument(
'--seed',
type=int,
default=1,
help="""Random seed."""
)
p.add_argument('--seed', type=int, default=None, help="""Random seed.""")
p.add_argument(
'--gpu_memory_fraction',
type=float,
default=0.7,
help="""Limit memory fraction used by training script for DALI"""
)
p.add_argument(
'--gpu_id',
type=int,
default=0,
default=0,
help="""Specify ID of the target GPU on multi-device platform. Effective only for single-GPU mode."""
)
FLAGS, unknown_args = p.parse_known_args()
if len(unknown_args) > 0:

22
TensorFlow/Classification/ConvNets/utils/dali_index.sh Normal file
View file

@ -0,0 +1,22 @@
#!/bin/bash
SRC_DIR=${1}
DST_DIR=${2}
echo "Creating training file indexes"
for file in ${SRC_DIR}/train-*; do
BASENAME=$(basename $file)
DST_NAME=$DST_DIR/$BASENAME
echo "Creating index $DST_NAME for $file"
tfrecord2idx $file $DST_NAME
done
echo "Creating validation file indexes"
for file in ${SRC_DIR}/validation-*; do
BASENAME=$(basename $file)
DST_NAME=$DST_DIR/$BASENAME
echo "Creating index $DST_NAME for $file"
tfrecord2idx $file $DST_NAME
done

122
TensorFlow/Classification/RN50v1.5/utils/dali_utils.py → TensorFlow/Classification/ConvNets/utils/dali_utils.py
View file

@ -30,24 +30,29 @@ __all__ = ["get_synth_input_fn", "normalized_inputs"]
class HybridPipe(dali.pipeline.Pipeline):
def __init__(self,
tfrec_filenames,
tfrec_idx_filenames,
height, width,
batch_size,
num_threads,
device_id,
shard_id,
num_gpus,
deterministic=False,
dali_cpu=True,
training=True):
def __init__(
self,
tfrec_filenames,
tfrec_idx_filenames,
height,
width,
batch_size,
num_threads,
device_id,
shard_id,
num_gpus,
deterministic=False,
dali_cpu=True,
training=True
):
kwargs = dict()
if deterministic:
kwargs['seed'] = 7 * (1 + hvd.rank())
super(HybridPipe, self).__init__(batch_size, num_threads, device_id, **kwargs)
self.training = training
self.input = dali.ops.TFRecordReader(
path=tfrec_filenames,
index_path=tfrec_idx_filenames,
@ -56,43 +61,39 @@ class HybridPipe(dali.pipeline.Pipeline):
num_shards=num_gpus,
initial_fill=10000,
features={
'image/encoded':dali.tfrecord.FixedLenFeature((), dali.tfrecord.string, ""),
'image/class/label':dali.tfrecord.FixedLenFeature([1], dali.tfrecord.int64, -1),
'image/class/text':dali.tfrecord.FixedLenFeature([ ], dali.tfrecord.string, ''),
'image/object/bbox/xmin':dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/ymin':dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/xmax':dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/ymax':dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0)})
if dali_cpu:
self.decode = dali.ops.HostDecoder(device="cpu", output_type=dali.types.RGB)
resize_device = "cpu"
else:
self.decode = dali.ops.nvJPEGDecoder(
device="mixed",
output_type=dali.types.RGB)
resize_device = "gpu"
'image/encoded': dali.tfrecord.FixedLenFeature((), dali.tfrecord.string, ""),
'image/class/label': dali.tfrecord.FixedLenFeature([1], dali.tfrecord.int64, -1),
'image/class/text': dali.tfrecord.FixedLenFeature([], dali.tfrecord.string, ''),
'image/object/bbox/xmin': dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/ymin': dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/xmax': dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0),
'image/object/bbox/ymax': dali.tfrecord.VarLenFeature(dali.tfrecord.float32, 0.0)
}
)
if training:
self.resize = dali.ops.RandomResizedCrop(
device=resize_device,
size=[height, width],
interp_type=dali.types.INTERP_LINEAR,
random_aspect_ratio=[0.8, 1.25],
random_area=[0.1, 1.0],
num_attempts=100)
if self.training:
self.decode = dali.ops.ImageDecoderRandomCrop(
device="cpu" if dali_cpu else "mixed",
output_type=dali.types.RGB,
random_aspect_ratio=[0.75, 1.33],
random_area=[0.05, 1.0],
num_attempts=100
)
self.resize = dali.ops.Resize(device="cpu" if dali_cpu else "gpu", resize_x=width, resize_y=height)
else:
self.decode = dali.ops.ImageDecoder(device="cpu" if dali_cpu else "mixed", output_type=dali.types.RGB)
# Make sure that every image > 224 for CropMirrorNormalize
self.resize = dali.ops.Resize (device=resize_device, resize_shorter=256)
self.resize = dali.ops.Resize(device="cpu" if dali_cpu else "gpu", resize_shorter=256)
self.normalize = dali.ops.CropMirrorNormalize(
device="gpu",
output_dtype=dali.types.FLOAT,
crop=(height, width),
image_type=dali.types.RGB,
mean=[121., 115., 100.],
std=[70., 68., 71.],
output_layout=dali.types.NHWC)
self.uniform = dali.ops.Uniform(range=(0.0, 1.0))
mean=[123.68, 116.28, 103.53],
std=[58.395, 57.120, 57.385],
output_layout=dali.types.NHWC
)
self.cast_float = dali.ops.Cast(device="gpu", dtype=dali.types.FLOAT)
self.mirror = dali.ops.CoinFlip()
self.iter = 0
@ -106,21 +107,26 @@ class HybridPipe(dali.pipeline.Pipeline):
# Decode and augmentation
images = self.decode(images)
images = self.resize(images)
images = self.normalize(images.gpu(), mirror=self.mirror())
images = self.normalize(images.gpu(), mirror=self.mirror() if self.training else None)
return (images, labels)
class DALIPreprocessor(object):
def __init__(self,
filenames,
idx_filenames,
height, width,
batch_size,
num_threads,
dtype=tf.uint8,
dali_cpu=True,
deterministic=False,
training=False):
def __init__(
self,
filenames,
idx_filenames,
height,
width,
batch_size,
num_threads,
dtype=tf.uint8,
dali_cpu=True,
deterministic=False,
training=False
):
device_id = hvd.local_rank()
shard_id = hvd.rank()
num_gpus = hvd.size()
@ -136,8 +142,9 @@ class DALIPreprocessor(object):
num_gpus=num_gpus,
deterministic=deterministic,
dali_cpu=dali_cpu,
training=training)
training=training
)
daliop = dali_tf.DALIIterator()
with tf.device("/gpu:0"):
@ -145,10 +152,11 @@ class DALIPreprocessor(object):
pipeline=pipe,
shapes=[(batch_size, height, width, 3), (batch_size, 1)],
dtypes=[tf.float32, tf.int64],
device_id=device_id)
device_id=device_id
)
def get_device_minibatches(self):
with tf.device("/gpu:0"):
self.labels -= 1 # Change to 0-based (don't use background class)
self.labels = tf.squeeze(self.labels)
return self.images, self.labels
self.labels -= 1 # Change to 0-based (don't use background class)
self.labels = tf.squeeze(self.labels, axis=-1)
return self.images, self.labels

54
TensorFlow/Classification/RN50v1.5/utils/data_utils.py → TensorFlow/Classification/ConvNets/utils/data_utils.py
View file

@ -27,9 +27,10 @@ from utils import dali_utils
__all__ = ["get_synth_input_fn", "normalized_inputs"]
_R_MEAN = 123.68
_G_MEAN = 116.78
_B_MEAN = 103.94
_G_MEAN = 116.28
_B_MEAN = 103.53
_CHANNEL_MEANS = [_R_MEAN, _G_MEAN, _B_MEAN]
_CHANNEL_STDS = [58.395, 57.120, 57.385]
_NUM_CHANNELS = 3
@ -109,12 +110,8 @@ def get_tfrecords_input_fn(filenames, batch_size, height, width, training, disto
def preproc_func(record, counter_):
return image_processing.preprocess_image_record(record, height, width, _NUM_CHANNELS, training)
ds = ds.cache()
if training:
ds = ds.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=shuffle_buffer_size, seed=seed))
else:
ds = ds.repeat()
@ -131,8 +128,9 @@ def get_tfrecords_input_fn(filenames, batch_size, height, width, training, disto
return ds
def get_inference_input_fn(filenames, height, width, num_threads):
ds = tf.data.Dataset.from_tensor_slices(filenames)
counter = tf.data.Dataset.range(sys.maxsize)
@ -140,38 +138,37 @@ def get_inference_input_fn(filenames, height, width, num_threads):
def preproc_func(record, counter_):
return image_processing.preprocess_image_file(record, height, width, _NUM_CHANNELS, is_training=False)
ds = ds.apply(
tf.data.experimental.map_and_batch(
map_func=preproc_func,
num_parallel_calls=num_threads,
batch_size=1
)
tf.data.experimental.map_and_batch(map_func=preproc_func, num_parallel_calls=num_threads, batch_size=1)
)
ds = ds.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)
return ds
def get_dali_input_fn(filenames, idx_filenames, batch_size, height, width, training, distort_color, num_threads, deterministic):
def get_dali_input_fn(
filenames, idx_filenames, batch_size, height, width, training, distort_color, num_threads, deterministic
):
if idx_filenames is None:
raise ValueError("Must provide idx_filenames for DALI's reader")
preprocessor = dali_utils.DALIPreprocessor(
filenames,
idx_filenames,
height, width,
height,
width,
batch_size,
num_threads,
dali_cpu=False,
deterministic=deterministic,
training=training)
training=training
)
images, labels = preprocessor.get_device_minibatches()
return (images, labels)
@ -189,12 +186,15 @@ def normalized_inputs(inputs):
means_per_channel = tf.reshape(_CHANNEL_MEANS, [1, 1, num_channels])
means_per_channel = tf.cast(means_per_channel, dtype=inputs.dtype)
inputs = tf.subtract(inputs, means_per_channel)
stds_per_channel = tf.reshape(_CHANNEL_STDS, [1, 1, num_channels])
stds_per_channel = tf.cast(stds_per_channel, dtype=inputs.dtype)
inputs = tf.subtract(inputs, means_per_channel)
return tf.divide(inputs, stds_per_channel)
return tf.divide(inputs, 255.0)
def get_serving_input_receiver_fn(batch_size, height, width, num_channels, data_format, dtype=tf.float32):
if data_format not in ["NHWC", "NCHW"]:
raise ValueError("Unknown data_format: %s" % str(data_format))
@ -202,9 +202,9 @@ def get_serving_input_receiver_fn(batch_size, height, width, num_channels, data_
input_shape = [batch_size] + [height, width, num_channels]
else:
input_shape = [batch_size] + [num_channels, height, width]
def serving_input_receiver_fn():
features = tf.placeholder(dtype=dtype, shape=input_shape, name='input_tensor')
return tf.estimator.export.TensorServingInputReceiver(features=features, receiver_tensors=features)
return serving_input_receiver_fn

Some files were not shown because too many files have changed in this diff Show more