Krzysztof Kudrynski 49e23b4597 Adding links to performance benchmark page

2021-07-21 14:39:48 +02:00

46 KiB

Raw Blame History

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.

Model overview
Setup
- Requirements
Quick Start Guide
Advanced
Performance
- Benchmarking
  - Training performance benchmark
  - Inference performance benchmark
- Results
Release notes
- Changelog
- Known issues

Model overview

The SE-ResNeXt101-32x4d is a ResNeXt101-32x4d model with added Squeeze-and-Excitation module introduced in the Squeeze-and-Excitation Networks paper.

The following performance optimizations were implemented in this model:

JIT graph compilation with XLA
Multi-GPU training with Horovod
Automated mixed precision AMP

This model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. 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

Here is a diagram of the Squeeze and Excitation module architecture for ResNet-type models:

Image source: Squeeze-and-Excitation Networks

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.
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 ResNet family networks.
- 250 Epochs -> best possible accuracy.
For 250 epoch training we also use MixUp regularization.

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	Yes
NVIDIA DALI	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.

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.

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 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 in Volta, and following with both the Turing and Ampere 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 previously required two steps:

Porting the model to use the FP16 data type where appropriate.
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 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 paper and Training With Mixed Precision documentation.
Techniques used for mixed precision training, see the Mixed-Precision Training of Deep Neural Networks blog.
How to access and enable AMP for TensorFlow, see Using TF-AMP from the TensorFlow User Guide.

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 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'

Enabling TF32

TensorFloat-32 (TF32) is the new math mode in NVIDIA A100 GPUs for handling the matrix math also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require high dynamic range for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x blog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.

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:

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:

For those unable to use the TensorFlow NGC container to set up the required environment or create your own container, see the versioned NVIDIA Container Support Matrix.

Quick Start Guide

To train your model using mixed precision or TF32 with Tensor Cores or FP32, perform the following steps using the default parameters of the SE-ResNext101-32x4d model on the ImageNet dataset. For the specifics concerning training and inference, see the Advanced section.

Clone the repository.

git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/Classification/ConvNets

Download and preprocess the dataset. The SE-ResNext101-32x4d script operates on ImageNet 1k, a widely popular image classification dataset from the ILSVRC challenge.

Download the images
Extract the training and validation data:

mkdir train && mv ILSVRC2012_img_train.tar train/ && cd train
tar -xvf ILSVRC2012_img_train.tar && rm -f ILSVRC2012_img_train.tar
find . -name "*.tar" | while read NAME ; do mkdir -p "${NAME%.tar}"; tar -xvf "${NAME}" -C "${NAME%.tar}"; rm -f "${NAME}"; done
cd ..
mkdir val && mv ILSVRC2012_img_val.tar val/ && cd val && tar -xvf ILSVRC2012_img_val.tar

Preprocess dataset to TFRecord form using script. Additional metadata from autors repository might be required.

Build the SE-ResNext101-32x4d TensorFlow NGC container.

docker build . -t nvidia_rn50

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

nvidia-docker run --rm -it -v <path to imagenet>:/data/tfrecords --ipc=host nvidia_rn50

(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 ./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.

Start training. To run training for a standard configuration (as described in 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/DGX1_SE-RNxt101-32x4d_AMP_90E.sh /path/to/result /data

Additionally, features like DALI data preprocessing or TensorFlow XLA can be enabled with following arguments when running those scripts:

bash ./se-resnext101-32x4d/training/DGX1_SE-RNxt101-32x4d_AMP_90E.sh /path/to/result /data/ --xla --dali

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> --results_dir <output location> [--xla] [--amp]

The optional --xla and --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}]
               [--export_dir EXPORT_DIR] [--to_predict TO_PREDICT]       
               --batch_size BATCH_SIZE [--num_iter NUM_ITER]  
               [--run_iter RUN_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] [--seed SEED]
               [--gpu_memory_fraction GPU_MEMORY_FRACTION] [--gpu_id GPU_ID]
               [--finetune_checkpoint FINETUNE_CHECKPOINT] [--use_final_conv]
               [--quant_delay QUANT_DELAY] [--quantize] [--use_qdq]        
               [--symmetric] [--data_dir DATA_DIR]         
               [--data_idx_dir DATA_IDX_DIR] [--dali]
               [--synthetic_data_size SYNTHETIC_DATA_SIZE] [--lr_init LR_INIT]
               [--lr_warmup_epochs LR_WARMUP_EPOCHS] 
               [--weight_decay WEIGHT_DECAY] [--weight_init {fan_in,fan_out}]
               [--momentum MOMENTUM] [--label_smoothing LABEL_SMOOTHING]
               [--mixup MIXUP] [--cosine_lr] [--xla]            
               [--data_format {NHWC,NCHW}] [--amp]
               [--static_loss_scale STATIC_LOSS_SCALE]
                                                            
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.                           
  --mode {train,train_and_evaluate,evaluate,predict,training_benchmark,inference_benchmark}
                        The execution mode of the script.
  --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.
  --run_iter RUN_ITER   Number of training iterations to run on single 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 model. If undefined,         
                        results dir 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.
  --seed SEED           Random seed.
  --gpu_memory_fraction GPU_MEMORY_FRACTION
                        Limit memory fraction used by training script for DALI.
  --gpu_id GPU_ID       Specify ID of the target GPU on multi-device platform.
                        Effective only for single-GPU mode.
  --finetune_checkpoint FINETUNE_CHECKPOINT
                        Path to pre-trained checkpoint which will be used for
                        fine-tuning.
  --use_final_conv      Use convolution operator instead of MLP as last layer.
  --quant_delay QUANT_DELAY
                        Number of steps to be run before quantization starts
                        to happen.
  --quantize            Quantize weights and activations during training.
                        (Defaults to Assymmetric quantization)
  --use_qdq             Use QDQV3 op instead of FakeQuantWithMinMaxVars op for
                        quantization. QDQv3 does only scaling.
  --symmetric           Quantize weights and activations during training using
                        symmetric quantization.

Dataset arguments:
  --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-*'.
  --dali                Enable DALI data input.
  --synthetic_data_size SYNTHETIC_DATA_SIZE
                        Dimension of image for synthetic dataset.

Training arguments:
  --lr_init LR_INIT     Initial value for the learning rate.
  --lr_warmup_epochs LR_WARMUP_EPOCHS
                        Number of warmup epochs for 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.
  --label_smoothing LABEL_SMOOTHING
                        The value of label smoothing.
  --mixup MIXUP         The alpha parameter for mixup (if 0 then mixup is not
                        applied).
  --cosine_lr           Use cosine learning rate schedule.

Generic optimization arguments:
  --xla                 Enable XLA (Accelerated Linear Algebra) computation
                        for improved performance.
  --data_format {NHWC,NCHW}
                        Data format used to do calculations.
  --amp                 Enable Automatic Mixed Precision to speedup
                        computation using tensor cores.

Automatic Mixed Precision arguments:
  --static_loss_scale STATIC_LOSS_SCALE
                        Use static loss scaling in FP32 AMP.

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 --xla and --amp flags control XLA and AMP during inference.

Performance

The performance measurements in this document were conducted at the time of publication and may not reflect the performance achieved from NVIDIA’s latest software release. For the most up-to-date performance measurements, go to NVIDIA Data Center Deep Learning Product 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 / TF32
  
  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>
- AMP
  
  python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --amp --warmup_steps 200 --batch_size <batch size> --data_dir=<path to imagenet> --results_dir=<path to results directory>
For multiple GPUs
- FP32 / TF32
  
  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>
- AMP
  
  mpiexec --allow-run-as-root --bind-to socket -np <num_gpus> python ./main.py --arch=se-resnext101-32x4d --mode=training_benchmark --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 --xla and --dali flags. For proper throughput reporting the value of --num_iter must be greater than --warmup_steps value. Suggested batch sizes for training are 96 for mixed precision training and 64 for single precision training per single V100 16 GB.

If no --data_dir=<path to imagenet> flag is specified then the benchmarks will use a synthetic dataset. The resolution of synthetic images used can be controlled with --synthetic_data_size flag.

Inference performance benchmark

To benchmark the inference performance on a specific batch size, run:

FP32 / TF32

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>

python ./main.py --arch=se-resnext101-32x4d --mode=inference_benchmark --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. If no --data_dir=<path to imagenet> flag is specified then the benchmarks will use a synthetic dataset.

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. For proper throughput reporting the value of --num_iter must be greater than --warmup_steps value.

For performance benchamrk of raw model, synthetic dataset can be used. To use synthetic dataset, use --synthetic_data_size flag instead of --data_dir to specify input image size.

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 A100 (8x A100 40GB)

Our results were obtained by running the /se-resnet50v1.5/training/DGXA100_RN50_{PRECISION}_90E.sh training script in the TensorFlow 20.06-tf1-py3 NGC container NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs.

Epochs	Batch Size / GPU	Accuracy - TF32 (top1)	Accuracy - mixed precision (top1)
90	128 (TF32) / 256 (AMP)	79.73	79.60

Training accuracy: NVIDIA DGX-1 (8x V100 16G)

Our results were obtained by running the /se-resnext101-32x4d/training/{/DGX1_RNxt101-32x4d_{PRECISION}_{EPOCHS}E.sh training script in the TensorFlow 20.06-tf1-py3 NGC container 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

Training performance results

Training performance: NVIDIA DGX A100 (8x A100 40GB)

Our results were obtained by running the se-resnext101-32x4d/training/training_perf.sh benchmark script in the TensorFlow 20.06-tf1-py3 NGC container NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs. Performance numbers (in images per second) were averaged over an entire training epoch.

GPUs	Batch Size / GPU	Throughput - TF32 + XLA	Throughput - mixed precision + XLA	Throughput speedup (TF32 - mixed precision)	Weak scaling - TF32 + XLA	Weak scaling - mixed precision + XLA
1	128 (TF) / 256 (AMP)	342 img/s	975 img/s	2.86x	1.00x	1.00x
8	128 (TF) / 256 (AMP)	2610 img/s	7230 img/s	2.77x	7.63x	7.41x

Training performance: NVIDIA DGX-1 (8x V100 16G)

Our results were obtained by running the se-resnext101-32x4d/training/training_perf.sh benchmark script in the TensorFlow 20.06-tf1-py3 NGC container 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 + XLA	Throughput - mixed precision + XLA	Throughput speedup (FP32 - mixed precision)	Weak scaling - FP32 + XLA	Weak scaling - mixed precision + XLA
1	64 (FP32) / 96 (AMP)	152 img/s	475 img/s	3.12x	1.00x	1.00x
8	64 (FP32) / 96 (AMP)	1120 img/s	3360 img/s	3.00x	7.37x	7.07x

Training performance: NVIDIA DGX-2 (16x V100 32G)

Our results were obtained by running the se-resnext101-32x4d/training/training_perf.sh benchmark script in the TensorFlow 20.06-tf1-py3 NGC container 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 + XLA	Throughput - mixed precision + XLA	Throughput speedup (FP32 - mixed precision)	Weak scaling - FP32 + XLA	Weak scaling - mixed precision + XLA
1	64 (FP32) / 96 (AMP)	158 img/s	472 img/s	2.98x	1.00x	1.00x
16	64 (FP32) / 96 (AMP)	2270 img/s	6580 img/s	2.89x	14.36x	13.94x

Training Time for 90 Epochs

Training time: NVIDIA DGX A100 (8x A100 40GB)

Our results were estimated based on the training performance results on NVIDIA DGX A100 with (8x A100 40G) GPUs.

GPUs	Time to train - mixed precision + XLA	Time to train - TF32 + XLA
1	~36h	~102h
8	~5h	~14h

Training time: NVIDIA DGX-1 (8x V100 16G)

Our results were estimated based on the training performance results on NVIDIA DGX-1 with (8x V100 16G) GPUs.

GPUs	Time to train - mixed precision + XLA	Time to train - FP32 + XLA
1	~68h	~210h
8	~10h	~29h

Training time: NVIDIA DGX-2 (16x V100 32G)

Our results were estimated based on the training performance results on NVIDIA DGX-2 with (16x V100 32G) GPUs.

GPUs	Time to train - mixed precision + XLA	Time to train - FP32 + XLA
1	~68h	~202h
16	~5h	~14h

Inference performance results

Inference performance: NVIDIA DGX A100 (1x A100 40GB)

Our results were obtained by running the inference_benchmark.sh inferencing benchmarking script in the TensorFlow 20.06-tf1-py3 NGC container NGC container on NVIDIA DGX A100 with (1x A100 40G) GPU.

TF32 Inference Latency

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	95.32 img/s	10.52 ms	10.52 ms	10.55 ms	11.10 ms
2	169.59 img/s	11.82 ms	11.83 ms	11.92 ms	12.56 ms
4	258.97 img/s	15.45 ms	15.70 ms	15.78 ms	16.22 ms
8	355.09 img/s	22.53 ms	22.74 ms	22.84 ms	23.17 ms
16	561.11 img/s	28.52 ms	28.85 ms	29.09 ms	29.50 ms
32	698.94 img/s	45.78 ms	46.36 ms	46.56 ms	46.87 ms
64	751.17 img/s	85.21 ms	86.74 ms	87.27 ms	87.95 ms
128	802.64 img/s	159.47 ms	160.01 ms	160.35 ms	161.42 ms
256	840.72 img/s	304.50 ms	305.87 ms	306.11 ms	306.57 ms

TF32 Inference Latency + XLA

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	92.46 img/s	10.84 ms	10.90 ms	10.96 ms	11.14 ms
2	161.55 img/s	12.40 ms	12.44 ms	12.51 ms	12.62 ms
4	237.41 img/s	16.88 ms	17.54 ms	17.79 ms	18.25 ms
8	358.39 img/s	22.35 ms	23.56 ms	24.29 ms	25.53 ms
16	577.33 img/s	27.72 ms	28.64 ms	28.92 ms	29.22 ms
32	800.81 img/s	39.97 ms	40.93 ms	41.42 ms	41.87 ms
64	921.00 img/s	69.64 ms	70.44 ms	70.90 ms	79.54 ms
128	1024.70 img/s	124.99 ms	125.70 ms	126.10 ms	138.57 ms
256	1089.80 img/s	234.90 ms	236.02 ms	236.37 ms	237.26 ms

Mixed Precision Inference Latency

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	84.06 img/s	11.92 ms	11.94 ms	11.96 ms	12.08 ms
2	170.38 img/s	11.76 ms	11.82 ms	11.87 ms	11.94 ms
4	336.09 img/s	11.93 ms	12.06 ms	12.17 ms	12.62 ms
8	669.91 img/s	11.94 ms	12.33 ms	12.47 ms	12.88 ms
16	1119.49 img/s	14.36 ms	14.86 ms	15.11 ms	16.11 ms
32	1482.46 img/s	21.66 ms	22.04 ms	22.38 ms	23.72 ms
64	1680.85 img/s	38.09 ms	39.02 ms	39.34 ms	41.02 ms
128	1728.27 img/s	74.30 ms	74.92 ms	75.22 ms	75.60 ms
256	1761.56 img/s	145.33 ms	146.54 ms	146.83 ms	147.34 ms

Mixed Precision Inference Latency + XLA

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	74.83 img/s	13.39 ms	13.45 ms	13.49 ms	13.57 ms
2	135.28 img/s	14.81 ms	14.98 ms	15.10 ms	16.19 ms
4	272.18 img/s	14.70 ms	15.07 ms	15.30 ms	15.80 ms
8	517.69 img/s	15.50 ms	16.63 ms	17.05 ms	18.10 ms
16	1050.03 img/s	15.38 ms	16.84 ms	17.49 ms	17.97 ms
32	1781.06 img/s	18.27 ms	19.54 ms	20.00 ms	25.94 ms
64	2551.55 img/s	25.26 ms	26.03 ms	26.62 ms	29.67 ms
128	2834.59 img/s	45.50 ms	46.85 ms	47.72 ms	54.91 ms
256	3367.18 img/s	76.03 ms	77.06 ms	77.36 ms	78.13 ms

Inference performance: NVIDIA DGX-1 (1x V100 16G)

Our results were obtained by running the inference_benchmark.sh inferencing benchmarking script in the TensorFlow 20.06-tf1-py3 NGC container 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	75.72 img/s	13.25 ms	13.38 ms	13.50 ms	13.66 ms
2	112.58 img/s	17.90 ms	20.74 ms	20.91 ms	21.87 ms
4	191.09 img/s	20.93 ms	21.05 ms	21.09 ms	21.27 ms
8	235.39 img/s	33.98 ms	34.14 ms	34.19 ms	34.28 ms
16	315.24 img/s	50.76 ms	50.96 ms	51.01 ms	51.32 ms
32	376.05 img/s	85.09 ms	85.56 ms	85.71 ms	86.40 ms
64	427.39 img/s	149.84 ms	150.08 ms	150.37 ms	161.87 ms
128	460.82 img/s	277.76 ms	278.97 ms	279.48 ms	280.95 ms

Mixed Precision Inference Latency

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	66.44 img/s	15.10 ms	15.17 ms	15.25 ms	16.01 ms
2	132.33 img/s	15.16 ms	15.32 ms	15.37 ms	15.50 ms
4	273.84 img/s	14.63 ms	15.14 ms	15.83 ms	17.38 ms
8	509.35 img/s	15.71 ms	16.10 ms	16.21 ms	16.55 ms
16	770.02 img/s	20.78 ms	20.96 ms	21.03 ms	21.24 ms
32	926.46 img/s	34.55 ms	34.88 ms	35.05 ms	36.32 ms
64	1039.74 img/s	61.55 ms	61.82 ms	61.99 ms	62.32 ms
128	1102.00 img/s	116.15 ms	116.62 ms	116.80 ms	116.97 ms

Mixed Precision Inference Latency + XLA

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	58.55 img/s	17.12 ms	17.21 ms	17.28 ms	17.42 ms
2	105.00 img/s	19.10 ms	19.29 ms	19.36 ms	19.67 ms
4	207.60 img/s	19.31 ms	19.59 ms	19.67 ms	19.84 ms
8	413.16 img/s	19.37 ms	19.77 ms	19.87 ms	20.24 ms
16	739.12 img/s	21.80 ms	24.48 ms	24.71 ms	26.93 ms
32	1196.83 img/s	26.99 ms	27.10 ms	27.49 ms	28.80 ms
64	1470.31 img/s	43.74 ms	44.02 ms	44.18 ms	46.28 ms
128	1683.63 img/s	76.03 ms	77.00 ms	77.23 ms	78.15 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 20.06-tf1-py3 NGC container 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	71.44 img/s	14.07 ms	14.22 ms	14.43 ms	16.44 ms
2	149.68 img/s	13.43 ms	13.79 ms	13.94 ms	16.63 ms
4	183.01 img/s	21.85 ms	22.12 ms	22.18 ms	22.44 ms
8	220.67 img/s	36.25 ms	36.84 ms	37.17 ms	37.43 ms
16	310.27 img/s	51.57 ms	51.88 ms	52.09 ms	53.37 ms
32	381.41 img/s	83.89 ms	84.30 ms	84.66 ms	85.04 ms
64	440.37 img/s	145.45 ms	145.49 ms	145.86 ms	147.53 ms
128	483.84 img/s	264.54 ms	265.04 ms	265.46 ms	266.43 ms

Mixed Precision Inference Latency

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	73.06 img/s	13.74 ms	14.07 ms	14.20 ms	14.35 ms
2	155.23 img/s	12.95 ms	13.13 ms	13.33 ms	15.49 ms
4	303.68 img/s	13.23 ms	13.38 ms	13.46 ms	14.34 ms
8	583.43 img/s	13.72 ms	13.90 ms	14.08 ms	15.47 ms
16	783.30 img/s	20.43 ms	20.66 ms	21.31 ms	21.97 ms
32	932.10 img/s	34.34 ms	34.71 ms	34.81 ms	35.70 ms
64	1058.07 img/s	60.48 ms	60.75 ms	60.94 ms	62.49 ms
128	1129.65 img/s	113.30 ms	113.53 ms	113.66 ms	114.81 ms

Mixed Precision Inference Latency + XLA

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	66.43 img/s	15.14 ms	15.24 ms	15.31 ms	19.18 ms
2	122.85 img/s	16.39 ms	18.28 ms	18.45 ms	20.33 ms
4	247.80 img/s	16.14 ms	16.44 ms	16.57 ms	17.24 ms
8	498.19 img/s	16.07 ms	16.26 ms	16.66 ms	17.70 ms
16	831.20 img/s	19.40 ms	19.30 ms	19.39 ms	25.41 ms
32	1223.75 img/s	26.42 ms	26.31 ms	26.70 ms	29.88 ms
64	1520.64 img/s	42.09 ms	42.45 ms	42.57 ms	42.84 ms
128	1739.61 img/s	73.58 ms	73.98 ms	74.17 ms	74.72 ms

Inference performance: NVIDIA T4 (1x T4 16G)

Our results were obtained by running the inference_benchmark.sh inferencing benchmarking script in the TensorFlow 20.06-tf1-py3 NGC container 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	27.39 img/s	36.68 ms	38.85 ms	39.01 ms	40.40 ms
2	44.56 img/s	44.96 ms	46.25 ms	46.92 ms	48.92 ms
4	65.11 img/s	61.43 ms	62.22 ms	62.93 ms	65.01 ms
8	80.09 img/s	99.88 ms	100.34 ms	100.85 ms	101.79 ms
16	93.98 img/s	170.24 ms	170.72 ms	171.27 ms	171.98 ms
32	99.86 img/s	320.42 ms	320.99 ms	321.37 ms	322.28 ms
64	103.31 img/s	619.44 ms	620.08 ms	620.55 ms	622.19 ms
128	105.16 img/s	1217.18 ms	1218.09 ms	1218.59 ms	1221.16 ms

Mixed Precision Inference Latency

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	57.21 img/s	17.57 ms	18.06 ms	18.15 ms	20.74 ms
2	80.34 img/s	24.97 ms	25.38 ms	25.69 ms	27.12 ms
4	115.12 img/s	34.77 ms	35.61 ms	36.74 ms	37.61 ms
8	147.51 img/s	54.24 ms	54.79 ms	55.28 ms	58.25 ms
16	173.83 img/s	92.04 ms	92.50 ms	93.26 ms	94.72 ms
32	182.19 img/s	175.64 ms	176.51 ms	177.44 ms	178.52 ms
64	193.20 img/s	331.25 ms	332.56 ms	333.34 ms	334.58 ms
128	195.17 img/s	655.82 ms	657.24 ms	658.79 ms	661.76 ms

Mixed Precision Inference Latency + XLA

Batch Size	Avg throughput	Avg latency	90% Latency	95% Latency	99% Latency
1	46.19 img/s	21.72 ms	21.90 ms	21.93 ms	23.64 ms
2	80.98 img/s	24.77 ms	24.99 ms	25.15 ms	25.63 ms
4	129.49 img/s	30.89 ms	31.26 ms	31.34 ms	32.31 ms
8	156.91 img/s	51.00 ms	52.17 ms	52.51 ms	53.32 ms
16	204.45 img/s	78.26 ms	79.58 ms	79.96 ms	80.44 ms
32	215.22 img/s	148.68 ms	149.63 ms	150.41 ms	151.62 ms
64	235.36 img/s	272.05 ms	273.56 ms	274.33 ms	275.86 ms
128	244.45 img/s	523.62 ms	525.12 ms	525.89 ms	528.42 ms

Release notes

Changelog

April 2020

Initial release August 2020
Updated command line argument names
Added support for syntetic dataset with different image size

Known issues

Performance without XLA enabled is low due to BN + ReLU fusion bug.

46 KiB Raw Blame History Unescape Escape

SE-ResNext101-32x4d for TensorFlow

Table Of Contents

Model overview

Model architecture

Default configuration

Optimizer

Data Augmentation

Feature support matrix

Features

Mixed precision training

Enabling mixed precision

Enabling TF32

Setup

Requirements

Quick Start Guide

Advanced

Scripts and sample code

Parameters

The main.py script

Inference process

Performance

Benchmarking

Training performance benchmark

Inference performance benchmark

Results

Training accuracy results

Training accuracy: NVIDIA DGX A100 (8x A100 40GB)

Training accuracy: NVIDIA DGX-1 (8x V100 16G)

Training performance results

Training performance: NVIDIA DGX A100 (8x A100 40GB)

Training performance: NVIDIA DGX-1 (8x V100 16G)

Training performance: NVIDIA DGX-2 (16x V100 32G)

Training Time for 90 Epochs

Training time: NVIDIA DGX A100 (8x A100 40GB)

Training time: NVIDIA DGX-1 (8x V100 16G)

Training time: NVIDIA DGX-2 (16x V100 32G)

Inference performance results

Inference performance: NVIDIA DGX A100 (1x A100 40GB)

Inference performance: NVIDIA DGX-1 (1x V100 16G)

Inference performance: NVIDIA DGX-2 (1x V100 32G)

Inference performance: NVIDIA T4 (1x T4 16G)

Release notes

Changelog

Known issues

46 KiB

Raw Blame History

The `main.py` script