The SSD320 v1.2 model is based on the [SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325) paper, which describes SSD as “a method for detecting objects in images using a single deep neural network”.
We have altered the network in order to improve accuracy and increase throughput. Changes we have made include:
- Replacing the VGG backbone with the more popular ResNet50.
- Adding multi-scale detection to the backbone using [Feature Pyramid Networks](https://arxiv.org/pdf/1612.03144.pdf).
- Replacing the original hard negative mining loss function with [Focal Loss](https://arxiv.org/pdf/1708.02002.pdf).
- Decreasing the input size to 320 x 320.
Our implementation is based on the existing [model from the TensorFlow models repository](https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync.config).
This model trains with mixed precision tensor cores on NVIDIA Volta GPUs, therefore you can get results much faster than training without tensor cores. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.
The following features were implemented in this model:
- Data-parallel multi-GPU training with Horovod.
- Mixed precision support with TensorFlow Automatic Mixed Precision (TF-AMP), which enables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.
- Tensor Core operations to maximize throughput using NVIDIA Volta GPUs.
- Dynamic loss scaling for tensor cores (mixed precision) training.
Because of these enhancements, the SSD320 v1.2 model achieves higher accuracy.
### Default configuration
We trained the model for 12500 steps (27 epochs) with the following setup:
- [SGDR](https://arxiv.org/pdf/1608.03983.pdf) with cosine decay learning rate
- Learning rate base = 0.16
- Momentum = 0.9
- Warm-up learning rate = 0.0693312
- Warm-up steps = 1000
- Batch size per GPU = 32
- Number of GPUs = 8
## Setup
The following section list the requirements in order to start training the SSD320 v1.2 model.
### Requirements
This repository contains `Dockerfile` which extends the TensorFlow NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following software:
Training the SSD model is implemented in the `object_detection/model_main.py` script.
All training parameters are set in the config files. Because evaluation is relatively time consuming,
it does not run every epoch. By default, evaluation is executed only once at the end of the training.
The model is evaluated using pycocotools distributed with the COCO dataset.
The number of evaluations can be changed using the `eval_count` parameter.
To run training with tensor cores, use `./examples/SSD320_FP16_{1,4,8}GPU.sh` scripts. For more details see [Enabling mixed precision](#enabling-mixed-precision) section below.
#### Data preprocessing
Before we feed data to the model, both during training and inference, we perform:
* Normalization
* Encoding bounding boxes
* Resize to 320x320
#### Data augmentation
During training we perform the following augmentation techniques:
* Random crop
* Random horizontal flip
* Color jitter
### Enabling mixed precision
[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](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html) training 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/sdk/mixed-precision-training/index.html) documentation.
- How to access and enable AMP for TensorFlow, see [Using TF-AMP](https://docs.nvidia.com/deeplearning/dgx/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.
## Benchmarking
The following section shows how to run benchmarks measuring the model performance in training and inference modes.
### Training performance benchmark
Training benchmark was run in various scenarios on V100 16G GPU. For each scenario, batch size was set to 32.
Batch size for the inference benchmark is controlled by the `--batch_size` argument,
while the checkpoint is provided to the script with the `--checkpoint_dir` argument.
The benchmark script provides extra arguments for extra control over the experiment.
We were using default values for the extra arguments during the experiments. For more details about them, please run:
```
bash examples/SSD320_FP16_inference.sh --help
```
## Results
The following sections provide details on how we achieved our performance and accuracy in training and inference.
### Training accuracy results
Our results were obtained by running the `./examples/SSD320_FP{16,32}_{1,4,8}GPU.sh` script in the TensorFlow-19.03-py3 NGC container on NVIDIA DGX-1 with 8x V100 16G GPUs.
All the results are obtained with batch size set to 32.
| **Number of GPUs** | **Mixed precision mAP** | **Training time with mixed precision** | **FP32 mAP** | **Training time with FP32** |
To achieve same results, follow the [Quick start guide](#quick-start-guide) outlined above.
Those results can be improved when [XLA](https://www.tensorflow.org/xla) is used
in conjunction with mixed precision, delivering up to 2x speedup over FP32 on a single GPU (~179 img/s).
However XLA is still considered experimental.
### Inference performance results
Our results were obtained by running the `examples/SSD320_FP{16,32}_inference.sh` script in the TensorFlow-19.03-py3 NGC container on NVIDIA DGX-1 with 1x V100 16G GPUs.
