This repository provides a script and recipe to train the Mask R-CNN model for Tensorflow to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.
Mask R-CNN is a convolution-based neural network for the task of object instance segmentation. The paper describing the model can be found [here](https://arxiv.org/abs/1703.06870). NVIDIA’s Mask R-CNN 20.06 is an optimized version of [Google's TPU implementation](https://github.com/tensorflow/tpu/tree/master/models/official/mask_rcnn), leveraging mixed precision arithmetic using Tensor Cores on NVIDIA Volta, Turing, and Ampere GPUs while maintaining target accuracy.
This model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results 2.2x 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.
This implementation of Mask-RCNN uses AMP to implement mixed precision training. It allows us to use FP16 training with FP32 master weights by modifying just a few lines of code.
Horovod is a distributed training framework for TensorFlow, Keras, PyTorch, and MXNet. The goal of Horovod is to make distributed deep learning fast and easy to use. For more information about how to get started with Horovod, see the [Horovod: Official repository](https://github.com/horovod/horovod).
Multi-GPU training with Horovod
Our model uses Horovod to implement efficient multi-GPU training with NCCL. For details, see example sources in this repository or see the [TensorFlow tutorial](https://github.com/horovod/horovod/#usage).
**XLA support (experimental)**
XLA is a domain-specific compiler for linear algebra that can accelerate TensorFlow models with potentially no source code changes. The results are improvements in speed and memory usage: most internal benchmarks run ~1.1-1.5x faster after XLA is enabled.
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 architecture, 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/performance/mixed-precision-training/index.html) previously required two steps:
1. Porting the model to use the FP16 data type where appropriate.
2. 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.
- 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.
- 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.
- 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.
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/sdk/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.
TensorFloat-32 (TF32) is the new math mode in [NVIDIA A100](https://www.nvidia.com/en-us/data-center/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](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/) blog post.
3. Start an interactive session in the NGC container to run training/inference.
Run the following command to launch the Docker container, the only argument is the *absolute path* to the
`data directory` which holds or will hold the `tfrecords` data. If data has not already been downloaded in the `data directory` then download it in step 4, else step 4 can be skipped.
The above script trains a model and performs an evaluation on the COCO 2017 dataset. By default, this training script:
- Uses 8 GPUs.
- Saves a checkpoint every 3696 iterations and at the end of training. All checkpoints, evaluation results and training logs are saved to the `/results` directory (in the container which can be mounted to a local directory).
- Mixed precision training with Tensor Cores.
6. Start validation/evaluation.
- For evaluation with AMP precision: `bash ./scripts/evaluation_AMP.sh`
-`dataset/` - A folder that contains shell scripts and Python files to download the dataset.
-`mask_rcnn_main.py` - Is the main function that is the starting point for the training and evaluation process.
-`docker/` - A folder that contains scripts to build a Docker image and start an interactive session.
### Parameters
#### `mask_rcnn_main.py` script parameters
You can modify the training behavior through the various flags in both the `train_net.py` script and through overriding specific parameters in the config files. Flags in the `mask_rcnn_main.py` script are as follows:
-`--mode` - Specifies the action to take like `train`, `train_and_eval` or `eval`.
-`--checkpoint` - The checkpoint of the backbone.
-`--eval_samples` - Number of samples to evaluate.
-`--init_learning_rate` - Initial learning rate.
-`--learning_rate_steps` - Specifies at which steps to reduce the learning rate.
-`--num_steps_per_eval` - Specifies after how many steps of training evaluation should be performed.
-`--total_steps` - Specifies the total number of steps for which training should be run.
-`--train_batch_size` - Training batch size per GPU.
-`--eval_batch_size` - Evaluation batch size per GPU.
To see the full list of available options and their descriptions, use the `-h` or `--help` command-line option, for example:
`python mask_rcnn_main.py --helpfull`
### Getting the data
The Mask R-CNN model was trained on the COCO 2017 dataset. This dataset comes with a training and validation set.
This repository contains the `./dataset/download_and_preprocess_coco.sh` script which automatically downloads and preprocesses the training and validation sets. The helper scripts are also present in the `dataset/` folder.
#### Dataset guidelines
The data should be organized into the following structure:
```bash
<data/dir>
annotations/
instances_train2017.json
instances_val2017.json
train2017/
COCO_train2017_*.jpg
val2017/
COCO_val2017_*.jpg
```
### Training process
Training is performed using the `mask_rcnn_main.py` script along with parameters defined in the config files.
The default config files can be found in the
`mask_rcnn_tf/mask_rcnn/mask_rcnn_params.py, mask_rcnn_tf/mask_rcnn/cmd_utils.py` files. To specify which GPUs to train on, `CUDA_VISIBLE_DEVICES` variable can be changed in the training scripts
provided in the `scripts` folder.
This script outputs results to the `/results` directory by default. The training log will contain information about:
- Loss, time per iteration, learning rate and memory metrics
- Performance values such as throughput per step
- Test accuracy and test performance values after evaluation
### Inference process
To run inference run `mask_rcnn_main.py` with commandline parameter
`mode=eval`. To run inference with a checkpoint, set the commandline
parameter `--model_dir` to `[absolute path of checkpoint folder]`.
The inference log will contain information about:
- Inference time per step
- Inference throughput per step
- Evaluation accuracy and performance values
## Performance
### Benchmarking
The following section shows how to run benchmarks measuring the model performance in training and inference modes.
##### Training accuracy: NVIDIA DGX A100 (8x A100 40GB)
Our results were obtained by building and launching the docker containers for TensorFlow 1.1x `./scripts/docker/build_tf1.sh`, `bash ./scripts/docker/launch_tf1.sh [data directory]` respectively and running the `scripts/train{_AMP}_{1,4,8}GPU.sh` training script on NVIDIA DGX A100 (8x A100 40GB) GPUs.
Our results were obtained by building and launching the docker containers for TensorFlow 1.1x `./scripts/docker/build_tf1.sh`, `bash ./scripts/docker/launch_tf1.sh [data directory]` respectively and running the `scripts/train{_AMP}_{1,4,8}GPU.sh` training script on NVIDIA DGX-1 with 8x V100 16GB GPUs.
Our results were obtained by running `python scripts/benchmark_training.py --gpus {1,4,8} --batch_size {2,4} [--amp]` benchmark script in the TensorFlow 1.1x 20.06-py3
NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs. Performance numbers (in images per second) were averaged over 200 steps omitting the first 100 warm-up steps.
Our results were obtained by running `python scripts/benchmark_training.py --gpus {1,4,8} --batch_size {2,4} [--amp]` benchmark script in the TensorFlow 1.1x 20.06-py3
NGC container on NVIDIA DGX-1 V100 (8x V100 16GB) GPUs. Performance numbers (in images per second) were averaged over 200 steps omitting the first 100 warm-up steps.
Our results were obtained by running `python scripts/benchmark_training.py --gpus {1,4,8} --batch_size {2,4} [--amp]` benchmark script in the TensorFlow 2.x 20.06-py3
NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs. Performance numbers (in images per second) were averaged over 200 steps omitting the first 100 warm-up steps.
Our results were obtained by running `python scripts/benchmark_training.py --gpus {1,4,8} --batch_size {2,4} [--amp]` benchmark script in the TensorFlow 2.x 20.06-py3
NGC container on NVIDIA DGX-1 V100 (8x V100 16GB) GPUs. Performance numbers (in images per second) were averaged over 200 steps omitting the first 100 warm-up steps.
Our results were obtained by running `python scripts/benchmark_inference.py --batch_size {2,4,8} [--amp]` benchmark script in the TensorFlow 1.1x 20.06-py3
NGC container on NVIDIA DGX A100 (1x A100 40GB) GPU.
Our results were obtained by running `python scripts/benchmark_inference.py --batch_size {2,4,8} [--amp]` benchmark script in the TensorFlow 1.1x 20.06-py3
NGC container on NVIDIA DGX-1 V100 (1x V100 16GB) GPU.
Our results were obtained by running `python scripts/benchmark_inference.py --batch_size {2,4,8} [--amp]` benchmark script in the TensorFlow 2.x 20.06-py3
NGC container on NVIDIA DGX A100 (1x A100 40GB) GPU.
Our results were obtained by running `python scripts/benchmark_inference.py --batch_size {2,4,8} [--amp]` benchmark script in the TensorFlow 2.x 20.06-py3
NGC container on NVIDIA DGX-1 V100 (1x V100 16GB) GPU.
