This repository provides a script and recipe to train the EfficientNet model to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.

• Model overview
  • Default configuration
  • Feature support matrix
    • Features
  • Mixed precision training
    • Enabling mixed precision
    • Enabling TF32
  • Quantization
    • Quantization-aware training
• Setup
  • Requirements
• Quick Start Guide
• Advanced
  • Scripts and sample code
  • Command-line options
  • Dataset guidelines
  • Training process
  • Inference process
    • NGC pretrained weights
  • QAT process
• Performance
  • Benchmarking
    • Training performance benchmark
    • Inference performance benchmark
  • Results
    • Training accuracy results
      • Training accuracy: NVIDIA A100 (8x A100 80GB)
      • Training accuracy: NVIDIA DGX-1 (8x V100 16GB)
      • Example plots
    • Training performance results
      • Training performance: NVIDIA A100 (8x A100 80GB)
      • Training performance: NVIDIA DGX-1 (8x V100 16GB)
      • Training performance: NVIDIA DGX-1 (8x V100 32GB)
  • Inference performance results
    • Inference performance: NVIDIA A100 (1x A100 80GB)
    • Inference performance: NVIDIA V100 (1x V100 16GB)
  • QAT results
    • QAT Training performance: NVIDIA DGX-1 (8x V100 32GB))
    • QAT Inference accuracy
• Release notes
  • Changelog
  • Known issues

EfficientNet is an image classification model family. It was first described in EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. The scripts provided enable you to train the EfficientNet-B0, EfficientNet-B4, EfficientNet-WideSE-B0 and, EfficientNet-WideSE-B4 models.

EfficientNet-WideSE models use Squeeze-and-Excitation layers wider than original EfficientNet models, the width of SE module is proportional to the width of Depthwise Separable Convolutions instead of block width.

WideSE models are slightly more accurate than original models.

This model is trained with mixed precision using Tensor Cores on Volta and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results over 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.

We use NHWC data layout when training using Mixed Precision.

The following sections highlight the default configurations for the EfficientNet models.

Optimizer

This model uses RMSprop with the following hyperparameters:

• Momentum (0.9)
• Learning rate (LR):
  • 0.08 for 4096 batch size for B0 models
  • 0.16 for 4096 batch size for B4 models scale the learning rate.
• Learning rate schedule - we use cosine LR schedule
• We use linear warmup of the learning rate during the first 16 epochs
• Weight decay (WD):
  • 1e-5 for B0 models
  • 5e-6 for B4 models
• We do not apply WD on Batch Norm trainable parameters (gamma/bias)
• Label smoothing = 0.1
• MixUp = 0.2
• We train for 400 epochs

Optimizer for QAT

This model uses SGD optimizer for B0 models and RMSPROP optimizer alpha=0.853 epsilon=0.00422 for B4 models. Other hyperparameters we used are:

• Momentum:
  • 0.89 for B0 models
  • 0.9 for B4 models
• Learning rate (LR):
  • 0.0125 for 128 batch size for B0 models
  • 4.09e-06 for 32 batch size for B4 models scale the learning rate.
• Learning rate schedule:
  • cosine LR schedule for B0 models
  • linear LR schedule for B4 models
• Weight decay (WD):
  • 4.50e-05 for B0 models
  • 9.714e-04 for B4 models
• We do not apply WD on Batch Norm trainable parameters (gamma/bias)
• We train for: *10 epochs for B0 models *2 epochs for B4 models

Data augmentation

This model uses the following data augmentation:

• For training:

  • Auto-augmentation
  • Basic augmentation:
    • Normalization
    • Random resized crop to target images size (depending on model version)
      • Scale from 8% to 100%
      • Aspect ratio from 3/4 to 4/3
    • Random horizontal flip

• For inference:

  • Normalization
  • Scale to target image size + 32
  • Center crop to target image size

The following features are supported by this model:

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 more information about DALI, refer to the DALI product documentation.

We use NVIDIA DALI, which speeds up data loading when CPU becomes a bottleneck. DALI can use CPU or GPU, and outperforms the PyTorch native dataloader.

Run training with --data-backends dali-gpu or --data-backends dali-cpu to enable DALI. For DGXA100 and DGX1 we recommend --data-backends dali-cpu.

DALI currently does not support Autoaugmentation, so for best accuracy it has to be disabled.

APEX

A PyTorch extension that contains utility libraries, such as Automatic Mixed Precision (AMP), which require minimal network code changes to leverage Tensor Cores performance. Refer to the Enabling mixed precision section for more details.

QAT

Quantization aware training (QAT) is a method for changing precision to INT8 which speeds up the inference process at the price of a slight decrease of network accuracy. Refer to the Quantization section for more details.

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 requires two steps:

1. Porting the model to use the FP16 data type where appropriate.
2. Adding loss scaling to preserve small gradient values.

The ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in CUDA 8 in the NVIDIA Deep Learning SDK.

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.
• APEX tools for mixed precision training, see the NVIDIA Apex: Tools for Easy Mixed-Precision Training in PyTorch.

Mixed precision is enabled in PyTorch by using the Automatic Mixed Precision (AMP), a library from APEX that 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 PyTorch, loss scaling can be easily applied by using scale_loss() method provided by AMP. The scaling value to be used can be dynamic or fixed.

For an in-depth walk through on AMP, check out sample usage here. APEX is a PyTorch extension that contains utility libraries, such as AMP, which require minimal network code changes to leverage Tensor Cores performance.

To enable mixed precision, you can:

• Import AMP from APEX:

  from apex import amp

• Wrap model and optimizer in amp.initialize:

  model, optimizer = amp.initialize(model, optimizer, opt_level="O1", loss_scale="dynamic")

• Scale loss before backpropagation:

  with amp.scale_loss(loss, optimizer) as scaled_loss:
    scaled_loss.backward()

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.

Quantization is the process of transforming deep learning models to use parameters and computations at a lower precision. Traditionally, DNN training and inference have relied on the IEEE single-precision floating-point format, using 32 bits to represent the floating-point model weights and activation tensors.

This compute budget may be acceptable at training as most DNNs are trained in data centers or in the cloud with NVIDIA V100 or A100 GPUs that have significantly large compute capability and much larger power budgets. However, during deployment, these models are most often required to run on devices with much smaller computing resources and lower power budgets at the edge. Running a DNN inference using the full 32-bit representation is not practical for real-time analysis given the compute, memory, and power constraints of the edge.

To help reduce the compute budget, while not compromising on the structure and number of parameters in the model, you can run inference at a lower precision. Initially, quantized inferences were run at half-point precision with tensors and weights represented as 16-bit floating-point numbers. While this resulted in compute savings of about 1.2–1.5x, there was still some compute budget and memory bandwidth that could be leveraged. In lieu of this, models are now quantized to an even lower precision, with an 8-bit integer representation for weights and tensors. This results in a model that is 4x smaller in memory and about 2–4x faster in throughput.

While 8-bit quantization is appealing to save compute and memory budgets, it is a lossy process. During quantization, a small range of floating-point numbers are squeezed to a fixed number of information buckets. This results in loss of information.

The minute differences which could originally be resolved using 32-bit representations are now lost because they are quantized to the same bucket in 8-bit representations. This is similar to rounding errors that one encounters when representing fractional numbers as integers. To maintain accuracy during inferences at a lower precision, it is important to try and mitigate errors arising due to this loss of information.

In QAT, the quantization error is considered when training the model. The training graph is modified to simulate the lower precision behavior in the forward pass of the training process. This introduces the quantization errors as part of the training loss, which the optimizer tries to minimize during the training. Thus, QAT helps in modeling the quantization errors during training and mitigates its effects on the accuracy of the model at deployment.

However, the process of modifying the training graph to simulate lower precision behavior is intricate. To run QAT, it is necessary to insert FakeQuantization nodes for the weights of the DNN Layers and Quantize-Dequantize (QDQ) nodes to the intermediate activation tensors to compute their dynamic ranges.

For more information, see this Quantization paper and Quantization-Aware Training documentation. Tutorial for pytoch-quantization library can be found here pytorch-quantization tutorial.

It is important to mention that EfficientNet is NN, which is hard to quantize because the activation function all across the network is the SiLU (called also the Swish), whose negative values lie in very short range, which introduce a large quantization error. More details can be found in Appendix D of the Quantization paper.

The following section lists the requirements that you need to meet in order to start training the EfficientNet model.

This repository contains Dockerfile which extends the PyTorch NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:

• NVIDIA Docker
• PyTorch 21.03-py3 NGC container or newer
• Supported GPUs:
  • NVIDIA Volta architecture
  • NVIDIA Turing architecture
  • NVIDIA Ampere architecture

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 DGX Documentation:

• Getting Started Using NVIDIA GPU Cloud
• Accessing And Pulling From The NGC Container Registry
• Running PyTorch

To set up the required environment or create your own container, as an alternative to the use of the PyTorch NGC container, see the versioned NVIDIA Container Support Matrix.

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

1. Clone the repository.

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

2. Download and pre-process the dataset.

The EfficientNet script operates on ImageNet 1k, a widely popular image classification dataset from the ILSVRC challenge.

PyTorch can work directly on JPEGs, therefore, pre-processing/augmentation is not needed.

3. Download the images.

4. Extract the training 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 ..

5. Extract the validation data and move the images to subfolders:

mkdir val && mv ILSVRC2012_img_val.tar val/ && cd val && tar -xvf ILSVRC2012_img_val.tar
wget -qO- https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh | bash

The directory in which the train/ and val/ directories are placed, is referred to as <path to imagenet> in this document.

6. Build the EfficientNet PyTorch NGC container.

docker build . -t nvidia_efficientnet

7. Start an interactive session in the NGC container to run training/inference.

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

8. Start training

To run training for a standard configuration (DGX A100/DGX-1V, AMP/TF32/FP32, 400 Epochs), run one of the scripts in the ./efficientnet/training directory called ./efficientnet/training/{AMP, TF32, FP32}/{ DGX A100, DGX-1V }_efficientnet-<version>_{AMP, TF32, FP32}_{ 400 }E.sh.

Ensure ImageNet is mounted in the /imagenet directory.

For example: bash ./efficientnet/training/AMP/DGXA100_efficientnet-b0_AMP.sh <path were to store checkpoints and logs>

9. Start inference

You can download pre-trained weights from NGC:

wget --content-disposition  -O 

unzip 

To run inference on ImageNet, run:

python ./main.py --arch efficientnet-<version> --evaluate --epochs 1 --pretrained -b <batch size> <path to imagenet>

To run inference on JPEG image using pre-trained weights, run:

python classify.py --arch efficientnet-<version> --pretrained --precision AMP|FP32 --image <path to JPEG image>

The following sections provide greater details of the dataset, running training and inference, and the training results.

For a non-standard configuration, run:

• For 1 GPU

  • FP32 python ./main.py --arch efficientnet-<version> --label-smoothing 0.1 <path to imagenet> python ./main.py --arch efficientnet-<version> --label-smoothing 0.1 --amp --static-loss-scale 256 <path to imagenet>

• For multiple GPUs

  • FP32 python ./multiproc.py --nproc_per_node 8 ./main.py --arch efficientnet-<version> --label-smoothing 0.1 <path to imagenet>
  • AMP python ./multiproc.py --nproc_per_node 8 ./main.py --arch efficientnet-<version> --label-smoothing 0.1 --amp --static-loss-scale 256 <path to imagenet>

Use python ./main.py -h to obtain the list of available options in the main.py script.

To see the full list of available options and their descriptions, use the -h or --help command-line option, for example:

python main.py -h

To use your own dataset, divide it into directories. For example:

• Training images - train/<class id>/<image>
• Validation images - val/<class id>/<image>

If your dataset has a number of classes different than 1000, you need to pass the --num_classes N flag to the training script.

All the results of the training will be stored in the directory specified with --workspace argument.

The script will store:

• the most recent checkpoint - checkpoint.pth.tar (unless --no-checkpoints flag is used).
• the checkpoint with the best validation accuracy - model_best.pth.tar (unless --no-checkpoints flag is used).
• the JSON log - in the file specified with the --raport-file flag.

Metrics gathered through training:

• train.loss - training loss
• train.total_ips - training speed measured in images/second
• train.compute_ips - training speed measured in images/second, not counting data loading
• train.data_time - time spent on waiting on data
• train.compute_time - time spent in forward/backward pass

To restart training from the checkpoint use the --resume option.

To start training from pretrained weights (for example, downloaded from NGC) use the --pretrained-from-file option.

The difference between --resume and --pretrained-from-file flags is that the pretrained weights contain only model weights, and checkpoints, apart from model weights, contain optimizer state, LR scheduler state.

Checkpoints are suitable for dividing the training into parts, for example, in order to divide the training job into shorter stages, or restart training after an infrastructure failure.

Pretrained weights can be used as a base for fine tuning the model to a different dataset, or as a backbone to detection models.

Validation is done every epoch, and can be also run separately on a checkpointed model.

python ./main.py --arch efficientnet-<version> --evaluate --epochs 1 --resume <path to checkpoint> -b <batch size> <path to imagenet>

Metrics gathered through training:

• val.loss - validation loss
• val.top1 - validation top1 accuracy
• val.top5 - validation top5 accuracy
• val.total_ips - inference speed measured in images/second
• val.compute_ips - inference speed measured in images/second, not counting data loading
• val.data_time - time spent on waiting on data
• val.compute_time - time spent on inference

To run inference on JPEG image, you have to first extract the model weights from checkpoint:

python checkpoint2model.py --checkpoint-path <path to checkpoint> --weight-path <path where weights will be stored>

Then, run the classification script:

python classify.py --arch efficientnet-<version> --pretrained-from-file <path to weights from previous step> --precision AMP|FP32 --image <path to JPEG image>

You can also run the ImageNet validation on pretrained weights:

python ./main.py --arch efficientnet-<version> --evaluate --epochs 1 --pretrained-from-file <path to pretrained weights> -b <batch size> <path to imagenet>

Pretrained weights can be downloaded from NGC:

wget <ngc weights url>

URL for each model can be found in the following table:

To run inference on ImageNet, run:

python ./main.py --arch efficientnet-<version> --evaluate --epochs 1 --pretrained -b <batch size> <path to imagenet>

To run inference on JPEG images using pretrained weights, run:

python classify.py --arch efficientnet-<version> --pretrained --precision AMP|FP32 --image <path to JPEG image>

EfficientNet-b0 and EfficientNet-b4 models can be quantized using the QAT process from running the quant_main.py script.

python ./quant_main.py <path to imagenet> --arch efficientnet-quant-<version> --epochs <# of QAT epochs> --pretrained-from-file <path to non-quantized model weights> <any other parameters for training such as batch, momentum etc.>

During the QAT process, evaluation is done in the same way as during standard training. quant_main.py works in the same way as the original main.py script, but with quantized models. It means that quant_main.py can be used to resume the QAT process with the flag --resume:

python ./quant_main.py <path to imagenet> --arch efficientnet-quant-<version> --resume <path to mid-training checkpoint> ...

or to evaluate a created checkpoint with the flag --evaluate:

python ./quant_main.py --arch efficientnet-quant-<version> --evaluate --epochs 1 --resume <path to checkpoint> -b <batch size> <path to imagenet>

It also can run on multi-GPU in an identical way as the standard main.py script:

python ./multiproc.py --nproc_per_node 8 ./quant_main.py --arch efficientnet-quant-<version> ... <path to imagenet>

There is also a possibility to transform trained models (quantized or not) into ONNX format, which is needed to convert it later into TensorRT, where quantized networks are much faster during inference. Conversion to TensorRT will be supported in the next release. The conversion to ONNX consists of two steps:

• translate checkpoint to pure weights:

python checkpoint2model.py --checkpoint-path <path to quant checkpoint> --weight-path <path where quant weights will be stored>

• translate pure weights to ONNX:

python model2onnx.py --arch efficientnet-quant-<version> --pretrained-from-file <path to model quant weights> -b <batch size> --trt True

Quantized models could also be used to classify new images using the classify.py flag. For example: python classify.py --arch efficientnet-quant-<version> --pretrained-from-file <path to quant weights> --image <path to JPEG image>

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.

The following section shows how to run benchmarks measuring the model performance in training and inference modes.

To benchmark training, run:

• For 1 GPU

  • FP32 (V100 GPUs only) python ./launch.py --model efficientnet-<version> --precision FP32 --mode benchmark_training --platform DGX1V <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100
  • TF32 (A100 GPUs only) python ./launch.py --model efficientnet-<version> --precision TF32 --mode benchmark_training --platform DGXA100 <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100
  • AMP python ./launch.py --model efficientnet-<version> --precision AMP --mode benchmark_training --platform <DGX1V|DGXA100> <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100

• For multiple GPUs

  • FP32 (V100 GPUs only) python ./launch.py --model efficientnet-<version> --precision FP32 --mode benchmark_training --platform DGX1V <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100
  • TF32 (A100 GPUs only) python ./multiproc.py --nproc_per_node 8 ./launch.py --model efficientnet-<version> --precision TF32 --mode benchmark_training --platform DGXA100 <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100
  • AMP python ./multiproc.py --nproc_per_node 8 ./launch.py --model efficientnet-<version> --precision AMP --mode benchmark_training --platform <DGX1V|DGXA100> <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100

Each of these scripts will run 100 iterations and save results in the benchmark.json file.

To benchmark inference, run:

• FP32 (V100 GPUs only)

python ./launch.py --model efficientnet-<version> --precision FP32 --mode benchmark_inference --platform DGX1V <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100

• TF32 (A100 GPUs only)

python ./launch.py --model efficientnet-<version> --precision TF32 --mode benchmark_inference --platform DGXA100 <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100

• AMP

python ./launch.py --model efficientnet-<version> --precision AMP --mode benchmark_inference --platform <DGX1V|DGXA100> <path to imagenet> --raport-file benchmark.json --epochs 1 --prof 100

Each of these scripts will run 100 iterations and save results in the benchmark.json file.

Our results were obtained by running the applicable training script in the pytorch-21.03 NGC container.

To achieve these same results, follow the steps in the Quick Start Guide.

Our results were obtained by running the applicable efficientnet/training/<AMP|TF32>/*.sh training script in the PyTorch 20.12 NGC container on NVIDIA DGX A100 (8x A100 80GB) GPUs.

Our results were obtained by running the applicable efficientnet/training/<AMP|FP32>/*.sh training script in the PyTorch 20.12 NGC container on NVIDIA DGX-1 (8x V100 16GB) GPUs.

The following images show an A100 run.

Our results were obtained by running the applicable efficientnet/training/<AMP|TF32>/*.sh training script in the PyTorch 21.03 NGC container on NVIDIA DGX A100 (8x A100 80GB) GPUs.

Our results were obtained by running the applicable efficientnet/training/<AMP|FP32>/*.sh training script in the PyTorch 21.03 NGC container on NVIDIA DGX-1 (8x V100 16GB) GPUs.

Our results were obtained by running the applicable efficientnet/training/<AMP|FP32>/*.sh training script in the PyTorch 21.03 NGC container on NVIDIA DGX-1 (8x V100 16GB) GPUs.

Our results were obtained by running the applicable efficientnet/inference/<AMP|FP32>/*.sh inference script in the PyTorch 21.03 NGC container on NVIDIA DGX-1 (8x V100 16GB) GPUs.

Our results were obtained by running the applicable efficientnet/inference/<AMP|FP32>/*.sh inference script in the PyTorch 21.03 NGC container on NVIDIA DGX-1 (8x V100 16GB) GPUs.

The best checkpoints generated during training were used as a base for the QAT.

1. April 2020

• Initial release

There are no known issues with this model.