Skip to content

Latest commit

 

History

History
161 lines (114 loc) · 8.72 KB

README.md

File metadata and controls

161 lines (114 loc) · 8.72 KB
Raw

Vision Transformer for PyTorch

This directory provides a script and recipe to train the Vision Transformer model to achieve state of the art accuracy, and is tested and maintained by Intel® Gaudi®. For further information on performance, refer to Intel Gaudi Model Performance Data page. Before you get started, make sure to review the Supported Configurations.

For further information on training deep learning models using Intel Gaudi AI accelerator, refer to developer.habana.ai.

Table of Contents

Model Overview

This is a PyTorch implementation of the Vision Transformer model described in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale paper. It is based on an earlier implementation from PyTorch Image Models(timm) and the official repository.

The paper shows that Transformers applied directly to image patches and pre-trained on large datasets perform well in image recognition task.

fig1

The Vision Transformer model achieves State-of-the-Art in image recognition task with the standard Transformer encoder and fixed-size patches. To perform classification, you can use the standard approach of adding an extra learnable "classification token" to the sequence.

fig2

Setup

Please follow the instructions provided in the Gaudi Installation Guide to set up the environment including the $PYTHON environment variable. To achieve the best performance, please follow the methods outlined in the Optimizing Training Platform Guide. The guides will walk you through the process of setting up your system to run the model on Gaudi.

Clone Intel Gaudi Model-References

In the docker container, clone this repository and switch to the branch that matches your Intel Gaudi software version. You can run the hl-smi utility to determine the Intel Gaudi software version.

git clone -b [Intel Gaudi software version] https://github.com/HabanaAI/Model-References /root/Model-References

Note: If Model-References repository path is not in the PYTHONPATH, make sure you update it:

export PYTHONPATH=$PYTHONPATH:/path/to/Model-References

Install Model Requirements

  1. Go to PyTorch Vision Transformer directory:
cd /path/to/ViT_directory
  1. Install the required packages using pip:
$PYTHON -m pip install -r requirements.txt

Download the Pre-trained Model

Below are the official checkpoints of Google:

  • Available models: ViT-B_16(85.8M), R50+ViT-B_16(97.96M), ViT-B_32(87.5M), ViT-L_16(303.4M), ViT-L_32(305.5M), ViT-H_14(630.8M)
    • imagenet21k pre-train models
      • ViT-B_16, ViT-B_32, ViT-L_16, ViT-L_32, ViT-H_14
    • imagenet21k pre-train + imagenet2012 fine-tuned models
      • ViT-B_16-224, ViT-B_16, ViT-B_32, ViT-L_16-224, ViT-L_16, ViT-L_32
    • Hybrid Model(Resnet50 + Transformer)
      • R50-ViT-B_16
# imagenet21k pre-train
wget https://storage.googleapis.com/vit_models/imagenet21k/{MODEL_NAME}.npz

# imagenet21k pre-train + imagenet2012 fine-tuning
wget https://storage.googleapis.com/vit_models/imagenet21k+imagenet2012/{MODEL_NAME}.npz

Prepare the Dataset

ImageNet 2012 dataset needs to be organized as per PyTorch requirements. For the specific requirements, refer to Data Processing section.

NOTE: It is assumed that the ImageNet dataset is downloaded and available at /data/pytorch/imagenet/ILSVRC2012/ path.

Training Examples

The Vision Transformer demos included in this release is Lazy mode training for different batch sizes with FP32 and BF16 mixed precision.

Single Card and Multi-Card Training Examples

Run training on 1 HPU:

  • 1 HPU, batch size 32, gradient accumulation 1, FP32:

    $PYTHON train.py --name imagenet1k_TF --dataset imagenet1K --data_path /data/pytorch/imagenet/ILSVRC2012 --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --num_steps 20000 --eval_every 1000 --train_batch_size 32 --gradient_accumulation_steps 1 --img_size 384 --learning_rate 0.06

  • 1 HPU, batch size 32, gradient accumulation=1, mixed precision BF16:

    $PYTHON train.py --name imagenet1k_TF --dataset imagenet1K --data_path /data/pytorch/imagenet/ILSVRC2012 --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --num_steps 20000 --eval_every 1000 --train_batch_size 32 --gradient_accumulation_steps 1 --img_size 384 --learning_rate 0.06 --autocast

  • 1 HPU, batch size 512, gradient accumulation 16, FP32:

    $PYTHON train.py --name imagenet1k_TF --dataset imagenet1K --data_path /data/pytorch/imagenet/ILSVRC2012 --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --num_steps 20000 --eval_every 1000 --train_batch_size 512 --gradient_accumulation_steps 16 --img_size 384 --learning_rate 0.06

  • 1 HPU, batch size 512, gradient accumulation 16, mixed precision BF16:

    $PYTHON train.py --name imagenet1k_TF --dataset imagenet1K --data_path /data/pytorch/imagenet/ILSVRC2012 --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --num_steps 20000 --eval_every 1000 --train_batch_size 512 --gradient_accumulation_steps 16 --img_size 384 --learning_rate 0.06 --autocast

Run training on 8 HPUs:

To run multi-card demo, make sure to set the following prior to the training:

  • The host machine has 512 GB of RAM installed.
  • The docker is installed and set up as per the Gaudi Setup and Installation Guide, so that the docker has access to all 8 cards required for multi-card demo.
  • All server network interfaces are up. You can change the state of each network interface managed by the habanalabs driver by running the following command: 
    sudo ip link set <interface_name> up

NOTE: To identify if a specific network interface is managed by the habanalabs driver type, run:

sudo ethtool -i <interface_name>

Run training on 8 HPUs, batch size 512, gradient accumulation 2, mixed precision BF16:

NOTE: mpirun map-by PE attribute value may vary on your setup. For the recommended calculation, refer to the instructions detailed in mpirun Configuration.

mpirun -n 8 --bind-to core --map-by socket:PE=6 --rank-by core --report-bindings --allow-run-as-root $PYTHON -u train.py --name imagenet1k_TF --dataset imagenet1K --data_path /data/pytorch/imagenet/ILSVRC2012 --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --num_steps 20000 --eval_every 1000 --train_batch_size 64 --gradient_accumulation_steps 2 --img_size 384 --learning_rate 0.06 --autocast

Supported Configurations

Validated on Intel Gaudi Software Version PyTorch Version Mode
Gaudi 1.10.0 2.0.1 Training

Changelog

Training Script Modifications

  • Added support for Gaudi devices:

    • Defined certain environment variables for Gaudi.
    • Added support to run training in Lazy mode in addition to the Eager mode.
    • mark_step() is performed to trigger execution.
    • Added support to use HPU accelerator plugin, DDP plugin for multi-card training and mixed precision plugin provided with installed PyTorch Lightning package.

  • Improved performance:

    • Enabled using fusedSGD instead of default SGD.
    • Moved the div before the matmul in attention module.

1.12.0

  • Removed HMP and switched to autocast.
  • Eager mode support is deprecated.

Known Issues

  • Placing mark_step() arbitrarily may lead to undefined behavior. It is recommended to keep mark_step() as shown in the provided scripts.
  • Only scripts and configurations mentioned in this README are supported and verified.