ViT Quantization Toolkit

This tutorial contains the guidance for ViT Quantization Toolkit.

1. Download Pre-trained model (Google's Official Checkpoint)

• 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

2 Usage

2.1 Environment setup

• Initialize submodule

  bashCopy

  git submodule update --init
  

• Run the container.

  You can choose pytorch version you want. Here, we list some possible images:

  • nvcr.io/nvidia/pytorch:22.09-py3 contains the PyTorch 1.13.0 and python 3.8

• Install additional dependencies (not included by container)

  bashCopy

  pip install timm==0.4.12 termcolor==1.1.0 pytorch-quantization ml_collections
  

2.2 Data preparation

We use standard ImageNet dataset, you can download it from http://image-net.org/. We provide the following two ways to load data:

• For standard folder dataset, move validation images to labeled sub-folders. The file structure should look like:

  bashCopy

  $ tree data
  imagenet
  ├── train
  │   ├── class1
  │   │   ├── img1.jpeg
  │   │   ├── img2.jpeg
  │   │   └── ...
  │   ├── class2
  │   │   ├── img3.jpeg
  │   │   └── ...
  │   └── ...
  └── val
      ├── class1
      │   ├── img4.jpeg
      │   ├── img5.jpeg
      │   └── ...
      ├── class2
      │   ├── img6.jpeg
      │   └── ...
      └── ...
  
  

• To boost the slow speed when reading images from massive small files, we also support zipped ImageNet, which includes four files:

  • train.zip, val.zip: which store the zipped folder for train and validate splits.
  • train_map.txt, val_map.txt: which store the relative path in the corresponding zip file and ground truth label. Make sure the data folder looks like this:
  bashCopy

  $ tree data
  data
  └── ImageNet-Zip
      ├── train_map.txt
      ├── train.zip
      ├── val_map.txt
      └── val.zip
  
  $ head -n 5 data/ImageNet-Zip/val_map.txt
  ILSVRC2012_val_00000001.JPEG	65
  ILSVRC2012_val_00000002.JPEG	970
  ILSVRC2012_val_00000003.JPEG	230
  ILSVRC2012_val_00000004.JPEG	809
  ILSVRC2012_val_00000005.JPEG	516
  
  $ head -n 5 data/ImageNet-Zip/train_map.txt
  n01440764/n01440764_10026.JPEG	0
  n01440764/n01440764_10027.JPEG	0
  n01440764/n01440764_10029.JPEG	0
  n01440764/n01440764_10040.JPEG	0
  n01440764/n01440764_10042.JPEG	0
  

2.3 Calibration & Evaluation

To calibrate and then evaluate a calibrated Vision Transformer on ImageNet val, run:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use> \
  --master_port 12345 main.py \
  --calib \
  --name vit \
  --pretrained_dir <checkpoint> \
  --data-path <imagenet-path> \
  --model_type <model-type> \
  --img_size  \
  --num-calib-batch <batch-number> \
  --calib-batchsz <batch-size> \
  --quant-mode <mode> \
  --calibrator <calibrator> \
  --percentile <percentile> \
  --calib-output-path <calib-output-path>

For example, to calibrate the ViT-B_16 with a single GPU: (For calibration, we only support using one single GPU). You can use sh calib.sh for simplicity.

export DATA_DIR=Path to the dataset
export CKPT_DIR=Path to the ViT checkpoints
python -m torch.distributed.launch --nproc_per_node 1 \
    --master_port 12345 main.py \
    --calib \
    --name vit \
    --pretrained_dir $CKPT_DIR/ViT-B_16.npz \
    --data-path $DATA_DIR \
    --model_type ViT-B_16 \
    --img_size 384 \
    --num-calib-batch 20 \
    --calib-batchsz 8 \
    --quant-mode ft2 \
    --calibrator percentile \
    --percentile 99.99 \
    --calib-output-path $CKPT_DIR

Difference between --quant-mode 1 and --quant-mode 2

--quant-mode 1 indicates that all GEMMs are quantized to be INT8-in-INT32-out, while --quant-mode 2 means quantizating all GEMMs to be INT8-in-INT8-out. This is a speed-versus-accuracy trade-off: mode=2 is faster in CUDA implementation but its accuracy is lower.

In order to narrow the accuracy gap for mode=2, QAT is a reasonable choice.

2.4 Quantization Aware Training (QAT)

To run QAT finetuning with a calibrated Vision Transformer, run:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use> \
    --master_port 12345 main.py \
    --train \
    --name vit \
    --pretrained_dir <calibrated-checkpoint> \
    --data-path <imagenet-path> \
    --model_type <model-type> \
    --quant-mode <mode> \
    --img_size  \
    --distill \
    --teacher <uncalibrated-checkpoint> \
    --output <qat-output-path> \
    --quant-mode <mode>\
    --batch-size <batch-size> \
    --num-epochs <num-of-epochs> \
    --qat-lr <learning-rate-of-QAT>

For example, to do QAT with ViT-B_16 by 4 GPU on a single node for 5 epochs, run: (You can see qat.sh for reference.)

export DATA_DIR=Path to the dataset
export CKPT_DIR=Path to the ViT checkpoints
python -m torch.distributed.launch --nproc_per_node 4 \
    --master_port 12345 main.py \
    --train \
    --name vit \
    --pretrained_dir $CKPT_DIR/ViT-B_16_calib.pth \
    --data-path $DATA_DIR \
    --model_type ViT-B_16 \
    --quant-mode ft2 \
    --img_size 384 \
    --distill \
    --teacher $CKPT_DIR/ViT-B_16.npz \
    --output qat_output \
    --quant-mode ft2\
    --batch-size 32 \
    --num-epochs 5 \
    --qat-lr 1e-4

Improvement brought by QAT (under mode=2)

As shown below, the accuracy gap is narrowed down from 2.15% to 0.65%.