BEiT: BERT Pre-Training of Image Transformers

Official PyTorch implementation and pretrained models of BEiT.

The code and pretrained models of BEiT v2 can be found at here.

The code and pretrained models of BEiT-3 can be found at here.

• March, 2023: release the code and pretrained models of BEiT-3
• March, 2023: BEiT-3 was accepted by CVPR 2023.
• Sept 2022: release the code and pretrained models of BEiT v2
• Aug 2022: release preprint Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks
• Aug 2022: release preprint BEiT v2: Masked Image Modeling with Vector-Quantized Visual Tokenizers
• June 2022: release preprint VL-BEiT: Generative Vision-Language Pretraining
• March, 2022: add linear probe examples
• January, 2022: BEiT was accepted by ICLR 2022 as Oral presentation (54 out of 3391).
• August 2021: BEiT is on HuggingFace
• July 2021: BEiT-large achieves state-of-the-art results on ADE20K (a big jump to 57.0 mIoU) for semantic segmentation.
• July 2021: BEiT-large achieves state-of-the-art ImageNet top-1 accuracy (88.6%) under the setting without extra data other than ImageNet-22k.
• July 2021: release the code and pretrained models of BEiT
• June 2021: release preprint BEiT: BERT Pre-Training of Image Transformers

Pretrained models

We provide four BEiT weights pretrained on ImageNet-22k. The models were pretrained with 224x224 resolution.

• BEiT-base: #layer=12; hidden=768; FFN factor=4x; #head=12; patch=16x16 (#parameters: 86M)
• BEiT-large: #layer=24; hidden=1024; FFN factor=4x; #head=16; patch=16x16 (#parameters: 304M)

Download checkpoints that are self-supervised pretrained and then intermediate fine-tuned on ImageNet-22k (recommended):

• BEiT-base: beit_base_patch16_224_pt22k_ft22k
• BEiT-large: beit_large_patch16_224_pt22k_ft22k

Download checkpoints that are self-supervised pretrained on ImageNet-22k:

• BEiT-base: beit_base_patch16_224_pt22k
• BEiT-large: beit_large_patch16_224_pt22k

Setup

alias=`whoami | cut -d'.' -f2`; docker run -it --rm --runtime=nvidia --ipc=host --privileged -v /home/${alias}:/home/${alias} pytorch/pytorch:1.7.1-cuda11.0-cudnn8-devel bash

First, clone the repo and install required packages:

git clone https://github.com/microsoft/unilm.git
cd unilm/beit
pip install -r requirements.txt

The required packages including: Pytorch version 1.7.1, torchvision version 0.8.2 and Timm version 0.3.2, etc.

For mixed-precision training, please install apex

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Fine-tuning on ImageNet-1k (image classification)

We summarize the validation results as follows. We also provide the fine-tuned weights and fine-tuning logs. The detailed instructions to reproduce the results can be found at get_started_for_image_classification.md.

Fine-tuning on ADE20K (semantic segmentation)

We summarize the validation results as follows. We also provide the fine-tuned weights and fine-tuning logs. The detailed instructions to reproduce the results can be found at semantic_segmentation/README.md.

Example: Pre-training BEiT-base on ImageNet-22k

The BEiT-base model can be pretrained on ImageNet-22k using a DGX-2 box (16 V100-32GB):

# Set the path to save checkpoints
OUTPUT_DIR=/path/to/save/your_model
# Download and extract ImageNet-22k
DATA_PATH=/path/to/imagenet22k
# Download the tokenizer weight from OpenAI's DALL-E
TOKENIZER_PATH=/path/to/save/dall_e_tokenizer_weight
mkdir -p $TOKENIZER_PATH
wget -o $TOKENIZER_PATH/encoder.pkl https://cdn.openai.com/dall-e/encoder.pkl
wget -o $TOKENIZER_PATH/decoder.pkl https://cdn.openai.com/dall-e/decoder.pkl

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=16 run_beit_pretraining.py \
        --data_path ${DATA_PATH} --output_dir ${OUTPUT_DIR} --num_mask_patches 75 \
        --model beit_base_patch16_224_8k_vocab --discrete_vae_weight_path ${TOKENIZER_PATH} \
        --batch_size 128 --lr 1.5e-3 --warmup_steps 10000 --epochs 150 \
        --clip_grad 3.0 --drop_path 0.1 --layer_scale_init_value 0.1

• --num_mask_patches: number of the input patches need be masked.
• --batch_size: batch size per GPU.
• Effective batch size = number of GPUs * --batch_size. So in the above example, the effective batch size is 128*16 = 2048.
• --lr: learning rate.
• --warmup_steps: learning rate warmup steps.
• --epochs: total pre-training epochs.
• --clip_grad: clip gradient norm.
• --drop_path: stochastic depth rate.
• --imagenet_default_mean_and_std: enable this for ImageNet-1k pre-training, i.e., (0.485, 0.456, 0.406) for mean and (0.229, 0.224, 0.225) for std. We use (0.5, 0.5, 0.5) for mean and (0.5, 0.5, 0.5) for std by default on other pre-training data.
• --layer_scale_init_value: 0.1 for base, 1e-5 for large, set 0 to disable layerscale.

Example: Pre-training BEiT-base on ImageNet-1k

The BEiT-base model can be pretrained on ImageNet-1k using a DGX-2 box (16 V100-32GB):

# Set the path to save checkpoints
OUTPUT_DIR=/path/to/save/your_model
# Download and extract ImageNet-1k
DATA_PATH=/path/to/imagenet1k_train_set
# Download the tokenizer weight from OpenAI's DALL-E
TOKENIZER_PATH=/path/to/save/dall_e_tokenizer_weight
mkdir -p $TOKENIZER_PATH
wget -o $TOKENIZER_PATH/encoder.pkl https://conversationhub.blob.core.windows.net/beit-share-public/dall-e_vae/encoder.pkl
wget -o $TOKENIZER_PATH/decoder.pkl https://conversationhub.blob.core.windows.net/beit-share-public/dall-e_vae/decoder.pkl

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=16 run_beit_pretraining.py \
        --data_path ${DATA_PATH} --output_dir ${OUTPUT_DIR} --num_mask_patches 75 \
        --model beit_base_patch16_224_8k_vocab --discrete_vae_weight_path ${TOKENIZER_PATH} \
        --batch_size 128 --lr 1.5e-3 --warmup_epochs 10 --epochs 800 \
        --clip_grad 3.0 --drop_path 0.1 --layer_scale_init_value 0.1 \
        --imagenet_default_mean_and_std

Example: Fine-tuning BEiT on ImageNet-22k

The BEiT-large model can be fine-tuned on ImageNet-22k using a DGX-2 box (16 V100-32GB):

# Set the path to save checkpoints
OUTPUT_DIR=/path/to/save/your_model
# Download and extract ImageNet-22k
DATA_PATH=/path/to/imagenet22k

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=16 run_class_finetuning.py \
    --model beit_large_patch16_224 --data_path $DATA_PATH \
    --nb_classes 21841 --data_set image_folder --disable_eval_during_finetuning \
    --finetune https://github.com/addf400/files/releases/download/v1.0/beit_large_patch16_224_pt22k.pth \
    --output_dir $OUTPUT_DIR --batch_size 64 --lr 2e-3 --update_freq 2 \
    --warmup_epochs 5 --epochs 90 --layer_decay 0.75 --drop_path 0.2 \
    --weight_decay 0.05 --enable_deepspeed --layer_scale_init_value 1e-5 --clip_grad 1.0

• --batch_size: batch size per GPU.
• Effective batch size = number of GPUs * --batch_size * --update_freq. So in the above example, the effective batch size is 16*64*2 = 2048.
• --lr: learning rate.
• --warmup_epochs: learning rate warmup epochs.
• --epochs: total pre-training epochs.
• --clip_grad: clip gradient norm.
• --drop_path: stochastic depth rate.
• --layer_scale_init_value: 0.1 for base, 1e-5 for large, set 0 to disable layerscale.

The BEiT-base can be fine-tuned on ImageNet-22k as follows:

# Set the path to save checkpoints
OUTPUT_DIR=/path/to/save/your_model
# Download and extract ImageNet-22k
DATA_PATH=/path/to/imagenet22k

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=16 run_class_finetuning.py \
    --model beit_base_patch16_224 --data_path $DATA_PATH \
    --nb_classes 21841 --data_set image_folder --disable_eval_during_finetuning \
    --finetune https://github.com/addf400/files/releases/download/v1.0/beit_base_patch16_224_pt22k.pth \
    --output_dir $OUTPUT_DIR --batch_size 256 --lr 3e-3 --update_freq 1 \
    --warmup_epochs 5 --epochs 90 --layer_decay 0.65 --drop_path 0.2 \
    --weight_decay 0.05 --enable_deepspeed --layer_scale_init_value 0.1 --clip_grad 3.0

Code for Analysis of Self-Attention Map

Pre-trained BEiT_base_patch16_224 on ImageNet-1k with 800 epochs, config: --disable_rel_pos_bias --abs_pos_emb --layer_scale_init_value 0

Code grouped in BEiTv2 Repo

If you find this repository useful, please consider citing our work:

@inproceedings{beit,
title={{BEiT}: {BERT} Pre-Training of Image Transformers},
author={Hangbo Bao and Li Dong and Songhao Piao and Furu Wei},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=p-BhZSz59o4}
}

Acknowledgement

This repository is built using the timm library, the DeiT repository and the Dino repository.

License

This project is licensed under the license found in the LICENSE file in the root directory of this source tree.

Microsoft Open Source Code of Conduct

Contact Information

For help or issues using BEiT models, please submit a GitHub issue.

For other communications, please contact Li Dong ([email protected]), Furu Wei ([email protected]).