CAE: Context AutoEncoder for Self-Supervised Representation Learning

<p align="center"> <img src='furnace/CAE.png'> </p>

This is a PyTorch implementation of CAE: Context AutoEncoder for Self-Supervised Representation Learning.

Highlights

• State-of-the-art MIM performance. Results in the paper are successfully reproduced.

Installation

Clone the repo and install required packages.

pip install -r requirements.txt

# 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" ./

Data Preparation

First, download ImageNet-1k from http://image-net.org/.

The directory structure is the standard layout of torchvision's datasets.ImageFolder. The training and validation data are expected to be in the train/ folder and val folder, respectively:

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

Second, download the pretrained tokenizer.

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

Pretraining

Here is an example that pretrains CAE-base on ImageNet-1K with 32 GPUs. Please see scripts/cae_base_800e.sh for complete script.

OMP_NUM_THREADS=1 $PYTHON -m torch.distributed.launch \
  --nproc_per_node=8 \
  tools/run_pretraining.py \
  --data_path ${DATA_PATH} \
  --output_dir ${OUTPUT_DIR} \
  --model cae_base_patch16_224_8k_vocab --discrete_vae_weight_path ${TOKENIZER_PATH} \
  --batch_size 64 --lr 1.5e-3 --warmup_epochs 20 --epochs 800 \
  --clip_grad 3.0 --layer_scale_init_value 0.1 \
  --imagenet_default_mean_and_std \
  --color_jitter 0 \
  --drop_path 0.1 \
  --sincos_pos_emb \
  --mask_generator block \
  --num_mask_patches 98 \
  --decoder_layer_scale_init_value 0.1 \
  --no_auto_resume \
  --save_ckpt_freq 100 \
  --exp_name $my_name \
  --regressor_depth 4 \
  --decoder_depth 4 \
  --align_loss_weight 2

• --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 64*32 = 2048.
• --lr: learning rate.
• --warmup_epochs: learning rate warmup epochs. Warm up [10, 20, 40] epochs for [300, 800, 1600] pretrain epochs respectively.
• --epochs: total pretraining epochs.
• --clip_grad: clip gradient norm.
• --drop_path: stochastic depth rate.
• --imagenet_default_mean_and_std: enable this for ImageNet-1k pretraining, i.e., (0.485, 0.456, 0.406) for mean and (0.229, 0.224, 0.225) for std. For other pretraining data, use (0.5, 0.5, 0.5) for mean and (0.5, 0.5, 0.5) for std by default.
• --layer_scale_init_value: 0.1 for base, 1e-5 for large, set 0 to disable layerscale. We set --decoder_layer_scale_init_value the same as this.
• --sincos_pos_emb: adopt sin-cos positional embedding during pretraining.
• --regressor_depth: length of the regressor.
• --decoder_depth: length of the decoder.
• --align_loss_weight: weight for alignment loss. 2 by default.

Warmup epochs for 300/800/1600 epochs pretraining are 10/20/40.

For CAE-large, please refer to scripts/cae_large_1600e.sh.

Results

Here provides the results of CAE-base/CAE-large for these evaluation tasks:

• Linear probing
• Attentive probing
• Fine-tuning
• Semantic segmentation
• Object detection and instance segmentation

Pretrained weights and logs are available (Google Drive, Baidu Cloud [Code: 4kil]). *: from CAE paper.

| Model | Pretraining data | #Epoch | Linear | Attentive | Fine-tuning | ADE Seg | COCO Det | COCO InstSeg | | ---------- | ---------------- | ------ | ------ | --------- | ----------- | ------- | -------- | ------------ | | MAE-base* | ImageNet-1K | 1600 | 67.8 | 74.2 | 83.6 | 48.1 | 48.4 | 42.6 | | MAE-large* | ImageNet-1K | 1600 | 76.0 | 78.8 | 86.0 | 53.6 | 54.0 | 47.1 | | CAE-base | ImageNet-1K | 300 | 64.5 | 74.0 | 83.6 | 48.1 | 48.3 | 42.7 | | CAE-base | ImageNet-1K | 800 | 68.9 | 75.9 | 83.8 | 49.7 | 49.9 | 43.9 | | CAE-base | ImageNet-1K | 1600 | 70.3 | 77.2 | 83.9 | 50.3 | 50.3 | 44.2 | | CAE-large | ImageNet-1K | 1600 | 77.8 | 81.2 | 86.2 | 54.9 | 54.5 | 47.5 |

Linear Probing

• Please refer to scripts/cae_base_800e.sh (32 GPUs).
• For CAE-large, just replace --model cae_base_patch16_224 with --model cae_large_patch16_224.

Attentive Probing

• Please refer to scripts/cae_base_800e.sh (32 GPUs).
• For CAE-large, just replace --model cae_base_patch16_224 with --model cae_large_patch16_224.

Fine-tuning

• Please refer to scripts/cae_base_finetune.sh (32 GPUs).
• For CAE-large, please refer to scripts/cae_large_finetune.sh (32 GPUs).

Segmentation & Detection

• Please refer to downstream_tasks dir to get started.

Acknowledgement

This repository is built using the BEiT and MMSelfSup, thanks for their open-source code! Thanks also to the CAE authors for their excellent work!

Citation

@article{ContextAutoencoder2022,
  title={Context Autoencoder for Self-Supervised Representation Learning},
  author={Chen, Xiaokang and Ding, Mingyu and Wang, Xiaodi and Xin, Ying and Mo, Shentong and Wang, Yunhao and Han, Shumin and Luo, Ping and Zeng, Gang and Wang, Jingdong},
  journal={arXiv preprint arXiv:2202.03026},
  year={2022}
}