Seeeeeyo/mae

Masked Autoencoders: A PyTorch Implementation

Steps

git clone https://github.com/Seeeeeyo/mae.git

0.1)

Use python 3.8 (should work). I tried only with python 3.10 but had to to the following changes:

In “/usr/local/lib/python3.10/dist-packages/timm/models/layers/helpers.py”, Add

import torch 
TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])

if TORCH_MAJOR == 1 and TORCH_MINOR < 8:
    from torch._six import container_abcs
else:
    import collections.abc as container_abcs

In “/content/mae/util/misc.py”, change from torch._six import inf By from torch import inf

1) 

*Download the data from the drive 'data_60k' (full medmnist dataset) and 'data_sampled.zip' (fractions of medmnist dataset).
*Download a medical classification dataset. I let you choose one, as you know these better than I do. Let's try to find one which 
 is kind of similar to MedMnist to hopefully reach some performances. The data structure should be as follow: 
	- eval_data
		- train 
			-class1
				-img1
				-img2
				-...
			-class2
			-...

cd mae

wget -nc https://dl.fbaipublicfiles.com/mae/finetune/mae_pretrain_vit_base.pth

!pip install submitit
!pip install timm==0.3.2

Evaluate the mae_vit_base on eval_data python main_finetune.py --eval --resume mae_finetuned_vit_base.pth --model vit_base_patch16 --batch_size 32 --data_path 'eval_data'

FINETUNE

python main_finetune.py \
    --accum_iter 1 \
    --batch_size 32 \
    --model vit_base_patch16 \
    --finetune 'mae_pretrain_vit_base.pth' \
    --epochs 50 \
    --blr 5e-4 --layer_decay 0.65 \
    --weight_decay 0.05 --drop_path 0.1 --mixup 0.8 --cutmix 1.0 --reprob 0.25 \
    --dist_eval --data_path 'AGI/data_60k'

OR (couldn't try it without cluster and multiple GPU):

python submitit_finetune.py \
	--job_dir ${JOB_DIR} \
    --nodes 4 \ TO CHANGE 
    --batch_size 32 \ 
    --model vit_base_patch14 \
    --finetune 'mae_pretrain_vit_base.pth' \
    --epochs 50 \
    --blr 1e-3 --layer_decay 0.75 \
    --weight_decay 0.05 --drop_path 0.3 --reprob 0.25 --mixup 0.8 --cutmix 1.0 \
    --dist_eval --data_path 'AGI/data_60k'

OR (if running on 1 node with 8 GPUs. Couldn't try it without cluster and multiple GPU):

MP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 main_finetune.py \
    --accum_iter 4 \
    --batch_size 32 \
    --model vit_base_patch16 \
    --finetune 'mae_pretrain_vit_base.pth' \
    --epochs 50 \
    --blr 5e-4 --layer_decay 0.65 \
    --weight_decay 0.05 --drop_path 0.1 --mixup 0.8 --cutmix 1.0 --reprob 0.25 \
    --dist_eval --data_path 'AGI/data_60k'

model_path = 'mae/output_dir/checkpoint-49.pth'

Evaluate the finetune mae_vit_base on eval_data

python main_finetune.py --eval --resume {model_path} --model vit_base_patch16 --batch_size TODO --data_path 'eval_data'

Repeat 6) and 7) for the different dataset sizes ('data_sampled_6k', 'data_sampled_36k', 'data_sampled_600').

IF NEEDED

9)a) In case the results are shitty, we might need to pretrain the model on MedMnist and then finetune on the eval_data.

python submitit_pretrain.py \
    --job_dir ${JOB_DIR} \
    --nodes 8 \
    --use_volta32 \
    --batch_size 64 \
    --model mae_vit_large_patch16 \
    --norm_pix_loss \
    --mask_ratio 0.75 \
    --epochs 800 \
    --warmup_epochs 40 \
    --blr 1.5e-4 --weight_decay 0.05 \
    --data_path 'AGI/data_60k'

9)b)

python main_finetune.py \
    --accum_iter 1 \
    --batch_size 32 \
    --model vit_base_patch16 \
    --finetune 'mae_pretrain_vit_base.pth' \
    --epochs 50 \
    --blr 5e-4 --layer_decay 0.65 \
    --weight_decay 0.05 --drop_path 0.1 --mixup 0.8 --cutmix 1.0 --reprob 0.25 \
    --dist_eval --data_path 'eval_data'

This is a PyTorch/GPU re-implementation of the paper Masked Autoencoders Are Scalable Vision Learners:

@Article{MaskedAutoencoders2021,
  author  = {Kaiming He and Xinlei Chen and Saining Xie and Yanghao Li and Piotr Doll{\'a}r and Ross Girshick},
  journal = {arXiv:2111.06377},
  title   = {Masked Autoencoders Are Scalable Vision Learners},
  year    = {2021},
}

The original implementation was in TensorFlow+TPU. This re-implementation is in PyTorch+GPU.
This repo is a modification on the DeiT repo. Installation and preparation follow that repo.
This repo is based on timm==0.3.2, for which a fix is needed to work with PyTorch 1.8.1+.

Catalog

Visualization demo
Pre-trained checkpoints + fine-tuning code
Pre-training code

Visualization demo

Run our interactive visualization demo using Colab notebook (no GPU needed):

Fine-tuning with pre-trained checkpoints

The following table provides the pre-trained checkpoints used in the paper, converted from TF/TPU to PT/GPU:

	ViT-Base	ViT-Large	ViT-Huge
pre-trained checkpoint	download	download	download
md5	`8cad7c`	`b8b06e`	`9bdbb0`

The fine-tuning instruction is in FINETUNE.md.

By fine-tuning these pre-trained models, we rank #1 in these classification tasks (detailed in the paper):

	ViT-B	ViT-L	ViT-H	ViT-H₄₄₈	prev best
ImageNet-1K (no external data)	83.6	85.9	86.9	87.8	87.1
following are evaluation of the same model weights (fine-tuned in original ImageNet-1K):
ImageNet-Corruption (error rate)	51.7	41.8	33.8	36.8	42.5
ImageNet-Adversarial	35.9	57.1	68.2	76.7	35.8
ImageNet-Rendition	48.3	59.9	64.4	66.5	48.7
ImageNet-Sketch	34.5	45.3	49.6	50.9	36.0
following are transfer learning by fine-tuning the pre-trained MAE on the target dataset:
iNaturalists 2017	70.5	75.7	79.3	83.4	75.4
iNaturalists 2018	75.4	80.1	83.0	86.8	81.2
iNaturalists 2019	80.5	83.4	85.7	88.3	84.1
Places205	63.9	65.8	65.9	66.8	66.0
Places365	57.9	59.4	59.8	60.3	58.0

Pre-training

The pre-training instruction is in PRETRAIN.md.

License

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.

Seeeeeyo / mae