Deit-Pytorch

This is the DeiT model code warehouse, which mainly includes models: DeiT, DeiTV2, ResMLP, CaiT, patchconvnet-models. The code is mainly derived from the official source code of facebookresearch, and has been modified based on it. Now it can be applied to your own image classification datasets.

Precautions

<1>Before you use the code to train your own data set, please first enter the train_gpu.py file and modify the data_root, batch_size and nb_classes parameters. If you want to draw the confusion matrix and ROC curve, you only need to remove the comments of Plot_ROC and Predictor at the end of the code. The comment of the function is enough, and the third parameter can be changed to the path of your own model weights file(.pth).

<2>If you want to use another model, import it in the train_gpu.py file, then find the following code and replace the name of model function.

model = deit_tiny_patch16_224(pretrained=False,
                              num_classes=args.nb_classes,
                              drop_rate=args.drop,
                              drop_path_rate=args.drop_path,
                              img_size=args.input_size
                              )

Train this model

train model with single-machine single-card：

python train_gpu.py

train model with single-machine multi-card：

torchrun --nproc_per_node=8 train_gpu.py

train model with single-machine multi-card:

(using a specified part of the cards: for example, I want to use the second and fourth cards)

CUDA_VISIBLE_DEVICES=1,3 torchrun --nproc_per_node=2 train_gpu.py

train model with multi-machine multi-card:

(For the specific number of GPUs on each machine, modify the value of --nproc_per_node. If you want to specify a certain card, just add CUDA_VISIBLE_DEVICES= to specify the index number of the card before each command. The principle is the same as single-machine multi-card training)

On the first machine: python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=0 --master_addr=<Master node IP address> --master_port=<Master node port number> train_gpu.py

On the second machine: python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=1 --master_addr=<Master node IP address> --master_port=<Master node port number> train_gpu.py

Paper

@InProceedings{pmlr-v139-touvron21a,
  title =     {Training data-efficient image transformers & distillation through attention},
  author =    {Touvron, Hugo and Cord, Matthieu and Douze, Matthijs and Massa, Francisco and Sablayrolles, Alexandre and Jegou, Herve},
  booktitle = {International Conference on Machine Learning},
  pages =     {10347--10357},
  year =      {2021},
  volume =    {139},
  month =     {July}
}