Agent Attention: On the Integration of Softmax and Linear Attention

This is a warehouse for Agent-Attention-Models based on pytorch framework, can be used to train your image datasets. The code mainly comes from official source code

Project Structure

├── datasets: Load datasets
    ├── my_dataset.py: Customize reading data sets and define transforms data enhancement methods
    ├── split_data.py: Define the function to read the image dataset and divide the training-set and test-set
    ├── threeaugment.py: Additional data augmentation methods
├── models: UniRepLKNet Model
    ├── agent_cswin.py: Construct "Agent-cswin" models
    ├── agent_deit.py: Construct "Agent-deit" models
    ├── agent_pvt.py: Construct "Agent-pvt" models
    ├── agent_swin.py: Construct "Agent-swin" models
    ├── deit.py: Construct "deit" models
├── util:
    ├── engine.py: Function code for a training/validation process
    ├── losses.py: Knowledge distillation loss, combined with teacher model (if any)
    ├── optimizer.py: Define Sophia optimizer
    ├── samplers.py: Define the parameter of "sampler" in DataLoader
    ├── utils.py: Record various indicator information and output and distributed environment
├── estimate_model.py: Visualized evaluation indicators ROC curve, confusion matrix, classification report, etc.
└── train_gpu.py: Training model startup file

Precautions

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, num_workers and nb_classes parameters. If you want to draw the confusion matrix and ROC curve, you only need to set the predict parameter to True

Use Sophia Optimizer (in util/optimizer.py)

You can use anther optimizer sophia, just need to change the optimizer in train_gpu.py, for this training sample, can achieve better results

# optimizer = create_optimizer(args, model_without_ddp)
optimizer = SophiaG(model.parameters(), lr=2e-4, betas=(0.965, 0.99), rho=0.01, weight_decay=args.weight_decay)

Train this model

Parameters Meaning:

1. nproc_per_node: <The number of GPUs you want to use on each node (machine/server)>
2. CUDA_VISIBLE_DEVICES: <Specify the index of the GPU corresponding to a single node (machine/server) (starting from 0)>
3. nnodes: <number of nodes (machine/server)>
4. node_rank: <node (machine/server) serial number>
5. master_addr: <master node (machine/server) IP address>
6. master_port: <master node (machine/server) port number>

Note:

If you want to use multiple GPU for training, whether it is a single machine with multiple GPUs or multiple machines with multiple GPUs, each GPU will divide the batch_size equally. For example, batch_size=4 in my train_gpu.py. If I want to use 2 GPUs for training, it means that the batch_size on each GPU is 4. Do not let batch_size=1 on each GPU, otherwise BN layer maybe report an error. If you recive an error like "unrecognized arguments: --local-rank=1___" when you use distributed multi-GPUs training, just replace the command "torch.distributed.launch" to "torch.distributed.run".

train model with single-machine single-GPU：

python train_gpu.py

train model with single-machine multi-GPU：

python -m torch.distributed.launch --nproc_per_node=8 train_gpu.py

train model with single-machine multi-GPU:

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

CUDA_VISIBLE_DEVICES=1,3 python -m torch.distributed.launch --nproc_per_node=2 train_gpu.py

train model with multi-machine multi-GPU:

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

Citation

@article{han2023agent,
  title={Agent Attention: On the Integration of Softmax and Linear Attention},
  author={Han, Dongchen and Ye, Tianzhu and Han, Yizeng and Xia, Zhuofan and Song, Shiji and Huang, Gao},
  journal={arXiv preprint arXiv:2312.08874},
  year={2023}
}