• Clone

git clone [email protected]:whlzy/SimHIT.git
cd PR_EXP

• You should modify the path of config/exp_mlp/test_hardswish.yml or config/exp_mlp/test_relu.yml to your local dataset path.

cd config/exp_mlp
cat test_relu.yml
...
...

• You can choose gpu or cpu in the config/exp_mlp/test_hardswish.yml or config/exp_mlp/test_relu.yml.

• You can run the script.

cd ../..
sh scripts/train_hardswish.sh
sh scripts/train_relu.sh

• You can change the net config in the config/exp_mlp/test_hardswish.yml or config/exp_mlp/test_relu.yml.

• You can add some new experiments with just adding new script in scripts and new yaml file in config.

• You can add new dataset in src/data, but maybe need to change some codes.

• You can rewrite a new training code like train_mlp.py using src/runner.py. src/runner.py is a class which assembles partial training process and config process. You just need to use the src/runner.py and rewrite set_data, set_model, train_one_epoch and test_one_epoch like train_mlp.py. Like the train_mlp.py, you can freely modify the network and modify the training process in train_one_epoch.

Note: If you want to use DP:

1. you need add "dp: True" to "basic" in your config.yml file.

Note: If you want to use DDP:

1. you need add "ddp: your port" to "basic" in your config.yml file.
2. please remember to add "if rank == 0:".
3. please remember to modify the sampler in torch.utils.data.DataLoader.
4. you need a slurm environment.

Note: If you don't need dp or ddp, you can just delete the "dp: True" and "ddp: your port".

EXP log is in the exp/*/test_hardswish.

The output.log is logged by mmcv logging in the exp/*/test_hardswish/output.log.

The config you used is written in the exp/*/test_hardswish/config.yaml.

The tensorboard log is in the exp/*/test_hardswish/logdir.

The best checkpoint is in the exp/*/test_hardswish/checkpoint/best/model_best.pth.

• train mlp.

sh scripts/train_hardswish.sh
sh scripts/train_relu.sh

• train alexnet.

sh scripts/train_AlexNet.sh

• train resnet18.

sh scripts/train_resnet18.sh

You need install following packages:

thop
yaml
tqdm
pytorch >= 1.6
skimage <= 0.16.2

This project is released under the Apache 2.0 license.

Our code is partially borrowed from MMCV and IMDN. Thanks Yiting Zhang and Junjing huang for their help.

If you find this project useful in your research, please consider to cite.

@misc{SimHIT,
    title={SimHIT: A Simple Framework for HIT Pattern Recognition Experiment},
    author={Zeyu Lu},
    howpublished = {\url{https://github.com/whlzy/SimHIT}},
    year={2022}
}