Awesome

ASFormer: Transformer for Action Segmentation

This repo provides training & inference code for BMVC 2021 paper: ASFormer: Transformer for Action Segmentation

Update

Thanks to @ddz16 . There is a small error in the code (L72, model.py), but it does not affect the main conclusions of this paper or the performance, see #2. In order to load the pre training model of the paper, we do not update the repo to fix the bug. However, we still suggest to make a manual change after downloading the code.

(L72, model.py) window_mask[:, :, i:i+self.bl] = 1  change to window_mask[:, i, i:i+self.bl] = 1

Enviroment

Pytorch == 1.1.0, torchvision == 0.3.0, python == 3.6, CUDA=10.1

Reproduce our results

1. Download the dataset data.zip at (https://mega.nz/#!O6wXlSTS!wcEoDT4Ctq5HRq_hV-aWeVF1_JB3cacQBQqOLjCIbc8) or (https://zenodo.org/record/3625992#.Xiv9jGhKhPY). 
2. Unzip the data.zip file to the current folder. There are three datasets in the ./data folder, i.e. ./data/breakfast, ./data/50salads, ./data/gtea
3. Download the pre-trained models at (https://pan.baidu.com/s/1zf-d-7eYqK-IxroBKTxDfg) or (https://drive.google.com/file/d/1xNykN3vXMHCpHIYT0eb5ZHnKSu3Y2K8r/view?usp=sharing). There are pretrained models for three datasets, i.e. ./models/50salads, ./models/breakfast, ./models/gtea
4. Run python main.py --action=predict --dataset=50salads/gtea/breakfast --split=1/2/3/4/5 to generate predicted results for each split.
5. Run python eval.py --dataset=50salads/gtea/breakfast --split=0/1/2/3/4/5 to evaluate the performance. **NOTE**: split=0 will evaulate the average results for all splits, It needs to be done after you complete all split predictions.

Train your own model

Also, you can retrain the model by yourself with following command.

python main.py --action=train --dataset=50salads/gtea/breakfast --split=1/2/3/4/5

The training process is very stable in our experiments. It convergences very fast and is not sensitive to the number of training epochs.

Demo for using ASFormer as your backbone

In our paper, we replace the original TCN-based backbone model MS-TCN in ASRF with our ASFormer. The new model achieves even higher results on the 50salads dataset than the original ASRF. Code is Here.

If you find our repo useful, please give us a star and cite

@inproceedings{chinayi_ASformer,  
	author={Fangqiu Yi and Hongyu Wen and Tingting Jiang}, 
	booktitle={The British Machine Vision Conference (BMVC)},   
	title={ASFormer: Transformer for Action Segmentation},
	year={2021},  
}

Feel free to raise a issue if you got trouble with our code.