Yuecong Min

Thank you for your attention. We aim to maintain simplicity in the released code, and this specific section primarily enhances boundary accuracy. You can implement GSBA by extending the pseudo...

@ethio-artifical You just need to comment `model.cuda()` line in [main.py](https://github.com/ycmin95/VAC_CSLR/blob/main/main.py#L118), and modify the pre-trained load process in [main.py](https://github.com/ycmin95/VAC_CSLR/blob/main/main.py#L122) to `state_dict = torch.load(weight_path, map_location=torch.device('cpu'))`. The command to evaluate the pre-trained model...

You can check whether you successfully install ctcdecode.

As mentioned in README, you can also try the pytorch version of CTC.

You can also try [decoder provided in tensorflow](https://www.tensorflow.org/api_docs/python/tf/nn/ctc_beam_search_decoder), which is also adopted by [recent work](https://github.com/FangyunWei/SLRT/blob/67ddead91f66a2757378a43a59294b9e5e5d6ba1/TwoStreamNetwork/modelling/recognition.py#L16).

Hi, the loss with in each epoch is recorded in the [loss_value](https://github.com/ycmin95/VAC_CSLR/blob/main/seq_scripts.py#L32) during training, you can maintain a global recorder to plot epoch vs loss.

The released code is based on Phoenix14, you can modify the data format according to your requirements.

Hi, I'd like to clarify that we do not currently utilize SGML files to visualize the alignment. However, you may refer to the discussion about the alignment in [this issue](https://github.com/ycmin95/VAC_CSLR/issues/4)....

Hi, I have not try this in windows system, perhaps you can find some useful information in [relevant issues here](https://github.com/parlance/ctcdecode/issues?q=windows).

Hi, all experiments are conducted on local severs and I am not familiar with google colab. You can find some tutorials about google colab first, and then try to modify...