Yuecong Min

Thank you for your interest in our work. Currently, I do not have any plan to provide a step-by-step tutorial, but I will consider doing so in the future (maybe...

You can check whether this file "dev.corpus.csv" exists, which is the annotation file of phienix2014 dataset.

Thank you for pointing out this typo and it has been corrected. The default configuration file use [two GPUs](https://github.com/ycmin95/VAC_CSLR/blob/main/configs/baseline.yaml#L11). You can modify it according to your actual situation. For example,...

Hi, @kido1412y2y，[similar issue](https://github.com/ycmin95/VAC_CSLR/issues/16) here, you can check whether processed images are found.

During preprocessing, frames are extracted from video, the glob function here is adopted to load frames. When it fails (the path is wrong or the frames are not extracted), the...

Perhaps you need to check the permission of `features` and `features/fullFrame-256x256px` directories and there is no need to run the script with sudo.

Hi, @Dikay1. I have not met this problem before, it seems like a problem of ctcdecoder, you can find some information from [this issue](https://github.com/parlance/ctcdecode/issues/98).

Relevant codes are [here](https://github.com/ycmin95/VAC_CSLR/blob/main/main.py#L169), you can check whether the training date are loaded successfully.

Software: Pytorch 1.8+ and ubuntu 16.04+ Hardware: the training process needs about 20G graphics memory, our experiments are conducted on two Nvidia titan X gpus or a single 3090 gpu....

It depends on the final goal of your project. If you want to obtain a better feature extractor for SLR, it would be better to adopt GPUs with larger memory,...