WIP - Rough Draft of README #157
Conversation
rbracco
commented
Aug 12, 2020
rbracco
commented
- Updates the README to explain the expected inputs, outputs of CTCDecode, as well as their shapes.
- Adds an installation command that works on Google Colab
- Adds a resources section for people less familiar with Connectionist Temporal Classification (CTC) and Beam Search.
|
Epic! I see your TODOs, will fill them in tomorrow |
… duplicate install instructions, formatted usage example
|
This is so awesome, thanks so much for doing this. I've added some changes on top of your branch, could you have a look? I took a lot of things from the original code: https://github.com/PaddlePaddle/DeepSpeech/blob/develop/decoders/swig_wrapper.py As well as some descriptions that I've made. The only one I'm still a bit unsure (but is an important param) is the beta parameter. I'll try dig up a better description but I think currently it's better than nothing! |
|
Thank you for the help. Those changes look great and clarified things a lot. I showed this to a friend and they suggested adding the parameter descriptions as a docstring for easier usage. Let me know which option you prefer from A. leaving things as they are B. Add parameters to docstring and leave them in the README. C. Add parameters to the docstring and remove them from the README. Also are there any other things you'd like to see in the README? I've done some timing tests on various lengths of audio and beam sizes. I could include the results of that as well. The one variable I didn't consider was the vocab size. Please let me know your thoughts and thanks again.
|
|
100% agreed, we could add docstrings here and I think it would be very beneficial: https://github.com/parlance/ctcdecode/blob/master/ctcdecode/__init__.py#L6 Is this something you can tackle? And for option A/B or C, which would you think makes sense? That graph is very informative, and something useful for sure! Feel free to add additional information :) |
|
Thanks for the input. I added the docstrings but wasn't entirely sure about the format. I ended up deciding to keep the docstrings shorter and drier and leave the readme longer. I think I'll leave the timing stuff out for now. Is there anything else you would like to see added? |
|
LGTM! thanks so much @rbracco :) |
|
Realised there was one TODO in their, will update and make a PR |
