Shu-wen Yang comments

Results 96 comments of


                                            Shu-wen Yang

TypeError: 'builtin_function_or_method' object does not support item assignment

Hi, can you try the latest way of using S3PRL features? https://s3prl.github.io/s3prl/tutorial/upstream_collection.html

Incorrect encoding detected in ./downstream/asr/char.dict, please rebuild the dataset

Could you try to re-install fairseq with `fairseq@git+https://github.com//pytorch/fairseq.git@b5a039c292facba9c73f59ff34621ec131d82341#egg=fairseq`

Incorrect encoding detected in ./downstream/asr/char.dict, please rebuild the dataset

I find the latest master occasionally remove the fairseq dependency information, might lead you to install the latest fairseq which might not be compatible.

Incorrect encoding detected in ./downstream/asr/char.dict, please rebuild the dataset

Oh, just check out the `b5a039c292facba9c73f59ff34621ec131d82341 ` commit, and maybe run the `pip install -e ./` again.

Incorrect encoding detected in ./downstream/asr/char.dict, please rebuild the dataset

git checkout b5a039c292facba9c73f59ff34621ec131d82341

Incorrect encoding detected in ./downstream/asr/char.dict, please rebuild the dataset

Hmm... I still can not reproduce. Did you try the solutions on other places? (like https://stackoverflow.com/questions/38518023/unicodedecodeerror-utf8-codec-cant-decode-byte-0x80-in-position-3131-invali) I guess this might be an env-specific issue.

Added the save and load features code.

Hi, Thanks for the proposal. I quickly go through it and I have a simple concern that since in the current coding style, not all dataset return the `name` of...

It would be better to save the features if upstream is not trainable

You are right. However, dumping all layers of feature for hubert or wav2vec2 requires huge amount of disk space. Also, I have tried that HDD would be a serious IO...

(WIP) a better version of enhancement and separation downstream

Hey @HuangZiliAndy ! Thanks for the great PR! The performance improvement is huge! I briefly look into it and I found there are quite some logic changes might be hard...

(WIP) a better version of enhancement and separation downstream

Hey @HuangZiliAndy, I can successfully run it. However, I find that you did not include any README in this PR. Would you like to merge it first or add the...