Andy T. Liu
Andy T. Liu
Hello, I believe you can fit Common Voice to our downstream ASR dataset by satisfying this requirement: https://github.com/s3prl/s3prl/blob/540a4f86e9f099b240c314a56ff80b069e91d4cd/s3prl/downstream/asr/dataset.py#L46-L47 The usage is pretty straightforward, you just have to define the sub-directory...
> I use checkpoints trained recently with downstream tasks. The directory of downstream experiment looks like: > > And I use any of `states-200000.ckpt`, `best-states-test.ckpt`, `best-states-dev.ckpt`. > Such checkpoint has...
Hi @myhrbeu, I'm not sure if I understand your question correctly, but you can change the number of layers by changing this line: https://github.com/s3prl/s3prl/blob/e52439edaeb1a443e82960e6401ae6ab4241def6/s3prl/pretrain/audio_albert/config_model.yaml#L4 For example: ```yaml num_hidden_layers: 7 ```...
> I found it is very slow when having 7 layers This is expected and normal. Because sharing layers (sharing weights across layers) won't make the model forward faster. you...
Hello, Apparently, #L463 needs to be fixed. However, our repo was not using #L463. This is how we prepare the transformer input: https://github.com/s3prl/s3prl/blob/ccf621e78edb80534ba32a86b5f7077f84b9a6fd/s3prl/upstream/mockingjay/builder.py#L178-L189 There are some more details that you...
and we will fix [#L463](https://github.com/s3prl/s3prl/blob/ccf621e78edb80534ba32a86b5f7077f84b9a6fd/s3prl/upstream/mockingjay/model.py#L463) in our next release!
Hello! That sounds nice, we will be supporting validation loss for pertaining in our next version. It has been implemented but is not yet ready to release, because we have...
Following from issue https://github.com/s3prl/s3prl/issues/342, it looks like you have modified our code, for: 1) Our [dataset](https://github.com/s3prl/s3prl/blob/master/s3prl/pretrain/bucket_dataset.py) does bucketing, so that in each batch all the utterances have the same length...
> I think this problem was caused by the old version because I downloaded s3prl last October, and only modified some interesting parts. > Further, in the last issue #342,...
> So the pre-trained model is performed until `1,500` sequences, right? This is partially correct. We will randomly sample a sub-sequence from the whole sequence: https://github.com/s3prl/s3prl/blob/6e4d75b4e149f662bb419a154109f0ac663f958d/s3prl/pretrain/bucket_dataset.py#L80-L84 so it is not...