Peter Plantinga
Peter Plantinga
The same speed perturbation is applied to the entire batch, so the relative lengths of each item should remain the same.
Its possible there is code in the toolkit that assumes the longest element of the batch has relative length of `1` but I see nothing in `make_masks` that assumes this....
How might I go about mixing teacher forcing with feed previous with code from tutorial #3 ?
I'm working on this, but I've been traveling this week and last. I hope to have a reasonable PR by the end of next week
Superseded by #2563
tl;dr this is not possible with the current way this is structured. Longer answer: There's no general way to extract the arguments used to construct an object. On the other...
The issue @weiwei-ww has brought up is actually a separate issue. In this case, what is desired is the dereference to happen before the include. In fact, the `!include` was...
I tried this recipe with a `peft` layer and it *just worked* to my amazement. Here's the exact change I made: ``` whisper: !new:speechbrain.nnet.adapters.AdaptedModel model_to_adapt: !ref - adapter_class: !name:speechbrain.nnet.adapters.LoRA -...
> @pplantinga have you tested it with pretrainer using for interfaces? I tested this and it worked, but had warnings due to loading only trained params. I have fixed this...
One epoch (100h) results for Whisper Small.en, published results are `test-clean=3.05` and `test-other=7.53`: ``` speechbrain.utils.train_logger - Epoch loaded: 1 - test loss: 9.73e-01, test CER: 1.03, test WER: 2.81 speechbrain.utils.train_logger...