Parcollet Titouan

icon indicating a website link www.darnault-parcollet.fr icon indicating an email [email protected]

icon company Samsung AI Cambridge icon location United-Kingdom icon location Research Scientist at Samsung AI Cambridge — Affiliated Lecturer at the University of Cambridge — Associate Professor on leave at the University of Avignon

Results 231 comments of Parcollet Titouan

Add type checking with PyRight to CI (from Samsung AI Center Cambridge)

@pplantinga will most likely be as interested as I am into that. It will take a good number of PR (Rogier is on it) to fix all warnings but he...

Weights only load failed. In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`.

Hum, I could not reproduce the issue locally, that is quite surprising. Is this with the latest develop branch commit? But I can see what can cause this, let me...

High Training Loss and WER in Transducer Recipe on LibriSpeech 960hrs Dataset

The problem with these issues is that we don't know why it happens and can't reproduce locally (I tried, it trains). One thing I notice here is that the number...

High Training Loss and WER in Transducer Recipe on LibriSpeech 960hrs Dataset

Hi @nutan235, it seems that you are not training on the 960 hours here. If you really ran this on one GPU, then it means that you have 150 *...

High Training Loss and WER in Transducer Recipe on LibriSpeech 960hrs Dataset

Hi @nutan235 , happy to see this. May I ask if your dynamic_batching is set to True or False? We do not recommend setting it to False for most recipe...

Solves #2846 Refactor run_opts into a @dataclass

@pplantinga / @Adel-Moumen if agreement from both of you, please merge!

Positional Rotary Embeddings for transformers

@shucongzhang will take care of that.

large CER

Hi, I am not sure about what you are trying to achieve. You can perform only testing with the --test_only flag, no need for commenting. The code for CER calculation...

large CER

Right, first thing to do then, have a look at our [inference tutorial](https://speechbrain.github.io/tutorial_advanced.html) that will give you more information about what you are doing and trying to do. For the...

large CER

Well it looks like your model hasn't been trained (i.e. no checkpoints to load)? Could you look into the results folder? There should be a save folder with CKPT-{smth} in...