sungfeng-huang / ssl-pretraining-separation Goto Github PK

Hi, @SungFeng-Huang! Awesome work, still top 3 SOTA even after two years :)
One question, may I ask for the hugging face link to the final trained models, please? Would be enough with the Libri2Mix PT-FT ConvTasNet.

Hey there, @SungFeng-Huang! Hoping you are all good. :)

We were trying your repo for fine-tuning ConvTasNet, which can be achieved by setting the --strategy = 'pretrained'. Looking at the train_general.py, there is the following snippet of code:

if known_args.strategy == "pretrained":
    parser.add_argument("--load_path", default=None, required=True, help="Checkpoint path to load for fine-tuning.")

1. Here, --load_path is not being used anywhere else in the code. Hence, the question is: how exactly fine-tuning is carried out? Ultimately, we couldn't tell if there is any difference between from_scratch and pretrained from the train_general.py code.

2. We also tried to find out whether you perform a freeze of some layers of a pre-trained ConvTasNet model before fine-tuning, but there is no such thing. Based on the paper, it seems that the ConvTasNet-trained model is loaded, and then --all-- these weights are learned again during fine-tuning. Is this correct?

Hope you can clarify our doubts, please! Thanks for your attention.

hi,
I tried to install requirements in anaconda with python=3.7.

But it went wrong.
ERROR: Could not find a version that satisfies the requirement torchaudio==0.8.0
ERROR: No matching distribution found for torchaudio==0.8.0

Is it must install torchaudio=0.8.0? Or is there any version to replace it?