search

ncarraz / AFILM

MLSP 2021 - Self-Attention for Audio Super-resolution - Keras implementation
MIT License
28 stars 8 forks source link

Pretrained model disappear #6

Closed yxlu-0102 closed 1 month ago

yxlu-0102 commented 10 months ago

I can't find the pre-trained model through the link you provided, can you check it out and re-upload them?

ncarraz commented 10 months ago

Unfortunately, the account associated with the google drive containing the model files has been deleted. We will re-train the models and re-upload them by the end of the year.

yxlu-0102 commented 10 months ago

I installed tensorflow-gpu=2.6.0 and keras=2.6.0 and ran in to this problem AttributeError: 'MaxPooling1D' object has no attribute '_metrics_lock'.

ncarraz commented 10 months ago

I recovered these files: VCTK-scale_2-Multi: https://drive.google.com/file/d/1F9WCHI-muWYQy2n1MpgAEtGKSIRUe3Go/view?usp=share_link, VCTK-scale_4-Multi: https://drive.google.com/file/d/1-5JfveMwPWS-3C2V8pJx7QJ4PzG24N55/view?usp=share_link

yxlu-0102 commented 10 months ago

Thanks a million!

On Fri, 10 Nov 2023 at 1:53 AM, ncarraz @.***> wrote:

I recovered these files: VCTK-scale_2-Multi: https://drive.google.com/file/d/1F9WCHI-muWYQy2n1MpgAEtGKSIRUe3Go/view?usp=share_link, VCTK-scale_4-Multi: https://drive.google.com/file/d/1-5JfveMwPWS-3C2V8pJx7QJ4PzG24N55/view?usp=share_link

— Reply to this email directly, view it on GitHub https://github.com/ncarraz/AFILM/issues/6#issuecomment-1804292565, or unsubscribe https://github.com/notifications/unsubscribe-auth/ATTJFEFV4DTLM5MYZAPNDELYDUJ75AVCNFSM6AAAAAA7EEAA3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBUGI4TENJWGU . You are receiving this because you authored the thread.Message ID: @.***>

yxlu-0102 commented 10 months ago

If I want to train your model, how many epochs should I set? I find the results of 20 epochs seems not that good.

yxlu-0102 commented 10 months ago

I recovered these files: VCTK-scale_2-Multi: https://drive.google.com/file/d/1F9WCHI-muWYQy2n1MpgAEtGKSIRUe3Go/view?usp=share_link, VCTK-scale_4-Multi: https://drive.google.com/file/d/1-5JfveMwPWS-3C2V8pJx7QJ4PzG24N55/view?usp=share_link

I found the spectrogram of the waveforms generated from the ckpt you provided have something run, this is because the audio-super-res use x_lr = np.array(x_hr[0::args.r]) instead of x_lr = decimate(x_hr, args.r) to downsample the x_hr by default, you should point it out in your test.py.

ncarraz commented 10 months ago

The models were trained for 50 epochs. Thank you for pointing it out.