About wavlm features
Kanraaaaa opened this issue · comments
Hi, thanks a lot for sharing your work. But I have met a problem about extracting wavlm feature.
I tried to extract query_feats, matching_set, but features of the last 2 layers (23, 24) are always NaN. My code is as follows:
test_utt = 'pretrained/M4_遇见.wav'
ref_utt = 'pretrained/M26.wav'
wavlm_encoder = WavLMEncoder(ckpt_path='pretrained/WavLM-Large.pt', device=device)
query_feats = wavlm_encoder.get_features(test_utt, weights=applied_weights)
matching_set = wavlm_encoder.get_features(ref_utt, weights=applied_weights)
synth_set = wavlm_encoder.get_features(ref_utt, output_layer=SPEAKER_INFORMATION_LAYER)
test_utt and ref_utt are downloaded from demo page.
Have you ever met this problem?
Hello, I tried extracting the WavLM features directly from the two audio samples on the demo page and did not encounter any NaN values. I suggest checking if there is any issue with the WavLM-Large checkpoint. You can obtain it directly from this link: WavLM-Large.pt.
Close this issue now, please feel free to reopen it if you have further questions.