[feature] add EnCodec model embeddings to FAD calculation #23
Conversation
| @@ -32,7 +34,8 @@ def load_audio_task(fname, sample_rate, dtype="float32"): | |||
| wav_data = wav_data / float(2**31) | |||
|
|
|||
| # Convert to mono | |||
| if len(wav_data.shape) > 1: | |||
| assert channels in [1, 2], "channels must be 1 or 2" | |||
There was a problem hiding this comment.
channels here is unspecified, I guess you miss a line before this to calculatechannels from wav_data
There was a problem hiding this comment.
Yes, it's fixed now. It's an input to the function, defaulting to 1 unless specified (as should be done for encodec 48khz)
| @@ -309,8 +360,8 @@ def update(*a): | |||
| for fname in os.listdir(dir): | |||
| res = pool.apply_async( | |||
| load_audio_task, | |||
| args=(os.path.join(dir, fname), self.sample_rate, dtype), | |||
| callback=update | |||
| args=(os.path.join(dir, fname), self.sample_rate, self.channels, dtype), | |||
There was a problem hiding this comment.
I suspect self.channels is not defined here, but please correct me if I am wrong.
There was a problem hiding this comment.
yep, it looks like I erased the definition when commiting
|
@ivanlmh Huge thanks to you contribution! Encodec is a very useful one. |
|
I think that's fixed... idk why the definitions disappeared, but anyway :) |
|
@ivanlmh Merged, huge thanks for your contribution once again! |
Hi Hao Hao, I've added Meta's EnCodec model (the 2 main versions they offer pre-trained) for FAD calculation
Let me know if there is something you think should be fixed or whatever :)