refactor!: change model to expect mel-band oriented tensors instead of time-oriented ones

[roedoejet]

See EveryVoiceTTS/EveryVoice#572

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
hfgl/utils.py	53% smaller
hfgl/cli.py	12% smaller

[SamuelLarkin]

make the code a little bit more readable

[SamuelLarkin]

Suggested change

	if time_oriented:
	data = data.transpose(0, 1)
	data_size = data.size()
	if (
	checkpoint["hyper_parameters"]["config"]["preprocessing"]["audio"]["n_mels"]
	not in data_size
	):
	raise ValueError(
	f"Your model expects a spectrogram of dimensions [K (Mel bands), T (frames)] where K == {checkpoint['hyper_parameters']['config']['preprocessing']['audio']['n_mels']} but you provided a tensor of size {data_size}"
	)
	if (
	data_size[0]
	!= checkpoint["hyper_parameters"]["config"]["preprocessing"]["audio"]["n_mels"]
	):
	raise ValueError(
	f"We expected the first dimension of your Mel spectrogram to correspond with the number of Mel bands declared by your model ({checkpoint['hyper_parameters']['config']['preprocessing']['audio']['n_mels']}). Instead, we found you model has the dimensions {data_size}. If your spectrogram is time-oriented, please re-run this command with the '--time-oriented' flag."
	)
	if time_oriented:
	data = data.transpose(0, 1)
	data_size = data.size()
	config_n_mels = checkpoint["hyper_parameters"]["config"]["preprocessing"]["audio"]["n_mels"]
	if (config_n_mels not in data_size):
	raise ValueError(
	f"Your model expects a spectrogram of dimensions [K (Mel bands), T (frames)] where K == {config_n_mels} but you provided a tensor of size {data_size}"
	)
	if (data_size[0] != config_n_mels):
	raise ValueError(
	f"We expected the first dimension of your Mel spectrogram to correspond with the number of Mel bands declared by your model ({config_n_mels}). Instead, we found you model has the dimensions {data_size}. If your spectrogram is time-oriented, please re-run this command with the '--time-oriented' flag."
	)

[joanise]

Yes, please accept this suggestion, the code is quite hard to read.

[joanise]

Please accept Sam's change request, otherwise this looks fine.

[joanise]

Yes, please accept this suggestion, the code is quite hard to read.

[joanise]

This warning message is good, probably giving the user the feedback I was asking for in the top-level PR.

[joanise]

Note: you'll need to rebase, and solve the merge conflict: we now have to use shell_complete, not autocompletion.