Add finetuning support for Speaker class

[ain-soph]

To allow tuning for speaker_embeds tensor. The PR includes:

1. Rename: _sample_random -> sample_random_tensor to allow usage out of class.
2. The spk_emb argument in Speaker.apply() changes str -> Union[str, torch.Tensor]
3. Add inplace argument to support gradient backward when setting to False
4. Update return type to return torch.Tensor

Besides, I don't quite see the advantage and motivation to encode the speaker_embeds tensor to string. To me, it only makes things more complex. I would personally prefer cpu torch.Tensor or np.ndarray

cc @fumiama

[fumiama]

Besides, I don't quite see the advantage and motivation to encode the speaker_embeds tensor to string.

People can still use the tensor but the .pt file can easily contains virus. The str encoding is equal to the numpy array dump but is easy to copy&paste.

[fumiama]

Also, you can save the str to a txt file (with the ext .pt if you like) and use it just like a .pt file if you like.

[ain-soph]

Also, you can save the str to a txt file (with the ext .pt if you like) and use it just like a .pt file if you like.

"use it just like a .pt file if you like" -> You mean we could directly read the tensor via torch.load() instead of using pybase16384 to encode/decode? I didn't know that.

[fumiama]

Also, you can save the str to a txt file (with the ext .pt if you like) and use it just like a .pt file if you like.

"use it just like a .pt file if you like" -> You mean we could directly read the tensor via torch.load() instead of using pybase16384 to encode/decode? I didn't know that.

Just treat the str as a kind of binary.

[fumiama]

Thanks!