assertion error: padded_mix.dtype == th.float64 in Separator.separate_tensor method

[charliemarshall1996]

🐛 Bug Report

When calling the separate_tensor method on an instance of the Separator class of the api.py module, with a wav value of a Tensor with a dtype of float32 (as per the method docstring), apply.py raises an assertion error for padded_mix.dtype == th.float64, within the apply_model function. I assumed the padded_mix.dtype should have been the same as the dtype of the input audio, so altered the assertion to be padded_mix.dtype == th.float32. This seemed to then work without issue.

To Reproduce

1. Synthesize or import audio as Tensor with dtype=torch.float32
2. Initialize Separator class
3. Call separate_tensor method with created Tensor object

Expected behavior

separate_tensor would return expected vals.

Actual Behavior

AssertionError raised:

Traceback (most recent call last):
File "path", line 19, in
separator.separate_tensor(reshaped_tensor)
File "path", line 271, in separate_tensor
out = apply_model(
^^^^^^^^^^^^
File "path", line 218, in apply_model
res = apply_model(sub_model, mix, **kwargs, callback_arg=callback_arg)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "path", line 273, in apply_model
res = apply_model(model, shifted, **kwargs, callback_arg=callback_arg)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "path", line 323, in apply_model
chunk_out = future.result() # Get the result from the future.
^^^^^^^^^^^^^^^
File "path", line 131, in result
return self.func(*self.args, **self.kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "path", line 365, in apply_model
assert padded_mix.dtype == th.float64
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Your Environment

• Python and PyTorch version:
  • py 3.11.4 64-bit
  • torch 2.0.1
• Operating system and version (desktop or mobile):
  • Windows 11
• Hardware (gpu or cpu, amount of RAM etc.):
  • cpu: AMD Ryzen 5
  • gpu: NVIDIA GeForce GTX 1650 SUPER
  • RAM: 32 GB

[CarlGao4]

Please try version 4.1.0a2. 4.1.0a1 has a lot of bugs with API, and I've fixed them in 4.1.0a2, which hasn't been merged. See #552

[charliemarshall1996]

Perfect. I'll give this a go. Thanks!