add causal option for HiFiGAN

[chomeyama]

Implemented causal option for HiFiGAN

[kan-bayashi]

Thank you for your contribution!
I have several requests:

1. Could you fix the CI errors?
2. Could you add the test code for causal calculation?

Example test code:

ParallelWaveGAN/test/test_parallel_wavegan.py

Lines 310 to 354 in edf35dd

	def test_causal_parallel_wavegan(upsample_net, aux_context_window):
	batch_size = 1
	batch_length = 4096
	args_g = make_generator_args(
	use_causal_conv=True,
	upsample_net=upsample_net,
	aux_context_window=aux_context_window,
	dropout=0.0,
	)
	model_g = ParallelWaveGANGenerator(**args_g)
	z = torch.randn(batch_size, 1, batch_length)
	c = torch.randn(
	batch_size,
	args_g["aux_channels"],
	batch_length // np.prod(args_g["upsample_params"]["upsample_scales"]),
	)
	z_ = z.clone()
	c_ = c.clone()
	z_[..., z.size(-1) // 2 :] = torch.randn(z[..., z.size(-1) // 2 :].shape)
	c_[..., c.size(-1) // 2 :] = torch.randn(c[..., c.size(-1) // 2 :].shape)
	c = torch.nn.ConstantPad1d(args_g["aux_context_window"], 0.0)(c)
	c_ = torch.nn.ConstantPad1d(args_g["aux_context_window"], 0.0)(c_)
	try:
	# check not equal
	np.testing.assert_array_equal(c.numpy(), c_.numpy())
	except AssertionError:
	pass
	else:
	raise AssertionError("Must be different.")
	try:
	# check not equal
	np.testing.assert_array_equal(z.numpy(), z_.numpy())
	except AssertionError:
	pass
	else:
	raise AssertionError("Must be different.")
	# check causality
	y = model_g(z, c)
	y_ = model_g(z_, c_)
	np.testing.assert_array_equal(
	y[..., : y.size(-1) // 2].detach().cpu().numpy(),
	y_[..., : y_.size(-1) // 2].detach().cpu().numpy(),
	)

ParallelWaveGAN/test/test_melgan.py

Lines 270 to 296 in edf35dd

	def test_causal_melgan(dict_g):
	batch_size = 4
	batch_length = 4096
	args_g = make_melgan_generator_args(**dict_g)
	upsampling_factor = np.prod(args_g["upsample_scales"])
	c = torch.randn(
	batch_size, args_g["in_channels"], batch_length // upsampling_factor
	)
	model_g = MelGANGenerator(**args_g)
	c_ = c.clone()
	c_[..., c.size(-1) // 2 :] = torch.randn(c[..., c.size(-1) // 2 :].shape)
	try:
	# check not equal
	np.testing.assert_array_equal(c.numpy(), c_.numpy())
	except AssertionError:
	pass
	else:
	raise AssertionError("Must be different.")
	# check causality
	y = model_g(c)
	y_ = model_g(c_)
	assert y.size(2) == c.size(2) * upsampling_factor
	np.testing.assert_array_equal(
	y[..., : c.size(-1) // 2 * upsampling_factor].detach().cpu().numpy(),
	y_[..., : c_.size(-1) // 2 * upsampling_factor].detach().cpu().numpy(),
	)

[kan-bayashi]

Oh sorry, the CI errors are not related to this PR.
I will fix it. Please ignore about it.

[kan-bayashi]

@chomeyama I fixed CI errors #327. Could you merge the current master?

[chomeyama]

I have implemented a test code for HiFiGAN and added some modifications to HiFiGAN Generator.

[kan-bayashi]

Cool, could you fix ci errors about black?

[kan-bayashi]

Thank you for your contribution!

[gafsd]

How do you find this change to hifigan output? Much different?