diff options
author | Jan Buethe <jbuethe@amazon.de> | 2023-09-13 17:57:28 +0300 |
---|---|---|
committer | Jan Buethe <jbuethe@amazon.de> | 2023-09-13 17:57:28 +0300 |
commit | 82f48d368b41d8bc4286e1375419daacbd10dbca (patch) | |
tree | 03c470187bf3b85779c345c5c939ed78fdbfb6e0 | |
parent | e7beaec3fb49df389b077799c5d1778ccb68610e (diff) |
removed trailing whitespace in fargan
Signed-off-by: Jan Buethe <jbuethe@amazon.de>
-rw-r--r-- | dnn/torch/fargan/fargan.py | 19 | ||||
-rw-r--r-- | dnn/torch/fargan/filters.py | 2 | ||||
-rw-r--r-- | dnn/torch/fargan/stft_loss.py | 16 | ||||
-rw-r--r-- | dnn/torch/fargan/test_fargan.py | 28 | ||||
-rw-r--r-- | dnn/torch/fargan/train_fargan.py | 4 |
5 files changed, 34 insertions, 35 deletions
diff --git a/dnn/torch/fargan/fargan.py b/dnn/torch/fargan/fargan.py index 91497ccd..e9cc687a 100644 --- a/dnn/torch/fargan/fargan.py +++ b/dnn/torch/fargan/fargan.py @@ -81,7 +81,7 @@ def gen_phase_embedding(periods, frame_size): class GLU(nn.Module): def __init__(self, feat_size): super(GLU, self).__init__() - + torch.manual_seed(5) self.gate = weight_norm(nn.Linear(feat_size, feat_size, bias=False)) @@ -89,16 +89,16 @@ class GLU(nn.Module): self.init_weights() def init_weights(self): - + for m in self.modules(): if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d)\ or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding): nn.init.orthogonal_(m.weight.data) def forward(self, x): - - out = x * torch.sigmoid(self.gate(x)) - + + out = x * torch.sigmoid(self.gate(x)) + return out class FWConv(nn.Module): @@ -160,21 +160,21 @@ class FARGANSub(nn.Module): self.subframe_size = subframe_size self.nb_subframes = nb_subframes self.cond_size = cond_size - + #self.sig_dense1 = nn.Linear(4*self.subframe_size+self.passthrough_size+self.cond_size, self.cond_size, bias=False) self.fwc0 = FWConv(4*self.subframe_size+80, self.cond_size) self.sig_dense2 = nn.Linear(self.cond_size, self.cond_size, bias=False) self.gru1 = nn.GRUCell(self.cond_size, self.cond_size, bias=False) self.gru2 = nn.GRUCell(self.cond_size, self.cond_size, bias=False) self.gru3 = nn.GRUCell(self.cond_size, self.cond_size, bias=False) - + self.dense1_glu = GLU(self.cond_size) self.dense2_glu = GLU(self.cond_size) self.gru1_glu = GLU(self.cond_size) self.gru2_glu = GLU(self.cond_size) self.gru3_glu = GLU(self.cond_size) self.ptaps_dense = nn.Linear(4*self.cond_size, 5) - + self.sig_dense_out = nn.Linear(4*self.cond_size, self.subframe_size, bias=False) self.gain_dense_out = nn.Linear(4*self.cond_size, 1) @@ -184,7 +184,7 @@ class FARGANSub(nn.Module): def forward(self, cond, prev, exc_mem, phase, period, states, gain=None): device = exc_mem.device #print(cond.shape, prev.shape) - + dump_signal(prev, 'prev_in.f32') idx = 256-torch.clamp(period[:,None], min=self.subframe_size+2, max=254) @@ -283,4 +283,3 @@ class FARGAN(nn.Module): prev = out states = [s.detach() for s in states] return sig, states - diff --git a/dnn/torch/fargan/filters.py b/dnn/torch/fargan/filters.py index 4a4a86f8..8ec97ea6 100644 --- a/dnn/torch/fargan/filters.py +++ b/dnn/torch/fargan/filters.py @@ -41,6 +41,6 @@ if __name__ == '__main__': A = toeplitz_from_filter(a) #print(A) R = filter_iir_response(a, 5) - + RA = toeplitz_from_filter(R) print(RA) diff --git a/dnn/torch/fargan/stft_loss.py b/dnn/torch/fargan/stft_loss.py index 98a60ec6..accf2f4a 100644 --- a/dnn/torch/fargan/stft_loss.py +++ b/dnn/torch/fargan/stft_loss.py @@ -17,7 +17,7 @@ def stft(x, fft_size, hop_size, win_length, window): Returns: Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1). """ - + #x_stft = torch.stft(x, fft_size, hop_size, win_length, window, return_complex=False) #real = x_stft[..., 0] #imag = x_stft[..., 1] @@ -83,26 +83,26 @@ class LogSTFTMagnitudeLoss(torch.nn.Module): var_x = torch.var(x, dim=1, keepdim=True) var_y = torch.var(y, dim=1, keepdim=True) - + std_x = torch.std(x, dim=1, keepdim=True) std_y = torch.std(y, dim=1, keepdim=True) - + x_minus_mean = x - mean_x y_minus_mean = y - mean_y - + pearson_corr = torch.sum(x_minus_mean * y_minus_mean, dim=1, keepdim=True) / \ (torch.sqrt(torch.sum(x_minus_mean ** 2, dim=1, keepdim=True) + 1e-7) * \ torch.sqrt(torch.sum(y_minus_mean ** 2, dim=1, keepdim=True) + 1e-7)) - + numerator = 2.0 * pearson_corr * std_x * std_y denominator = var_x + var_y + (mean_y - mean_x)**2 - + ccc = numerator/denominator - + ccc_loss = F.l1_loss(1.0 - ccc, torch.zeros_like(ccc))''' return error_loss #+ ccc_loss#+ ccc_loss - + class STFTLoss(torch.nn.Module): """STFT loss module.""" diff --git a/dnn/torch/fargan/test_fargan.py b/dnn/torch/fargan/test_fargan.py index 8a6d2c25..76e1f854 100644 --- a/dnn/torch/fargan/test_fargan.py +++ b/dnn/torch/fargan/test_fargan.py @@ -55,35 +55,35 @@ nb_frames = features.shape[1] gamma = checkpoint['model_kwargs']['gamma'] def lpc_synthesis_one_frame(frame, filt, buffer, weighting_vector=np.ones(16)): - + out = np.zeros_like(frame) filt = np.flip(filt) - + inp = frame[:] - - + + for i in range(0, inp.shape[0]): - + s = inp[i] - np.dot(buffer*weighting_vector, filt) - + buffer[0] = s - + buffer = np.roll(buffer, -1) - + out[i] = s - + return out def inverse_perceptual_weighting (pw_signal, filters, weighting_vector): - + #inverse perceptual weighting= H_preemph / W(z/gamma) - + signal = np.zeros_like(pw_signal) buffer = np.zeros(16) num_frames = pw_signal.shape[0] //160 assert num_frames == filters.shape[0] for frame_idx in range(0, num_frames): - + in_frame = pw_signal[frame_idx*160: (frame_idx+1)*160][:] out_sig_frame = lpc_synthesis_one_frame(in_frame, filters[frame_idx, :], buffer, weighting_vector) signal[frame_idx*160: (frame_idx+1)*160] = out_sig_frame[:] @@ -97,11 +97,11 @@ if __name__ == '__main__': features = torch.tensor(features).to(device) #lpc = torch.tensor(lpc).to(device) periods = torch.tensor(periods).to(device) - + sig, _ = model(features, periods, nb_frames - 4) weighting_vector = np.array([gamma**i for i in range(16,0,-1)]) sig = sig.detach().numpy().flatten() sig = inverse_perceptual_weighting(sig, lpc[0,:,:], weighting_vector) - + pcm = np.round(32768*np.clip(sig, a_max=.99, a_min=-.99)).astype('int16') pcm.tofile(signal_file) diff --git a/dnn/torch/fargan/train_fargan.py b/dnn/torch/fargan/train_fargan.py index 20cf2d2b..4ab20045 100644 --- a/dnn/torch/fargan/train_fargan.py +++ b/dnn/torch/fargan/train_fargan.py @@ -141,9 +141,9 @@ if __name__ == '__main__': loss.backward() optimizer.step() - + #model.clip_weights() - + scheduler.step() running_specc += specc_loss.detach().cpu().item() |