x = np.mean(x,axis=-1)
#perform stft
S = util.stft_real(x, blockSize= blockSize,hopSize=hopSize)
magnitude = np.abs(S).astype(np.float32)
angle = np.angle(S).astype(np.float32)
#initialize the model
model = Model.ModelSingleStep(blockSize)
#load the pretrained model
checkpoint = torch.load("savedModel_fcnn_best.pt", map_location=lambda storage, loc:storage)
model.load_state_dict(checkpoint['state_dict'])
#switch to eval mode
model.eval()
###################################
#Run your Model here to obtain a mask
###################################
magnitude_masked = model.process(magnitude)
###################################
#perform reconstruction
y = util.istft_real(magnitude_masked * np.exp(1j * angle), blockSize=blockSize, hopSize=hopSize)
#save the result
util.wavwrite(sys.argv[2], y,fs)
x, fs = util.wavread(sys.argv[1])
#downmix to single channel
x = np.mean(x, axis=-1)
#perform stft
S = util.stft_real(x, blockSize=blockSize, hopSize=hopSize)
magnitude = np.abs(S).astype(np.float32)
angle = np.angle(S).astype(np.float32)
#initialize the model
model = Model_fcn.ModelSingleStep(blockSize)
#load the pretrained model
model.load_state_dict(
torch.load("Modelfcn.pt", map_location=lambda storage, loc: storage))
#switch to eval mode
model.eval()
###################################
#Run your Model here to obtain a mask
###################################
spectro_pred = model.process(magnitude)
###################################
#perform reconstruction
y = util.istft_real(spectro_pred * np.exp(1j * angle),
blockSize=blockSize,
hopSize=hopSize)
#save the result
util.wavwrite(sys.argv[2], y, fs)
with torch.no_grad():
nFrame = magnitude.shape[1]
GRUout = []
for i in range(nFrame):
h = model.encode(mixture[i, :, :])
h_d.append(h)
if i < 5: #delay is 5
if i == 0:
outp, GRUo = model.forward(mixture[i, :, :], None,
torch.zeros_like(h))
else:
outp, GRUo = model.forward(mixture[i, :, :], GRUout[i - 1],
torch.zeros_like(h))
else:
outp, GRUo = model.forward(mixture[i, :, :], GRUout[i - 1],
h_d[i - 5])
GRUout.append(GRUo)
output.append(outp)
mag_voice[:, i] = magnitude[:, i] * output[i].numpy()
###################################
#perform reconstruction
y = util.istft_real(mag_voice * np.exp(1j * angle),
blockSize=blockSize,
hopSize=hopSize)
#save the result
util.wavwrite(sys.argv[2], y, fs)