hop_length=hop_length,
onesided=True,
center=True,
window=window)
if total_audio is None:
total_audio = audio_out_denoised
else:
total_audio = torch.cat((total_audio, audio_out_denoised),
0)
else:
total_audio = torch.istft(audio_spec,
n_fft,
hop_length=hop_length,
onesided=True,
center=True,
window=window)
scipy.io.wavfile.write(
ARGS.output_dir + "/denoised_" + str(i) + "_" +
filename[0].split("/")[-1].split(".")[0] + ".wav", sr,
total_audio.numpy().T)
#before or after writing intensity scaling to chose dB value
scipy.io.wavfile.write(
ARGS.output_dir + "/denoised_" + str(i) + "_" +
filename[0].split("/")[-1].split(".")[0] + ".wav", sr,
total_audio.numpy().T)
log.debug("Finished proccessing")
log.close()
print(vscore)
print("Viterbi paths:")
print(vpaths)
logger.log(step, tr_nll.data[0], val_nll.data[0],
tr_nll.data[0] / batch_size,
val_nll.data[0] / batch_size)
# serialize model occasionally:
if step % save_every == 0: logger.save(step, crf)
step += 1
if step > max_iters: raise StopIteration
del dataset
#--- handle keyboard interrupts:
except KeyboardInterrupt:
del dataset
logger.close()
print("-" * 80)
print("Halted training; reached {} training iterations.".format(step))
except StopIteration:
del dataset
logger.close()
print("-" * 80)
print("Finished training; reached max iterations of {}.".format(max_iters))
except Exception as e:
print("Something went wrong:")
print(e)
del dataset
logger.close()
