def evaluate_model(model, data_loader, checkpoint_dir, limit_eval_to=5):
"""evaluate model and save generated wav and plot
"""
test_path = data_loader.dataset.test_path
test_mel_path = data_loader.dataset.test_mel_path
test_mel_files = sorted(os.listdir(test_mel_path))
test_wav_path = data_loader.dataset.test_mel_path
test_wav_files = sorted(os.listdir(test_wav_path))
counter = 0
output_dir = os.path.join(checkpoint_dir, 'eval')
for f in test_mel_files:
mel = np.load(os.path.join(test_mel_path, f))
wav = model.generate(mel)
# save wav
wav_path = os.path.join(
output_dir,
"checkpoint_step{:09d}_wav_{}.wav".format(global_step, counter))
wavfile.write(wav_path, hp.sample_rate, soundsc(wav))
#librosa.output.write_wav(wav_path, wav, sr=hp.sample_rate)
# save wav plot
fig_path = os.path.join(
output_dir,
"checkpoint_step{:09d}_wav_{}.png".format(global_step, counter))
fig = plt.plot(wav.reshape(-1))
plt.savefig(fig_path)
# clear fig to drawing to the same plot
plt.clf()
counter += 1
# stop evaluation early via limit_eval_to
if counter >= limit_eval_to:
break
def evaluate_model(model, data_loader, checkpoint_dir, limit_eval_to=5):
"""evaluate model and save generated wav and plot
"""
test_path = data_loader.dataset.test_path
test_mel_path = data_loader.dataset.test_mel_path
test_mel_files = sorted(os.listdir(test_mel_path))
test_wav_path = data_loader.dataset.test_mel_path
test_wav_files = sorted(os.listdir(test_wav_path))
counter = 0
output_dir = os.path.join(checkpoint_dir,'eval')
for f in test_mel_files:
mel = np.load(os.path.join(test_mel_path,f))
mel_r, _ = data_loader.dataset.split(mel)
# manually do the interpolation
m_in = mel_r
rs = hp.time_resample_factor
if rs != 1:
# interpolate back to same timescale
t = [np.interp(np.arange(rs * m_in.shape[1]) / rs, np.arange(m_in.shape[1]), m_in[i, :]) for i in range(m_in.shape[0])]
m_in = np.vstack(t)
wav = model.generate(m_in)
# save wav
wav_path = os.path.join(output_dir,"checkpoint_step{:09d}_wav_{}.wav".format(global_step,counter))
wavfile.write(wav_path, hp.sample_rate, soundsc(wav))
#librosa.output.write_wav(wav_path, wav, sr=hp.sample_rate)
# save wav plot
fig_path = os.path.join(output_dir,"checkpoint_step{:09d}_wav_{}.png".format(global_step,counter))
fig = plt.plot(wav.reshape(-1))
plt.savefig(fig_path)
# clear fig to drawing to the same plot
plt.clf()
counter += 1
# stop evaluation early via limit_eval_to
if counter >= limit_eval_to:
break
post_time = 0
joint_time = 0
for jj in range(n_plot):
# use extra steps from earlier
pjj = preds[:(finished_step[jj] + extra_steps), jj]
spectrogram = pjj * saved_std + saved_mean
mel_dump = "sample_{}_mels.npz".format(jj)
np.savez(mel_dump, mels=spectrogram)
prename = "sample_{}_pre.wav".format(jj)
this_time = time.time()
reconstructed_waveform = sonify(spectrogram,
len(spectrogram) * step, logmel)
end_this_time = time.time()
wavfile.write(prename, sample_rate, soundsc(reconstructed_waveform))
elapsed = end_this_time - this_time
print("Elapsed pre sampling time {} s".format(elapsed))
pre_time += elapsed
joint_time += elapsed
fftsize = 512
substep = 32
postname = "sample_{}_post.wav".format(jj)
this_time = time.time()
rw_s = np.abs(
stft(reconstructed_waveform,
fftsize=fftsize,
step=substep,