t = x[ni, 0]
t = t[:, 0]
orig_buf[ni * step:(ni * step) + cut] += t
x_o = orig_buf
x_r = rec_buf
f, axarr = plt.subplots(2, 1)
axarr[0].plot(x_r)
axarr[0].set_title("Reconstruction")
axarr[1].plot(x_o)
axarr[1].set_title("Original")
plt.savefig("vq_vae_generation_results")
plt.close()
f, axarr = plt.subplots(2, 1)
specplot(specgram(x_r), axarr[0])
axarr[0].set_title("Reconstruction")
specplot(specgram(x_o), axarr[1])
axarr[1].set_title("Original")
plt.savefig("vq_vae_generation_results_spec")
plt.close()
wavfile.write("original_wav.wav", 8000, soundsc(x_o))
wavfile.write("reconstructed_wav.wav", 8000, soundsc(x_r))
from IPython import embed
embed()
raise ValueError()
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 * itr._std + itr._mean
prename = "{}_{}_sample_{}_pre.wav".format(file_names[jj], args.input_type, jj)
this_time = time.time()
if os.path.exists(prename):
print("{} file already found, not redoing calculations".format(prename))
fs, reconstructed_waveform = wavfile.read(prename)
else:
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(file_names[jj], args.input_type, jj)
this_time = time.time()
if not os.path.exists(postname):
rw_s = np.abs(stft(reconstructed_waveform, fftsize=fftsize, step=substep, real=False,
compute_onesided=False))
rw = iterate_invert_spectrogram(rw_s, fftsize, substep, n_iter=gl_steps, verbose=True)
end_this_time = time.time()
wavfile.write(postname, sample_rate, soundsc(rw))
# put all the shorter traces into one long sequence, no overlap add or anything
i_rec = final_hidden_indices
rec_buf = np.zeros((len(x_rec) * step + 2 * cut))
i_buf = np.zeros((len(x_rec) * step + 2 * cut))
for ni in range(len(x_rec)):
t = x_rec[ni, 0]
t = t[:, 0]
rec_buf[ni * step:(ni * step) + cut] += t
i_buf[ni * step:(ni * step) + cut] = i_rec[ni]
x_r = rec_buf
i_r = i_buf
f, axarr = plt.subplots(2, 1)
axarr[0].plot(x_r)
axarr[0].set_title("Generation")
axarr[1].plot(i_r, color="r")
plt.savefig("vq_vae_generation_results")
plt.close()
f, axarr = plt.subplots(1, 1)
specplot(specgram(x_r), axarr)
axarr.set_title("Generation")
#axarr[1].plot(i_r, color="r")
plt.savefig("vq_vae_generation_results_spec")
plt.close()
wavfile.write("generated_wav.wav", 8000, soundsc(x_r))
with tf.Session():
spectrogram = logmel(waveform).eval()
spectrogram2 = logmel2(waveform)
spectrogram = (spectrogram - spectrogram.min()) / float(spectrogram.max() -
spectrogram.min())
spectrogram2 = (spectrogram2 - spectrogram2.min()) / float(spectrogram2.max() -
spectrogram2.min())
f, axarr = plt.subplots(1, 2)
axarr[0].imshow(spectrogram)
axarr[1].imshow(spectrogram2)
plt.savefig("tmpspec")
reconstructed_waveform = sonify(spectrogram, len(waveform), logmel)
wavfile.write("tmp.wav", sample_rate, soundsc(reconstructed_waveform))
reconstructed_waveform2 = sonify(spectrogram2, len(waveform), logmel)
wavfile.write("tmp2.wav", sample_rate, soundsc(reconstructed_waveform2))
fftsize = 512
substep = 32
rw_s = np.abs(
stft(reconstructed_waveform,
fftsize=fftsize,
step=substep,
real=False,
compute_onesided=False))
rw = iterate_invert_spectrogram(rw_s,
fftsize,
substep,
n_iter=100,