def invert_to_sound (specname):
specinfo = np.load(specname)
original = utils.ispecgram(specinfo,
n_fft=512,
hop_length=None,
mask=True,
log_mag=True,
re_im=False,
dphase=True,
mag_only=False,
num_iters=1000)
plt.plot(original)
plt.show()
# scale the original array to audible sound
write(specname +'Inverse.wav', 16000, original)
# samples = sess.run(G_sample, feed_dict={Z: sample_Z(16, Z_dim)})
# # fig = plot(samples)
# # plt.savefig('out/{}.png'.format(str(i).zfill(3)), bbox_inches='tight')
# i += 1
# # plt.close(fig)
# every 1000 iter test
if it % 1000 == 0:
samples = sess.run(G_sample, feed_dict={Z: sample_Z('3230')})
samples = np.transpose(samples)
XXX = concat_to_spec(samples)
ori = utils.ispecgram(XXX,
n_fft=512,
hop_length=None,
mask=True,
log_mag=True,
re_im=False,
dphase=True,
mag_only=False,
num_iters=1000)
write('out/{}.wav'.format(str(i).zfill(3)), 16000, ori)
i += 1
# print samples.shape
# fig = plot(samples)
# plt.savefig('out/{}.png'.format(str(i).zfill(3)), bbox_inches='tight')
# i += 1
# plt.close(fig)
V_mb, A_mb = data_gen.next()
# print 'V_mb', V_mb
def main():
video_all = [
'3229', '3230', '3231', '3232', '3233', '3234', '3235', '3236', '3237'
]
save_folder = '/home/yuanxin/code/bitplanes-tracking/'
for video_index in video_all:
data_folder = '/home/yuanxin/Downloads/output/test_' + video_index
specname = './spec.npy'
centername = './center.npy'
os.chdir(data_folder)
feat_map = []
for file in sorted(glob.glob("*.prob")):
print file
f = open(file, "rb")
size = array.array("i") # L is the typecode for uint32
size.fromfile(f, 5)
f.seek(20, os.SEEK_SET)
total = size[0] * size[1] * size[2] * size[3] * size[4]
data = array.array("f")
data.fromfile(f, total)
data = np.array(data)
data = data.reshape(size[0], size[1], size[2], size[3], size[4])
# print data.shape
feat_map.append(data)
f.close()
feat_map = np.array(feat_map)
max_index = []
for i in range(feat_map.shape[0]):
max_index.append(np.argmax(feat_map[i][0].reshape(16)))
os.chdir(save_folder)
audioname = './VGG_out/inverse_' + video_index + '.wav'
X = np.load(specname)
center = np.load(centername)
label = max_index
label = np.lib.pad(label, (7, 7),
'constant',
constant_values=(label[0], label[-1]))
# interp label
# print X.shape
# print center.shape
# print label.shape
pdb.set_trace()
Y = center_to_spec(X[0:len(label)], center, label)
LabelT = np.transpose(Y)
print 'LabelT', LabelT.shape
# reshape to the correct rainbowgram format
XXX = concat_to_spec(LabelT)
print 'XXX', XXX.shape
ori = utils.ispecgram(XXX,
n_fft=512,
hop_length=None,
mask=True,
log_mag=True,
re_im=False,
dphase=True,
mag_only=False,
num_iters=1000)
# plt.plot(ori)
# plt.show()
# scale the original array to audible sound
write(audioname, 15360, ori)
def time_windowed_stft_transfer(content_file,
style_file,
window_seconds=5,
hop_seconds=2.5):
rate, content = wavfile.read(content_file)
if total_seconds:
content = content[:rate * total_seconds]
total_samples = int(1.0 * params.sample_rate / rate * len(content))
print('Resampling content')
content = resample(content, total_samples)
content = content.astype(np.float32)
librosa.output.write_wav('output/content.wav',
sr=params.sample_rate,
y=content.astype(np.int16),
norm=False)
x = np.random.normal(0, 1, total_samples)
x = x.astype(np.float32)
x /= np.max(np.abs(x))
x *= (amplitude - 1)
# x *= 0.01
x = content
unnormal_noise = x.flatten()
librosa.output.write_wav('output/x.wav',
sr=params.sample_rate,
y=unnormal_noise.astype(np.int16),
norm=False)
rate, style = wavfile.read(style_file)
if total_seconds:
style = style[:rate * total_seconds]
total_samples2 = int(len(style) / rate) * params.sample_rate
print('Resampling style')
style = resample(style, total_samples2)
style = style.astype(np.float32)
librosa.output.write_wav('output/style.wav',
sr=params.sample_rate,
y=style.astype(np.int16),
norm=False)
session = tf.Session()
style_stft = specgram(style,
n_fft=params.fft_size,
log_mag=audio_utils.log,
hop_length=256,
mag_only=True)
style_stft = style_stft.reshape([style_stft.shape[0], style_stft.shape[1]])
style_stft = style_stft.T
style_stft = style_stft[np.newaxis, :, :]
print('Style: ')
print(
f'Min: {np.min(style_stft):.4f}, Max: {np.max(style_stft):.4f}, Mean: {np.mean(style_stft):.4f}'
)
content_stft = specgram(content,
n_fft=params.fft_size,
log_mag=audio_utils.log,
hop_length=256,
mag_only=True)
content_stft = content_stft.reshape(
[content_stft.shape[0], content_stft.shape[1]])
content_stft = content_stft.T
content_stft = content_stft[np.newaxis, :, :]
print('Content: ')
print(
f'Min: {np.min(content_stft):.4f}, Max: {np.max(content_stft):.4f}, Mean: {np.mean(content_stft):.4f}'
)
result_stft = content_stft[:, :, :]
hop_samples = int(content_stft.shape[1] /
(len(content) / params.sample_rate) * hop_seconds)
window_samples = int(content_stft.shape[1] /
(len(content) / params.sample_rate) * window_seconds)
transferrer = StftTransfer(
session,
(content_stft.shape[0], window_samples, content_stft.shape[2]))
print('Starting window shifting')
first = True
for j in range(3):
content_stft = result_stft[:]
for i in range(0 if first else random.randint(0, hop_samples),
result_stft.shape[1], hop_samples):
start_time = time.time()
# next_hop = int(hop_samples[i + 1]+1) if len(hop_samples) > i + 1 else content_stft.shape[1]
current_content = content_stft[:, i:i + window_samples, :]
current_initial = result_stft[:, i:i + window_samples, :]
if current_content.shape[1] != window_samples:
continue
current_style = get_best_style(transferrer, current_initial[:],
style_stft, hop_samples // 4)
current_result_stft = transferrer.stft_transfer(
content_stft=current_content,
style_stft=current_style,
initial_stft=current_initial,
maxiter=20)
result_stft[:, i:i + window_samples, :] = current_result_stft
print(
f'Epoch {j} last sample: {(i/result_stft.shape[1]):.4f} elapsed time: {(time.time() - start_time):.4f}'
)
result = result_stft[0].T[:, :, np.newaxis]
result = ispecgram(result,
params.fft_size,
hop_length=256,
log_mag=audio_utils.log,
mag_only=True,
num_iters=20)
result = np.clip(result, -1, 1)
result *= amplitude
# result = stft_transfer(content=content, style=style, initial=x)
librosa.output.write_wav(f'output/result{j}.wav',
sr=params.sample_rate,
y=result.astype(np.int16),
norm=False)
print('Written result.wav')
# librosa.output.write_wav('output/result.wav', sr=sample_rate, y=result.astype(np.int16), norm=False)
# print('Written result.wav')
first = False
print('Inverting specgram')
result = result_stft[0].T[:, :, np.newaxis]
result = ispecgram(result,
params.fft_size,
hop_length=256,
log_mag=audio_utils.log,
mag_only=True,
num_iters=100)
result = np.clip(result, -1, 1)
result *= amplitude
# result = stft_transfer(content=content, style=style, initial=x)
librosa.output.write_wav('output/result.wav',
sr=params.sample_rate,
y=result.astype(np.int16),
norm=False)
print('Written result.wav')