-
Notifications
You must be signed in to change notification settings - Fork 0
/
style_transfer.py
92 lines (79 loc) · 3.13 KB
/
style_transfer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
import time
import tensorflow as tf
import os
from data_manager import DataManager
from data_manager import get_data
import argparse
import librosa
import numpy as np
import train
def print_tags(header, result):
tag_prob_list = []
for i in range(train.N_CLASSES):
tag_prob_list.append((header[i], result[i]))
top_tags = sorted(tag_prob_list, key=lambda x: x[1])[::-1]
for a, b in top_tags[:10]:
print a, b
print '...'
for a, b in top_tags[-11:-1]:
print a, b
def style_transfer(model, song_fname, start_in_seconds):
# First we restore the weights from the checkpoint
header, _, _, _, _ = get_data(train.N_CLASSES, train.MERGE_TAGS, train.SPLIT_RANDOMLY)
weights, biases = train.get_vars()
saver = tf.train.Saver()
sess = tf.Session()
saver.restore(sess, model)
print 'loading song'
audio = librosa.load(song_fname, sr=train.SAMPLE_RATE)[0]
start_idx = start_in_seconds * train.SAMPLE_RATE
# Clip to 3 seconds
clip = audio[start_idx : start_idx + train.SECONDS_OF_AUDIO * train.SAMPLE_RATE]
clip = np.expand_dims(np.array(clip), 0)
print clip.shape
x_var = tf.Variable(clip, name='x')
pred = train.net(x_var, weights, biases, train.keep_prob) # no sigmoid applied
# Print the predictions for the clip
sess.run(x_var.initializer)
librosa.output.write_wav('./cropped.wav', x_var.eval(session=sess).flatten(), train.SAMPLE_RATE)
result = sess.run(pred, feed_dict={train.keep_prob: 1}).flatten()
print_tags(header, result)
print ''
target = raw_input('Which tag to boost? ')
assert target in header
target_idx = header.index(target)
target_vec = result.copy()
# Make the target the largest logit
target_vec[target_idx] = np.amax(target_vec) + 1
print 'OK, will try to get logit equal to', target_vec[target_idx]
target_vec = np.expand_dims(target_vec, 0)
print target_vec.shape
# Optimize with respect to input
cost = tf.reduce_mean(tf.nn.l2_loss(pred - target_vec))
opt = tf.train.GradientDescentOptimizer(learning_rate=0.001)
opt_op = opt.minimize(cost, var_list=[x_var])
print 'starting optimization'
step = 0
start = time.time()
while True:
sess.run(opt_op, feed_dict={train.keep_prob: 1})
if step % 500 == 0:
loss = sess.run(cost, feed_dict={train.keep_prob: 1})
result = sess.run(pred, feed_dict={train.keep_prob: 1}).flatten()
# print_tags(header, result)
print 'Logit is currently', result[target_idx]
print 'Step', step, 'Loss', loss, 'Time', time.time() - start
if step % 5000 == 0:
librosa.output.write_wav('./transfer' + str(step) + '.wav', x_var.eval(session=sess).flatten(), train.SAMPLE_RATE)
print 'saved'
step += 1
sess.close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model', required=True)
parser.add_argument('--song', required=True)
parser.add_argument('--start', required=True)
args = parser.parse_args()
style_transfer(args.model, args.song, int(args.start))
if __name__ == '__main__':
main()