-
Notifications
You must be signed in to change notification settings - Fork 1
/
decodeimagesacresnet.py
159 lines (132 loc) · 6.69 KB
/
decodeimagesacresnet.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
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
from datetime import datetime
from dataloader.outdoor_data_mfcc import ActionsDataLoader as SoundDataLoader
from dataloader.actions_data_old import ActionsDataLoader
from models.unet_acresnet import UNetAc
from models.vision import ResNet50Model
import numpy as np
import tensorflow as tf
import os
import matplotlib.pyplot as plt
flags = tf.app.flags
slim = tf.contrib.slim
flags.DEFINE_string('model', None, 'Model type, it can AudioCoeff')
flags.DEFINE_string('datatype', 'outdoor', 'music or outdoor or old')
flags.DEFINE_string('train_file', None, 'File for training data')
flags.DEFINE_string('init_checkpoint', None, 'Checkpoint file for model initialization')
flags.DEFINE_integer('batch_size', 2, 'Batch size choose')
flags.DEFINE_integer('sample_length', 1, 'Length in seconds of a sequence sample')
FLAGS = flags.FLAGS
'''Plot reconstructed MFCC'''
def main(_):
plotdecodeimages()
def plotdecodeimages():
dataset = FLAGS.train_file.split('/')[-1]
dataset = dataset.split('.')[0]
s = FLAGS.init_checkpoint.split('/')[-1]
name = (s.split('_')[1]).split('.ckpt')[0]
name = '{}_{}_{}_{}'.format(FLAGS.model, dataset, 'Acoustic', name)
data_dir = str.join('/', FLAGS.init_checkpoint.split('/')[:-1] + [name])
random_pick = True
build_spectrogram = True
normalize = False
# Create data loaders according to the received program arguments
print('{} - Creating data loaders'.format(datetime.now()))
modalities = []
modalities.append(0)
modalities.append(1)
modalities.append(2)
with tf.device('/cpu:0'):
if FLAGS.datatype == 'old':
train_data = ActionsDataLoader(FLAGS.train_file, 'testing', batch_size=FLAGS.batch_size, num_epochs=1, sample_length=1,
datakind=FLAGS.datatype, buffer_size=10, shuffle=False, embedding=1,
normalize=normalize, build_spectrogram=build_spectrogram, correspondence=0,
random_pick=random_pick, modalities=modalities, nr_frames=1)
elif FLAGS.datatype == 'outdoor':
train_data = SoundDataLoader(FLAGS.train_file, 'testing', batch_size=FLAGS.batch_size, num_epochs=1, sample_length=1,
datakind=FLAGS.datatype, buffer_size=10, shuffle=False, embedding=1,
normalize=normalize, build_spectrogram=build_spectrogram, correspondence=0,
random_pick=random_pick, modalities=modalities, nr_frames=1)
# Build model
print('{} - Building model'.format(datetime.now()))
with tf.device('/gpu:0'):
model = UNetAc(input_shape=[36, 48, 12])
model_video = ResNet50Model(input_shape=[224, 298, 3], num_classes=None)
handle = tf.placeholder(tf.string, shape=())
iterator = tf.data.Iterator.from_string_handle(handle, train_data.data.output_types,
train_data.data.output_shapes)
train_iterat = train_data.data.make_initializable_iterator()
next_batch = iterator.get_next()
mfcc = tf.reshape(next_batch[1], shape=[-1, 12])
images = tf.reshape(next_batch[2], shape=[-1, 224, 298, 3])
acoustic = tf.reshape(next_batch[0], shape=[-1, 36, 48, 12])
# mfcc = mfcc - tf.reduce_min(mfcc, axis=[1], keep_dims=True)
# mfcc = mfcc / tf.reduce_max(mfcc, axis=[1], keep_dims=True)
mfccmap = tf.reshape(mfcc, (-1, 1, 12))
mfccmap = tf.tile(mfccmap, (1, 36 * 48, 1))
mfccmap = tf.reshape(mfccmap, (-1, 36, 48, 12))
model_video._build_model(images)
model._build_model(mfccmap, model_video.output)
output = model.output
var_list1 = slim.get_variables(model_video.scope + '/')
var_list2 = slim.get_variables(model.scope + '/')
var_list = var_list2 + var_list1
if os.path.exists(data_dir):
print("Features already computed!")
else:
os.makedirs(data_dir) # mkdir creates one directory, makedirs all intermediate directories
total_size = 0
batch_count = 0
num = 0
print('{} - Starting'.format(datetime.now()))
namesimage = ['Acoustic image', 'Reconstructed']
with tf.Session(
config=tf.ConfigProto(allow_soft_placement=True, gpu_options=tf.GPUOptions(allow_growth=True))) as session:
train_handle = session.run(train_iterat.string_handle())
# Initialize student model
if FLAGS.init_checkpoint is None:
print('{} - Initializing student model'.format(datetime.now()))
model.init_model(session, FLAGS.init_checkpoint)
print('{} - Done'.format(datetime.now()))
else:
print('{} - Restoring student model'.format(datetime.now()))
saver = tf.train.Saver(var_list=var_list)
saver.restore(session, FLAGS.init_checkpoint)
print('{} - Done'.format(datetime.now()))
#variables_in_checkpoint = tf.train.list_variables(FLAGS.init_checkpoint)
session.run(train_iterat.initializer)
while True:
try:
data, reconstructed = session.run(
[acoustic, output],
feed_dict={handle: train_handle,
model.network['keep_prob']: 1.0,
model.network['is_training']: 0,
model_video.network['keep_prob']: 1.0,
model_video.network['is_training']: 0
})
total_size += reconstructed.shape[0]
for h in range(np.shape(reconstructed)[0]):
# original and reconstructed
fig, axs = plt.subplots(4, 2, figsize=(6, 2.9 * 4))
plt.tight_layout(pad=1.0)
fig.suptitle('Reconstructed image')
imagesvideo = np.stack((data, reconstructed), 0)
for i in range(2):
for j in range(4):
x = j
y = i
axs[x, y].imshow(imagesvideo[i, h, :, :, j * 3:(j + 1) * 3])
axs[x, y].axis('off')
axs[x, y].set_title('{}'.format(namesimage[i]))
outImage_path = '{}/{}_images_{}.png'.format(data_dir, dataset, num)
plt.savefig(outImage_path)
plt.clf()
num = num + 1
print(total_size)
except tf.errors.OutOfRangeError:
break
batch_count += 1
print('{} - Completed, got {} samples'.format(datetime.now(), total_size))
if __name__ == '__main__':
flags.mark_flags_as_required(['train_file'])
tf.app.run()