-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathprednet.py
187 lines (127 loc) · 6.74 KB
/
prednet.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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import tensorflow as tf
from tensorboard import program
import numpy as np
import cv2
import os, sys, time, argparse, shutil
from tqdm import tqdm
################################### HELPER FUNCTIONS ###################################
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('-model', type=str, default="my_model")
parser.add_argument('-gpu', type=str, default="0")
parser.add_argument('-path', type=str, default="./")
parser.add_argument('-drop', type=float, default=0.0)
parser.add_argument('-reg', type=float, default=0.0)
parser.add_argument('-lr', type=float, default=0.001)
args = parser.parse_args()
return args
def tb_init(logdir = "./logs/"):
if os.path.exists(logdir):
shutil.rmtree(logdir)
train_writer = tf.summary.create_file_writer(logdir + "train/")
test_writer = tf.summary.create_file_writer(logdir+ "test/")
tb = program.TensorBoard()
tb.configure(argv=[None, '--logdir', "./logs/", '--samples_per_plugin', "images=0"])
url = tb.launch()
print("TensorBoard started at URL: {}".format(url))
return train_writer, test_writer
args = parse_args() #options [model, gpu, path, drop, reg, lr]
model_name = args.model
os.environ["CUDA_VISIBLE_DEVICES"]= args.gpu
train_writer, test_writer = tb_init("./logs/" + model_name + "/")
####################################### DEFINE MODELS #######################################
class PredNet (tf.keras.Model):
def __init__(self):
super(PredNet, self).__init__()
self.batch_len = 64
self.hidden_units = 2048
# layer 1
self.ConvLSTM1 = tf.keras.layers.ConvLSTM2D(filters = 6, kernel_size = (3,3), strides = 1, padding = 'same', stateful = False)
self.targetConv1 = tf.keras.layers.Conv2D(filters = 3, kernel_size = (3,3), strides = 1, padding = 'same')
self.predConv1 = tf.keras.layers.Conv2D(filters = 3, kernel_size = (3,3), strides = 1, padding = 'same')
# layer 2
self.ConvLSTM2 = tf.keras.layers.ConvLSTM2D(filters = 128, kernel_size = (3,3), strides = 1, padding = 'same', stateful = False)
self.targetConv2 = tf.keras.layers.Conv2D(filters = 64, kernel_size = (3,3), strides = 1, padding = 'same')
self.predConv2 = tf.keras.layers.Conv2D(filters = 64, kernel_size = (3,3), strides = 1, padding = 'same')
# layer 3
self.ConvLSTM3 = tf.keras.layers.ConvLSTM2D(filters = 256, kernel_size = (3,3), strides = 1, padding = 'same', stateful = False)
self.targetConv3 = tf.keras.layers.Conv2D(filters = 128, kernel_size = (3,3), strides = 1, padding = 'same')
self.predConv3 = tf.keras.layers.Conv2D(filters = 128, kernel_size = (3,3), strides = 1, padding = 'same')
self.relu = tf.keras.layers.ReLU()
self.unpool = tf.keras.layers.UpSampling2D(size = (2,2))
self.pool = tf.keras.layers.MaxPool2D(pool_size = (2,2))
def call(self, x, E1, E2, E3, R1, R2, R3):
# top down updates
output_3 = self.ConvLSTM3(tf.concat([E3, R3], axis = -1))
output_2 = self.ConvLSTM2(tf.concat([E2, R2, tf.expand_dims(self.unpool(output_3), axis = 0)], axis = -1))
output_1 = self.ConvLSTM1(tf.concat([E1, R1, tf.expand_dims(self.unpool(output_2), axis = 0)], axis = -1))
# Regular Updates layer 1
pred1 = self.predConv1(output_1)
pred1 = self.relu(pred1)
pred1_clip = tf.clip_by_value(pred1, 0.0, 1.0)
E1 = self.relu(tf.subtract(pred1, x))
# Regular Updates layer 2
pred2 = self.predConv2(output_2)
pred2 = self.relu(pred2)
E2 = self.relu(tf.subtract(pred2, self.pool(self.relu(self.targetConv2(E1)))))
# Regular Updates layer 3
pred3 = self.predConv3(output_3)
pred3 = self.relu(pred3)
E3 = self.relu(tf.subtract(pred3, self.pool(self.relu(self.targetConv3(E2)))))
return tf.expand_dims(E1, axis = 0), tf.expand_dims(E2, axis = 0), tf.expand_dims(E3, axis = 0), tf.expand_dims(output_1, axis = 0), tf.expand_dims(output_2, axis = 0), tf.expand_dims(output_3, axis = 0), tf.reduce_mean(pred1_clip, axis = 0)
def reset_states(self):
E1 = tf.random.normal((1, 1, 512, 512, 3))
E2 = tf.random.normal((1, 1, 256, 256, 64))
E3 = tf.random.normal((1, 1, 128, 128, 128))
R1 = tf.random.normal((1, 1, 512, 512, 6))
R2 = tf.random.normal((1, 1, 256, 256, 128))
R3 = tf.random.normal((1, 1, 128, 128, 256))
return E1, E2, E3, R1, R2, R3
####################################### BUILD MODELS ########################################
# PredNet model
predictModel = PredNet()
# Optimizer for training
optimizer = tf.keras.optimizers.Adam(learning_rate = args.lr)
####################################### TRAIN STEP #########################################
# @tf.function
def train_step(weight, img, E1, E2, E3, R1, R2, R3):
loss = 0
with tf.GradientTape() as tape:
E1, E2, E3, R1, R2, R3, pred = predictModel(img, E1, E2, E3, R1, R2, R3)
loss = weight * (tf.reduce_mean(E1) + 0.0 * tf.reduce_mean(E2) + 0.0 * tf.reduce_mean(E3))
trainable_variables = predictModel.trainable_variables
gradients = tape.gradient(loss, trainable_variables)
optimizer.apply_gradients(zip(gradients, trainable_variables))
return 1.0, loss, E1, E2, E3, R1, R2, R3, pred
####################################### PREPARE DATASET #####################################
frame_count = 10
snippet_size = 10
vidcap = cv2.VideoCapture(args.path)
num_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
steps = int(num_frames/(snippet_size))
weight = 0.0
E1, E2, E3, R1, R2, R3 = predictModel.reset_states()
training_pbar = tqdm(range(steps), desc='Train Step {:>2}/{} (loss _.___)'.format(0, steps), unit = 'steps', total = steps)
for curr_step in training_pbar:
# Read one snippet
video_snippet = []
for i in range(snippet_size):
success, img = vidcap.read()
img = cv2.resize(img, (512,512))
if success:
video_snippet.append(img)
if not success:
break
for i in range(snippet_size):
# print(tf.image.per_image_standardization(tf.cast(video_snippet[i], tf.float32)))
weight, loss, E1, E2, E3, R1, R2, R3, pred = train_step(weight, tf.image.per_image_standardization(tf.cast(video_snippet[i], tf.float32)), E1, E2, E3, R1, R2, R3)
with train_writer.as_default():
tf.summary.scalar("Loss", loss, step = frame_count)
tf.summary.image("Prediction", tf.expand_dims(pred * 255, axis = 0) , step = frame_count)
tf.summary.image("Target", tf.expand_dims(video_snippet[i], axis = 0) , step = frame_count)
train_writer.flush()
frame_count += snippet_size
train_writer.close()
test_writer.close()
vidcap.release()
print('All done', flush=True)