-
Notifications
You must be signed in to change notification settings - Fork 66
/
Copy pathgenerate.py
196 lines (166 loc) · 9.56 KB
/
generate.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
188
189
190
191
192
193
194
195
196
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import errno
import json
import os
import random
import cv2
import numpy as np
import tensorflow as tf
import sys
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(os.path.realpath(__file__)),'..')))
from video_prediction import datasets, models
from video_prediction.utils.ffmpeg_gif import save_gif
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--input_dir", type=str, required=True, help="either a directory containing subdirectories "
"train, val, test, etc, or a directory containing "
"the tfrecords")
parser.add_argument("--results_dir", type=str, default='results', help="ignored if output_gif_dir is specified")
parser.add_argument("--results_gif_dir", type=str, help="default is results_dir. ignored if output_gif_dir is specified")
parser.add_argument("--results_png_dir", type=str, help="default is results_dir. ignored if output_png_dir is specified")
parser.add_argument("--output_gif_dir", help="output directory where samples are saved as gifs. default is "
"results_gif_dir/model_fname")
parser.add_argument("--output_png_dir", help="output directory where samples are saved as pngs. default is "
"results_png_dir/model_fname")
parser.add_argument("--checkpoint", help="directory with checkpoint or checkpoint name (e.g. checkpoint_dir/model-200000)")
parser.add_argument("--mode", type=str, choices=['val', 'test'], default='val', help='mode for dataset, val or test.')
parser.add_argument("--dataset", type=str, help="dataset class name")
parser.add_argument("--dataset_hparams", type=str, help="a string of comma separated list of dataset hyperparameters")
parser.add_argument("--model", type=str, help="model class name")
parser.add_argument("--model_hparams", type=str, help="a string of comma separated list of model hyperparameters")
parser.add_argument("--batch_size", type=int, default=8, help="number of samples in batch")
parser.add_argument("--num_samples", type=int, help="number of samples in total (all of them by default)")
parser.add_argument("--num_epochs", type=int, default=1)
parser.add_argument("--num_stochastic_samples", type=int, default=5)
parser.add_argument("--gif_length", type=int, help="default is sequence_length")
parser.add_argument("--fps", type=int, default=4)
parser.add_argument("--gpu_mem_frac", type=float, default=0, help="fraction of gpu memory to use")
parser.add_argument("--seed", type=int, default=7)
args = parser.parse_args()
if args.seed is not None:
tf.set_random_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
args.results_gif_dir = args.results_gif_dir or args.results_dir
args.results_png_dir = args.results_png_dir or args.results_dir
dataset_hparams_dict = {}
model_hparams_dict = {}
if args.checkpoint:
checkpoint_dir = os.path.normpath(args.checkpoint)
if not os.path.isdir(args.checkpoint):
checkpoint_dir, _ = os.path.split(checkpoint_dir)
if not os.path.exists(checkpoint_dir):
raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT), checkpoint_dir)
with open(os.path.join(checkpoint_dir, "options.json")) as f:
print("loading options from checkpoint %s" % args.checkpoint)
options = json.loads(f.read())
args.dataset = args.dataset or options['dataset']
args.model = args.model or options['model']
try:
with open(os.path.join(checkpoint_dir, "dataset_hparams.json")) as f:
dataset_hparams_dict = json.loads(f.read())
except FileNotFoundError:
print("dataset_hparams.json was not loaded because it does not exist")
try:
with open(os.path.join(checkpoint_dir, "model_hparams.json")) as f:
model_hparams_dict = json.loads(f.read())
except FileNotFoundError:
print("model_hparams.json was not loaded because it does not exist")
args.output_gif_dir = args.output_gif_dir or os.path.join(args.results_gif_dir, os.path.split(checkpoint_dir)[1])
args.output_png_dir = args.output_png_dir or os.path.join(args.results_png_dir, os.path.split(checkpoint_dir)[1])
else:
if not args.dataset:
raise ValueError('dataset is required when checkpoint is not specified')
if not args.model:
raise ValueError('model is required when checkpoint is not specified')
args.output_gif_dir = args.output_gif_dir or os.path.join(args.results_gif_dir, 'model.%s' % args.model)
args.output_png_dir = args.output_png_dir or os.path.join(args.results_png_dir, 'model.%s' % args.model)
print('----------------------------------- Options ------------------------------------')
for k, v in args._get_kwargs():
print(k, "=", v)
print('------------------------------------- End --------------------------------------')
VideoDataset = datasets.get_dataset_class(args.dataset)
dataset = VideoDataset(
args.input_dir,
mode=args.mode,
num_epochs=args.num_epochs,
seed=args.seed,
hparams_dict=dataset_hparams_dict,
hparams=args.dataset_hparams)
VideoPredictionModel = models.get_model_class(args.model)
hparams_dict = dict(model_hparams_dict)
hparams_dict.update({
'context_frames': dataset.hparams.context_frames,
'sequence_length': dataset.hparams.sequence_length,
'repeat': dataset.hparams.time_shift,
})
model = VideoPredictionModel(
mode=args.mode,
hparams_dict=hparams_dict,
hparams=args.model_hparams)
sequence_length = model.hparams.sequence_length
context_frames = model.hparams.context_frames
future_length = sequence_length - context_frames
if args.num_samples:
if args.num_samples > dataset.num_examples_per_epoch():
raise ValueError('num_samples cannot be larger than the dataset')
num_examples_per_epoch = args.num_samples
else:
num_examples_per_epoch = dataset.num_examples_per_epoch()
if num_examples_per_epoch % args.batch_size != 0:
raise ValueError('batch_size should evenly divide the dataset size %d' % num_examples_per_epoch)
inputs = dataset.make_batch(args.batch_size)
input_phs = {k: tf.placeholder(v.dtype, v.shape, '%s_ph' % k) for k, v in inputs.items()}
with tf.variable_scope(''):
model.build_graph(input_phs)
for output_dir in (args.output_gif_dir, args.output_png_dir):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
with open(os.path.join(output_dir, "options.json"), "w") as f:
f.write(json.dumps(vars(args), sort_keys=True, indent=4))
with open(os.path.join(output_dir, "dataset_hparams.json"), "w") as f:
f.write(json.dumps(dataset.hparams.values(), sort_keys=True, indent=4))
with open(os.path.join(output_dir, "model_hparams.json"), "w") as f:
f.write(json.dumps(model.hparams.values(), sort_keys=True, indent=4))
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_mem_frac)
config = tf.ConfigProto(gpu_options=gpu_options, allow_soft_placement=True)
sess = tf.Session(config=config)
sess.graph.as_default()
model.restore(sess, args.checkpoint)
sample_ind = 0
while True:
if args.num_samples and sample_ind >= args.num_samples:
break
try:
input_results = sess.run(inputs)
except tf.errors.OutOfRangeError:
break
print("evaluation samples from %d to %d" % (sample_ind, sample_ind + args.batch_size))
feed_dict = {input_ph: input_results[name] for name, input_ph in input_phs.items()}
for stochastic_sample_ind in range(args.num_stochastic_samples):
gen_images = sess.run(model.outputs['gen_images'], feed_dict=feed_dict)
# only keep the future frames
gen_images = gen_images[:, -future_length:]
for i, gen_images_ in enumerate(gen_images):
context_images_ = (input_results['images'][i] * 255.0).astype(np.uint8)
gen_images_ = (gen_images_ * 255.0).astype(np.uint8)
gen_images_fname = 'gen_image_%05d_%02d.gif' % (sample_ind + i, stochastic_sample_ind)
context_and_gen_images = list(context_images_[:context_frames]) + list(gen_images_)
if args.gif_length:
context_and_gen_images = context_and_gen_images[:args.gif_length]
save_gif(os.path.join(args.output_gif_dir, gen_images_fname),
context_and_gen_images, fps=args.fps)
gen_image_fname_pattern = 'gen_image_%%05d_%%02d_%%0%dd.png' % max(2, len(str(len(gen_images_) - 1)))
for t, gen_image in enumerate(gen_images_):
gen_image_fname = gen_image_fname_pattern % (sample_ind + i, stochastic_sample_ind, t)
if gen_image.shape[-1] == 1:
gen_image = np.tile(gen_image, (1, 1, 3))
else:
gen_image = cv2.cvtColor(gen_image, cv2.COLOR_RGB2BGR)
cv2.imwrite(os.path.join(args.output_png_dir, gen_image_fname), gen_image)
sample_ind += args.batch_size
if __name__ == '__main__':
main()