1 """Train a voxel flow model on ucf101 dataset."""
2 from __future__ import absolute_import
3 from __future__ import division
4 from __future__ import print_function
7 from utils.prefetch_queue_shuffle import PrefetchQueue
10 import tensorflow as tf
11 import tensorflow.contrib.slim as slim
12 from datetime import datetime
14 from random import shuffle
15 from voxel_flow_model import Voxel_flow_model
16 from utils.image_utils import imwrite
17 from functools import partial
20 FLAGS = tf.app.flags.FLAGS
22 # Define necessary FLAGS
23 tf.app.flags.DEFINE_string('train_dir', './voxel_flow_checkpoints/',
24 """Directory where to write event logs """
25 """and checkpoint.""")
26 tf.app.flags.DEFINE_string('train_image_dir', './voxel_flow_train_image/',
27 """Directory where to output images.""")
28 tf.app.flags.DEFINE_string('test_image_dir', './voxel_flow_test_image/',
29 """Directory where to output images.""")
30 tf.app.flags.DEFINE_string('subset', 'train',
31 """Either 'train' or 'validation'.""")
32 tf.app.flags.DEFINE_string('pretrained_model_checkpoint_path', './voxel_flow_checkpoints/',
33 """If specified, restore this pretrained model """
34 """before beginning any training.""")
35 tf.app.flags.DEFINE_integer('max_steps', 10000000,
36 """Number of batches to run.""")
37 tf.app.flags.DEFINE_integer(
38 'batch_size', 32, 'The number of samples in each batch.')
39 tf.app.flags.DEFINE_float('initial_learning_rate', 0.0003,
40 """Initial learning rate.""")
42 def _read_image(filename):
43 image_string = tf.read_file(filename)
44 image_decoded = tf.image.decode_image(image_string, channels=3)
45 image_decoded.set_shape([256, 256, 3])
46 return tf.cast(image_decoded, dtype=tf.float32) / 127.5 - 1.0
48 def train(dataset_frame1, dataset_frame2, dataset_frame3):
50 with tf.Graph().as_default():
52 data_list_frame1 = dataset_frame1.read_data_list_file()
53 dataset_frame1 = tf.data.Dataset.from_tensor_slices(tf.constant(data_list_frame1))
54 dataset_frame1 = dataset_frame1.repeat().shuffle(buffer_size=1000000, seed=1).map(_read_image)
56 data_list_frame2 = dataset_frame2.read_data_list_file()
57 dataset_frame2 = tf.data.Dataset.from_tensor_slices(tf.constant(data_list_frame2))
58 dataset_frame2 = dataset_frame2.repeat().shuffle(buffer_size=1000000, seed=1).map(_read_image)
60 data_list_frame3 = dataset_frame3.read_data_list_file()
61 dataset_frame3 = tf.data.Dataset.from_tensor_slices(tf.constant(data_list_frame3))
62 dataset_frame3 = dataset_frame3.repeat().shuffle(buffer_size=1000000, seed=1).map(_read_image)
64 batch_frame1 = dataset_frame1.batch(FLAGS.batch_size).make_initializable_iterator()
65 batch_frame2 = dataset_frame2.batch(FLAGS.batch_size).make_initializable_iterator()
66 batch_frame3 = dataset_frame3.batch(FLAGS.batch_size).make_initializable_iterator()
68 # Create input and target placeholder.
69 input_placeholder = tf.concat([batch_frame1.get_next(), batch_frame3.get_next()], 3)
70 target_placeholder = batch_frame2.get_next()
72 # input_resized = tf.image.resize_area(input_placeholder, [128, 128])
73 # target_resized = tf.image.resize_area(target_placeholder,[128, 128])
76 model = Voxel_flow_model()
77 prediction = model.inference(input_placeholder)
78 # reproduction_loss, prior_loss = model.loss(prediction, target_placeholder)
79 reproduction_loss = model.loss(prediction, target_placeholder)
80 # total_loss = reproduction_loss + prior_loss
81 total_loss = reproduction_loss
83 # Perform learning rate scheduling.
84 learning_rate = FLAGS.initial_learning_rate
86 # Create an optimizer that performs gradient descent.
87 opt = tf.train.AdamOptimizer(learning_rate)
88 grads = opt.compute_gradients(total_loss)
89 update_op = opt.apply_gradients(grads)
92 summaries = tf.get_collection(tf.GraphKeys.SUMMARIES)
93 summaries.append(tf.summary.scalar('total_loss', total_loss))
94 summaries.append(tf.summary.scalar('reproduction_loss', reproduction_loss))
95 # summaries.append(tf.summary.scalar('prior_loss', prior_loss))
96 summaries.append(tf.summary.image('Input Image (before)', input_placeholder[:, :, :, 0:3], 3));
97 summaries.append(tf.summary.image('Input Image (after)', input_placeholder[:, :, :, 3:6], 3));
98 summaries.append(tf.summary.image('Output Image', prediction, 3))
99 summaries.append(tf.summary.image('Target Image', target_placeholder, 3))
102 saver = tf.train.Saver(tf.all_variables())
104 # Build the summary operation from the last tower summaries.
105 summary_op = tf.summary.merge_all()
107 # Restore checkpoint from file.
108 if FLAGS.pretrained_model_checkpoint_path:
110 assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)
111 ckpt = tf.train.get_checkpoint_state(
112 FLAGS.pretrained_model_checkpoint_path)
113 restorer = tf.train.Saver()
114 restorer.restore(sess, ckpt.model_checkpoint_path)
115 print('%s: Pre-trained model restored from %s' %
116 (datetime.now(), ckpt.model_checkpoint_path))
118 # Build an initialization operation to run below.
119 init = tf.initialize_all_variables()
121 sess.run([init, batch_frame1.initializer, batch_frame2.initializer, batch_frame3.initializer])
124 summary_writer = tf.summary.FileWriter(
128 data_size = len(data_list_frame1)
129 epoch_num = int(data_size / FLAGS.batch_size)
131 for step in range(0, FLAGS.max_steps):
132 batch_idx = step % epoch_num
134 # Run single step update.
135 _, loss_value = sess.run([update_op, total_loss])
138 print('Epoch Number: %d' % int(step / epoch_num))
141 print("Loss at step %d: %f" % (step, loss_value))
145 summary_str = sess.run(summary_op)
146 summary_writer.add_summary(summary_str, step)
149 # Run a batch of images
150 prediction_np, target_np = sess.run([prediction, target_placeholder])
151 for i in range(0,prediction_np.shape[0]):
152 file_name = FLAGS.train_image_dir+str(i)+'_out.png'
153 file_name_label = FLAGS.train_image_dir+str(i)+'_gt.png'
154 imwrite(file_name, prediction_np[i,:,:,:])
155 imwrite(file_name_label, target_np[i,:,:,:])
158 if step % 5000 == 0 or (step +1) == FLAGS.max_steps:
159 checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
160 saver.save(sess, checkpoint_path, global_step=step)
162 def validate(dataset_frame1, dataset_frame2, dataset_frame3):
163 """Performs validation on model.
168 def test(dataset_frame1, dataset_frame2, dataset_frame3):
169 """Perform test on a trained model."""
170 with tf.Graph().as_default():
171 # Create input and target placeholder.
172 input_placeholder = tf.placeholder(tf.float32, shape=(None, 256, 256, 6))
173 target_placeholder = tf.placeholder(tf.float32, shape=(None, 256, 256, 3))
175 # input_resized = tf.image.resize_area(input_placeholder, [128, 128])
176 # target_resized = tf.image.resize_area(target_placeholder,[128, 128])
179 model = Voxel_flow_model(is_train=True)
180 prediction = model.inference(input_placeholder)
181 # reproduction_loss, prior_loss = model.loss(prediction, target_placeholder)
182 reproduction_loss = model.loss(prediction, target_placeholder)
183 # total_loss = reproduction_loss + prior_loss
184 total_loss = reproduction_loss
186 # Create a saver and load.
187 saver = tf.train.Saver(tf.all_variables())
190 # Restore checkpoint from file.
191 if FLAGS.pretrained_model_checkpoint_path:
192 assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)
193 ckpt = tf.train.get_checkpoint_state(
194 FLAGS.pretrained_model_checkpoint_path)
195 restorer = tf.train.Saver()
196 restorer.restore(sess, ckpt.model_checkpoint_path)
197 print('%s: Pre-trained model restored from %s' %
198 (datetime.now(), ckpt.model_checkpoint_path))
200 # Process on test dataset.
201 data_list_frame1 = dataset_frame1.read_data_list_file()
202 data_size = len(data_list_frame1)
203 epoch_num = int(data_size / FLAGS.batch_size)
205 data_list_frame2 = dataset_frame2.read_data_list_file()
207 data_list_frame3 = dataset_frame3.read_data_list_file()
212 for id_img in range(0, data_size):
214 line_image_frame1 = dataset_frame1.process_func(data_list_frame1[id_img])
215 line_image_frame2 = dataset_frame2.process_func(data_list_frame2[id_img])
216 line_image_frame3 = dataset_frame3.process_func(data_list_frame3[id_img])
218 batch_data_frame1 = [dataset_frame1.process_func(ll) for ll in data_list_frame1[0:63]]
219 batch_data_frame2 = [dataset_frame2.process_func(ll) for ll in data_list_frame2[0:63]]
220 batch_data_frame3 = [dataset_frame3.process_func(ll) for ll in data_list_frame3[0:63]]
222 batch_data_frame1.append(line_image_frame1)
223 batch_data_frame2.append(line_image_frame2)
224 batch_data_frame3.append(line_image_frame3)
226 batch_data_frame1 = np.array(batch_data_frame1)
227 batch_data_frame2 = np.array(batch_data_frame2)
228 batch_data_frame3 = np.array(batch_data_frame3)
230 feed_dict = {input_placeholder: np.concatenate((batch_data_frame1, batch_data_frame3), 3),
231 target_placeholder: batch_data_frame2}
232 # Run single step update.
233 prediction_np, target_np, loss_value = sess.run([prediction,
236 feed_dict = feed_dict)
237 print("Loss for image %d: %f" % (i,loss_value))
238 file_name = FLAGS.test_image_dir+str(i)+'_out.png'
239 file_name_label = FLAGS.test_image_dir+str(i)+'_gt.png'
240 imwrite(file_name, prediction_np[-1,:,:,:])
241 imwrite(file_name_label, target_np[-1,:,:,:])
243 PSNR += 10*np.log10(255.0*255.0/np.sum(np.square(prediction_np-target_np)))
244 print("Overall PSNR: %f db" % (PSNR/len(data_list)))
246 if __name__ == '__main__':
248 os.environ["CUDA_VISIBLE_DEVICES"] = "0"
250 if FLAGS.subset == 'train':
252 data_list_path_frame1 = "data_list/ucf101_train_files_frame1.txt"
253 data_list_path_frame2 = "data_list/ucf101_train_files_frame2.txt"
254 data_list_path_frame3 = "data_list/ucf101_train_files_frame3.txt"
256 ucf101_dataset_frame1 = dataset.Dataset(data_list_path_frame1)
257 ucf101_dataset_frame2 = dataset.Dataset(data_list_path_frame2)
258 ucf101_dataset_frame3 = dataset.Dataset(data_list_path_frame3)
260 train(ucf101_dataset_frame1, ucf101_dataset_frame2, ucf101_dataset_frame3)
262 elif FLAGS.subset == 'test':
264 data_list_path_frame1 = "data_list/ucf101_test_files_frame1.txt"
265 data_list_path_frame2 = "data_list/ucf101_test_files_frame2.txt"
266 data_list_path_frame3 = "data_list/ucf101_test_files_frame3.txt"
268 ucf101_dataset_frame1 = dataset.Dataset(data_list_path_frame1)
269 ucf101_dataset_frame2 = dataset.Dataset(data_list_path_frame2)
270 ucf101_dataset_frame3 = dataset.Dataset(data_list_path_frame3)
272 test(ucf101_dataset_frame1, ucf101_dataset_frame2, ucf101_dataset_frame3)