git.sesse.net Git - voxel-flow/blob - voxel_flow_train.py

   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
   5
   6 import dataset
   7 from utils.prefetch_queue_shuffle import PrefetchQueue
   8 import numpy as np
   9 import os
  10 import tensorflow as tf
  11 import tensorflow.contrib.slim as slim
  12 from datetime import datetime
  13 import random
  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
  18 import pdb
  19
  20 FLAGS = tf.app.flags.FLAGS
  21
  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.""")
  41
  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
  47
  48 def train(dataset_frame1, dataset_frame2, dataset_frame3):
  49   """Trains a model."""
  50   with tf.Graph().as_default():
  51     # Create input.
  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)
  55
  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)
  59
  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)
  63
  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()
  67
  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()
  71
  72     # input_resized = tf.image.resize_area(input_placeholder, [128, 128])
  73     # target_resized = tf.image.resize_area(target_placeholder,[128, 128])
  74
  75     # Prepare model.
  76     model = Voxel_flow_model()
  77     prediction, flow = 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
  82
  83     # Perform learning rate scheduling.
  84     learning_rate = FLAGS.initial_learning_rate
  85
  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)
  90
  91     # Create summaries
  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))
 100     summaries.append(tf.summary.image('Flow', flow, 3))
 101
 102     # Create a saver.
 103     saver = tf.train.Saver(tf.all_variables())
 104
 105     # Build the summary operation from the last tower summaries.
 106     summary_op = tf.summary.merge_all()
 107
 108     # Restore checkpoint from file.
 109     if FLAGS.pretrained_model_checkpoint_path:
 110       sess = tf.Session()
 111       assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)
 112       ckpt = tf.train.get_checkpoint_state(
 113                FLAGS.pretrained_model_checkpoint_path)
 114       restorer = tf.train.Saver()
 115       restorer.restore(sess, ckpt.model_checkpoint_path)
 116       print('%s: Pre-trained model restored from %s' %
 117         (datetime.now(), ckpt.model_checkpoint_path))
 118     else:
 119       # Build an initialization operation to run below.
 120       init = tf.initialize_all_variables()
 121       sess = tf.Session()
 122       sess.run([init, batch_frame1.initializer, batch_frame2.initializer, batch_frame3.initializer])
 123
 124     # Summary Writter
 125     summary_writer = tf.summary.FileWriter(
 126       FLAGS.train_dir,
 127       graph=sess.graph)
 128
 129     data_size = len(data_list_frame1)
 130     epoch_num = int(data_size / FLAGS.batch_size)
 131
 132     for step in range(0, FLAGS.max_steps):
 133       batch_idx = step % epoch_num
 134
 135       # Run single step update.
 136       _, loss_value = sess.run([update_op, total_loss])
 137
 138       if batch_idx == 0:
 139         print('Epoch Number: %d' % int(step / epoch_num))
 140
 141       if step % 10 == 0:
 142         print("Loss at step %d: %f" % (step, loss_value))
 143
 144       if step % 100 == 0:
 145         # Output Summary
 146         summary_str = sess.run(summary_op)
 147         summary_writer.add_summary(summary_str, step)
 148
 149       if step % 500 == 0:
 150         # Run a batch of images
 151         prediction_np, target_np = sess.run([prediction, target_placeholder])
 152         for i in range(0,prediction_np.shape[0]):
 153           file_name = FLAGS.train_image_dir+str(i)+'_out.png'
 154           file_name_label = FLAGS.train_image_dir+str(i)+'_gt.png'
 155           imwrite(file_name, prediction_np[i,:,:,:])
 156           imwrite(file_name_label, target_np[i,:,:,:])
 157
 158       # Save checkpoint
 159       if step % 5000 == 0 or (step +1) == FLAGS.max_steps:
 160         checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
 161         saver.save(sess, checkpoint_path, global_step=step)
 162
 163 def validate(dataset_frame1, dataset_frame2, dataset_frame3):
 164   """Performs validation on model.
 165   Args:
 166   """
 167   pass
 168
 169 def test(dataset_frame1, dataset_frame2, dataset_frame3):
 170   """Perform test on a trained model."""
 171   with tf.Graph().as_default():
 172                 # Create input and target placeholder.
 173     input_placeholder = tf.placeholder(tf.float32, shape=(None, 256, 256, 6))
 174     target_placeholder = tf.placeholder(tf.float32, shape=(None, 256, 256, 3))
 175
 176     # input_resized = tf.image.resize_area(input_placeholder, [128, 128])
 177     # target_resized = tf.image.resize_area(target_placeholder,[128, 128])
 178
 179     # Prepare model.
 180     model, flow = Voxel_flow_model(is_train=True)
 181     prediction = model.inference(input_placeholder)
 182     # reproduction_loss, prior_loss = model.loss(prediction, target_placeholder)
 183     reproduction_loss = model.loss(prediction, target_placeholder)
 184     # total_loss = reproduction_loss + prior_loss
 185     total_loss = reproduction_loss
 186
 187     # Create a saver and load.
 188     saver = tf.train.Saver(tf.all_variables())
 189     sess = tf.Session()
 190
 191     # Restore checkpoint from file.
 192     if FLAGS.pretrained_model_checkpoint_path:
 193       assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)
 194       ckpt = tf.train.get_checkpoint_state(
 195                FLAGS.pretrained_model_checkpoint_path)
 196       restorer = tf.train.Saver()
 197       restorer.restore(sess, ckpt.model_checkpoint_path)
 198       print('%s: Pre-trained model restored from %s' %
 199         (datetime.now(), ckpt.model_checkpoint_path))
 200
 201     # Process on test dataset.
 202     data_list_frame1 = dataset_frame1.read_data_list_file()
 203     data_size = len(data_list_frame1)
 204     epoch_num = int(data_size / FLAGS.batch_size)
 205
 206     data_list_frame2 = dataset_frame2.read_data_list_file()
 207
 208     data_list_frame3 = dataset_frame3.read_data_list_file()
 209
 210     i = 0
 211     PSNR = 0
 212
 213     for id_img in range(0, data_size):
 214       # Load single data.
 215       line_image_frame1 = dataset_frame1.process_func(data_list_frame1[id_img])
 216       line_image_frame2 = dataset_frame2.process_func(data_list_frame2[id_img])
 217       line_image_frame3 = dataset_frame3.process_func(data_list_frame3[id_img])
 218
 219       batch_data_frame1 = [dataset_frame1.process_func(ll) for ll in data_list_frame1[0:63]]
 220       batch_data_frame2 = [dataset_frame2.process_func(ll) for ll in data_list_frame2[0:63]]
 221       batch_data_frame3 = [dataset_frame3.process_func(ll) for ll in data_list_frame3[0:63]]
 222
 223       batch_data_frame1.append(line_image_frame1)
 224       batch_data_frame2.append(line_image_frame2)
 225       batch_data_frame3.append(line_image_frame3)
 226
 227       batch_data_frame1 = np.array(batch_data_frame1)
 228       batch_data_frame2 = np.array(batch_data_frame2)
 229       batch_data_frame3 = np.array(batch_data_frame3)
 230
 231       feed_dict = {input_placeholder: np.concatenate((batch_data_frame1, batch_data_frame3), 3),
 232                     target_placeholder: batch_data_frame2}
 233       # Run single step update.
 234       prediction_np, target_np, loss_value = sess.run([prediction,
 235                                                       target_placeholder,
 236                                                       total_loss],
 237                                                       feed_dict = feed_dict)
 238       print("Loss for image %d: %f" % (i,loss_value))
 239       file_name = FLAGS.test_image_dir+str(i)+'_out.png'
 240       file_name_label = FLAGS.test_image_dir+str(i)+'_gt.png'
 241       imwrite(file_name, prediction_np[-1,:,:,:])
 242       imwrite(file_name_label, target_np[-1,:,:,:])
 243       i += 1
 244       PSNR += 10*np.log10(255.0*255.0/np.sum(np.square(prediction_np-target_np)))
 245     print("Overall PSNR: %f db" % (PSNR/len(data_list)))
 246
 247 if __name__ == '__main__':
 248
 249   os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 250
 251   if FLAGS.subset == 'train':
 252
 253     data_list_path_frame1 = "data_list/ucf101_train_files_frame1.txt"
 254     data_list_path_frame2 = "data_list/ucf101_train_files_frame2.txt"
 255     data_list_path_frame3 = "data_list/ucf101_train_files_frame3.txt"
 256
 257     ucf101_dataset_frame1 = dataset.Dataset(data_list_path_frame1)
 258     ucf101_dataset_frame2 = dataset.Dataset(data_list_path_frame2)
 259     ucf101_dataset_frame3 = dataset.Dataset(data_list_path_frame3)
 260
 261     train(ucf101_dataset_frame1, ucf101_dataset_frame2, ucf101_dataset_frame3)
 262
 263   elif FLAGS.subset == 'test':
 264
 265     data_list_path_frame1 = "data_list/ucf101_test_files_frame1.txt"
 266     data_list_path_frame2 = "data_list/ucf101_test_files_frame2.txt"
 267     data_list_path_frame3 = "data_list/ucf101_test_files_frame3.txt"
 268
 269     ucf101_dataset_frame1 = dataset.Dataset(data_list_path_frame1)
 270     ucf101_dataset_frame2 = dataset.Dataset(data_list_path_frame2)
 271     ucf101_dataset_frame3 = dataset.Dataset(data_list_path_frame3)
 272
 273     test(ucf101_dataset_frame1, ucf101_dataset_frame2, ucf101_dataset_frame3)