def trainLowMemory(self):
print("TRAINING IS STARTING RIGHT NOW!")
for epoch_num in range(1, MAX_EPOCHS + 1):
epoch_start_time = time.time()
currentSaveRate = SAVE_EVERY
for step in range(self.dataset.num_batches_train):
start_time = time.time()
# get data
batch_x, batch_y = self.sess.run(
[self.dataset.train_images, self.dataset.train_labels])
# train towers logits
logits = []
for sequence_image in range(self.dataset.frames):
feed_dict = {self.towerImage: batch_x[:, sequence_image]}
eval_tensors = self.towerNet
logits.append(self.sess.run(eval_tensors, feed_dict))
logits = np.transpose(np.array(logits), [1, 0, 2])
feed_dict = {
self.towerLogits: logits,
self.Y: batch_y,
self.is_training: True
}
if (step + 1) * BATCH_SIZE % 5000 == 0:
# optimiza + rest of network + summaries
eval_tensors = [
self.loss, self.train_op, self.merged_summary_op
]
loss_val, _, merged_ops = self.sess.run(
eval_tensors, feed_dict=feed_dict)
self.summary_train_train_writer.add_summary(
merged_ops, self.global_step.eval(session=self.sess))
else:
# optimize + rest of network
eval_tensors = [self.loss, self.train_op]
loss_val, _ = self.sess.run(eval_tensors, feed_dict)
duration = time.time() - start_time
util.log_step(epoch_num, step, duration, loss_val, BATCH_SIZE,
self.dataset.num_train_examples, LOG_EVERY)
if (step / self.dataset.num_batches_train) >= currentSaveRate:
self.saver.save(self.sess, self.ckptPrefix)
currentSaveRate += SAVE_EVERY
epoch_time = time.time() - epoch_start_time
print("Total epoch time training: {}".format(epoch_time))
# self.startValidation(epoch_num, epoch_time)
self.finishTraining()
def trainLowMemory(self):
print("TRAINING IS STARTING RIGHT NOW!")
for epoch_num in range(1, MAX_EPOCHS + 1):
epoch_start_time = time.time()
currentSaveRate = SAVE_EVERY
for step in range(self.dataset.num_batches_train):
start_time = time.time()
# get data
batch_x, batch_y = self.sess.run([self.dataset.train_images, self.dataset.train_labels])
# train towers logits
logits = []
for sequence_image in range(self.dataset.frames):
feed_dict = {self.towerImage: batch_x[:, sequence_image], self.is_training: True}
eval_tensors = self.towerNet
logits.append(self.sess.run(eval_tensors, feed_dict))
logits = np.transpose(np.array(logits), [1, 0, 2, 3, 4])
# logits
feed_dict = {self.towerLogits: logits, self.is_training: False}
eval_tensors = self.logits
logits = self.sess.run(eval_tensors, feed_dict)
# optimize
feed_dict = {self.optLogits: logits, self.Y: batch_y, self.is_training: True}
if (step + 1) * BATCH_SIZE % 5000 == 0:
eval_tensors = [self.loss, self.train_op, self.merged_summary_op]
loss_val, _, merged_ops = self.sess.run(eval_tensors, feed_dict=feed_dict)
self.summary_train_train_writer.add_summary(merged_ops, self.global_step.eval(session=self.sess))
else:
eval_tensors = [self.loss, self.train_op]
loss_val, _ = self.sess.run(eval_tensors, feed_dict)
duration = time.time() - start_time
util.log_step(epoch_num, step, duration, loss_val, BATCH_SIZE, self.dataset.num_train_examples,
LOG_EVERY)
if (step / self.dataset.num_batches_train) >= currentSaveRate:
self.saver.save(self.sess, self.ckptPrefix)
currentSaveRate += SAVE_EVERY
epoch_time = time.time() - epoch_start_time
print("Total epoch time training: {}".format(epoch_time))
self.startValidationLowMemory(epoch_num, epoch_time)
self.finishTraining()
def train(self):
print("TRAINING IS STARTING RIGHT NOW!")
for epoch_num in range(1, MAX_EPOCHS + 1):
epoch_start_time = time.time()
currentSaveRate = SAVE_EVERY
for step in range(self.dataset.num_batches_train):
start_time = time.time()
batch_x, batch_y = self.sess.run(
[self.dataset.train_images, self.dataset.train_labels])
feed_dict = {
self.is_training: True,
self.X: batch_x,
self.Y: batch_y
}
eval_tensors = [self.loss, self.train_op]
if (step + 1) * BATCH_SIZE % LOG_EVERY == 0:
eval_tensors += [self.merged_summary_op]
eval_ret = self.sess.run(eval_tensors, feed_dict=feed_dict)
eval_ret = dict(zip(eval_tensors, eval_ret))
loss_val = eval_ret[self.loss]
if self.merged_summary_op in eval_tensors:
self.summary_train_train_writer.add_summary(
eval_ret[self.merged_summary_op],
self.global_step.eval(session=self.sess))
duration = time.time() - start_time
util.log_step(epoch_num, step, duration, loss_val, BATCH_SIZE,
self.dataset.num_train_examples, LOG_EVERY)
if (step /
self.dataset.num_batches_train) >= (currentSaveRate):
self.saver.save(self.sess, self.ckptPrefix)
currentSaveRate += SAVE_EVERY
epoch_time = time.time() - epoch_start_time
print("Total epoch time training: {}".format(epoch_time))
self.startValidation(epoch_num, epoch_time)
self.finishTraining()
def train(self):
print("TRAINING IS STARTING RIGHT NOW!")
for epoch_num in range(1, MAX_EPOCHS + 1):
epoch_start_time = time.time()
currentSaveRate = SAVE_EVERY
for step in range(self.dataset.num_batches_train):
start_time = time.time()
batch_x, batch_y = self.sess.run([self.dataset.train_images, self.dataset.train_labels])
feed_dict = {self.is_training: True, self.X: batch_x, self.Y: batch_y}
eval_tensors = [self.loss, self.train_op]
if (step + 1) * BATCH_SIZE % LOG_EVERY == 0:
eval_tensors += [self.merged_summary_op]
eval_ret = self.sess.run(eval_tensors, feed_dict=feed_dict)
eval_ret = dict(zip(eval_tensors, eval_ret))
loss_val = eval_ret[self.loss]
if self.merged_summary_op in eval_tensors:
self.summary_train_train_writer.add_summary(eval_ret[self.merged_summary_op],
self.global_step.eval(session=self.sess))
duration = time.time() - start_time
util.log_step(epoch_num, step, duration, loss_val, BATCH_SIZE, self.dataset.num_train_examples,
LOG_EVERY)
if (step / self.dataset.num_batches_train) >= (currentSaveRate):
self.saver.save(self.sess, self.ckptPrefix)
currentSaveRate += SAVE_EVERY
epoch_time = time.time() - epoch_start_time
print("Total epoch time training: {}".format(epoch_time))
self.startValidation(epoch_num, epoch_time)
self.finishTraining()
