def create_graph(self, image_shape, num_classes):
start = time.time()
self.image_shape = image_shape
self.num_classes = num_classes
self.num_layers = len(self.config.keys()) - 1
self.global_step = tf.Variable(0,
name='last_successful_epoch',
trainable=False,
dtype=tf.int32)
self.last_epoch = tf.assign(self.global_step,
self.global_step + 1,
name='assign_updated_epoch')
# Step 1: Creating placeholders for inputs
self.make_placeholders_for_inputs()
# Step 2: Creating initial parameters for the variables
self.make_parameters()
# Step 3: Make predictions for the data
self.make_predictions()
# Step 4: Perform optimization operation
self.make_optimization()
# Step 5: Calculate accuracies
self.make_accuracy()
# Step 6: Initialize all the required variables
with tf.device(self.device):
self.init_var = tf.global_variables_initializer()
# Step 7: Initiate Session
config = tf.ConfigProto(
log_device_placement=True,
allow_soft_placement=True,
)
# config.gpu_options.allow_growth = True
# config.gpu_options.per_process_gpu_memory_fraction = 0.4
if self.session_type == 'default':
self.session = tf.Session(config=config)
if self.session_type == 'interactive':
self.session = tf.InteractiveSession(config=config)
print('Session: ' + str(self.session))
self.session.run(self.init_var)
# Step 8: Initiate logs
if self.logging is True:
file_utils.mkdir_p(self.logging_dir)
self.merged_summary_op = tf.summary.merge_all()
if self.restore is False:
file_utils.delete_all_files_in_dir(self.logging_dir)
self.summary_writer = \
tf.summary.FileWriter(self.logging_dir, graph=self.session.graph)
if self.save_checkpoint is True:
self.model = tf.train.Saver(max_to_keep=1)
# Step 9: Restore model
if self.restore is True:
self.restore_model()
epoch = self.session.run(self.global_step)
print('Model has been trained for %d iterations' % epoch)
end = time.time()
print('Tensorflow graph created in %.4f seconds' % (end - start))
return True
def start_session(self, graph):
# Step 1: Initialize all the required variables
with graph.as_default():
self.global_step = tf.Variable(0,
name='last_successful_epoch',
trainable=False,
dtype=tf.int32)
self.last_epoch = tf.assign(self.global_step,
self.global_step + 1,
name='assign_updated_epoch')
with tf.device(self.device):
self.init_var = tf.global_variables_initializer()
# Step 2: Initiate Session
config = tf.ConfigProto(
log_device_placement=True,
allow_soft_placement=True,
)
if self.session_type == 'default':
self.session = tf.Session(config=config, graph=graph)
if self.session_type == 'interactive':
self.session = tf.InteractiveSession(config=config, graph=graph)
print('Session: ' + str(self.session))
self.session.run(self.init_var)
# Step 3: Initiate logs
if self.logging is True:
file_utils.mkdir_p(self.logging_dir)
self.merged_summary_op = tf.summary.merge_all()
if self.restore is False:
file_utils.delete_all_files_in_dir(self.logging_dir)
self.summary_writer = \
tf.summary.FileWriter(self.logging_dir, graph=graph)
if self.save_model is True:
self.model = tf.train.Saver(max_to_keep=2)
# Step 4: Restore model
if self.restore is True:
self.restore_model()
epoch = self.session.run(self.global_step)
print('Model has been trained for %d iterations' % epoch)
return True
def download_and_extract_data(self, data_directory):
"""
:param data_directory:
:return:
"""
print('Downloading and extracting CIFAR 100 file')
## Step 1: Make the directories './datasets/cifar100/' if they do not exist
if not os.path.exists(data_directory):
if self.verbose is True:
print('Creating the directory \'%s\'' % data_directory)
file_utils.mkdir_p(data_directory)
else:
if self.verbose is True:
print('Directory \'%s\' already exists' % data_directory)
## Step 2: Check if './datasets/cifar100/cifar-100-python.tar.gz' exists
tar_file = data_directory + 'cifar-100.tar.gz'
make_tar = False
if not os.path.exists(tar_file):
make_tar = True
elif os.path.exists(tar_file) and not file_utils.verify_md5(
tar_file, self.file_md5):
if self.verbose is True:
print('Removing the wrong file \'%s\'' % tar_file)
os.remove(tar_file)
make_tar = True
else:
if self.verbose is True:
print('CIFAR 100 tarfile exists and MD5 sum is verified')
## Step 3: Download CIFAR 100 dataset from 'http://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz'
if make_tar is True:
result = file_utils.download(self.file_url,
tar_file,
verbose=self.verbose)
if result is False:
if self.verbose is True:
raise FileNotFoundError(
'Download of CIFAR 100 dataset failed')
return False
result = file_utils.verify_md5(tar_file,
self.file_md5,
verbose=self.verbose)
if result is False:
if self.verbose is True:
raise FileNotFoundError(
'Downloaded CIFAR 100 dataset failed md5sum check')
return False
## Step 4: Extract the dataset
make_extract = False
batches_directory = data_directory + 'cifar-100-batches'
if not os.path.exists(batches_directory):
make_extract = True
else:
num_files = sum(
os.path.isfile(os.path.join(batches_directory, f))
for f in os.listdir(batches_directory))
if num_files != 8:
shutil.rmtree(batches_directory)
make_extract = True
else:
if self.verbose is True:
print('Directory %s already exists' % batches_directory)
if make_extract is True:
print('Extracting file %s to %s' % (tar_file, batches_directory))
result = file_utils.extract(tar_file)
shutil.move('./cifar-100-python', batches_directory)
if result is False:
if self.verbose is True:
print('Extraction of CIFAR 100 dataset failed')
return False
else:
if self.verbose is True:
print('Extraction of CIFAR 100 dataset success')
return True
def run(self,
train_data,
train_labels,
train_classes,
validate_data=None,
validate_labels=None,
validate_classes=None,
test_data=None,
test_labels=None,
test_classes=None):
if self.train_validate_split is not None:
feed_dict_validate = {
self.x: validate_data,
self.y_true: validate_labels,
self.y_true_cls: validate_classes
}
if self.test_log is True:
feed_dict_test = {
self.x: test_data,
self.y_true: test_labels,
self.y_true_cls: test_classes
}
epoch = self.session.run(self.global_step)
print('Last successful epoch: ' + str(epoch))
converged = False
prev_cost = 0
start = time.time()
end_batch_index = 0
num_batches = int(train_data.shape[0] / self.batch_size)
while (epoch != self.max_iterations) and converged is False:
start_batch_index = 0
for batch in range(num_batches):
# print('Training on batch %d' %batch)
end_batch_index = start_batch_index + self.batch_size
if end_batch_index < train_data.shape[0]:
train_batch_data = train_data[
start_batch_index:end_batch_index, :]
train_batch_labels = train_labels[
start_batch_index:end_batch_index, :]
train_batch_classes = train_classes[
start_batch_index:end_batch_index]
else:
train_batch_data = train_data[start_batch_index:, :]
train_batch_labels = train_labels[start_batch_index:, :]
train_batch_classes = train_classes[start_batch_index:]
feed_dict_train = {
self.x: train_batch_data,
self.y_true: train_batch_labels,
self.y_true_cls: train_batch_classes
}
_, cost, train_acc, curr_epoch = self.session.run(
[
self.optimizer, self.loss, self.train_accuracy,
self.last_epoch
],
feed_dict=feed_dict_train)
train_loss_summary = self.session.run(
self.train_loss_summary, feed_dict=feed_dict_train)
train_acc_summary = self.session.run(self.train_acc_summary,
feed_dict=feed_dict_train)
start_batch_index += self.batch_size
if self.train_validate_split is not None:
validate_acc, validate_summary = \
self.session.run([self.validate_accuracy, self.validate_acc_summary],
feed_dict=feed_dict_validate)
if self.test_log is True:
test_acc, test_summary = \
self.session.run([self.test_accuracy, self.test_acc_summary],
feed_dict=feed_dict_test)
if self.separate_writer is False:
self.summary_writer.add_summary(train_loss_summary, epoch)
self.summary_writer.add_summary(train_acc_summary, epoch)
self.summary_writer.add_summary(validate_summary, epoch)
self.summary_writer.add_summary(test_summary, epoch)
else:
self.train_writer.add_summary(train_loss_summary, epoch)
self.train_writer.add_summary(train_acc_summary, epoch)
if self.train_validate_split is not None:
self.validate_writer.add_summary(validate_summary, epoch)
if self.test_log is True:
self.test_writer.add_summary(test_summary, epoch)
if epoch % self.display_step == 0:
duration = time.time() - start
if self.train_validate_split is not None and self.test_log is False:
print(
'>>> Epoch [%*d/%*d] | Error: %.4f | Train Acc.: %.4f | Validate Acc.: %.4f | '
'Duration: %.4f seconds' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations, cost,
train_acc, validate_acc, duration))
elif self.train_validate_split is not None and self.test_log is True:
print(
'>>> Epoch [%*d/%*d] | Error: %.4f | Train Acc.: %.4f | Validate Acc.: %.4f | '
'Test Acc.: %.4f | Duration: %.4f seconds' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations, cost,
train_acc, validate_acc, test_acc, duration))
elif self.train_validate_split is None and self.test_log is True:
print(
'>>> Epoch [%*d/%*d] | Error: %.4f | Train Acc.: %.4f | '
'Test Acc.: %.4f | Duration: %.4f seconds' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations, cost,
train_acc, test_acc, duration))
else:
print(
'>>> Epoch [%*d/%*d] | Error: %.4f | Train Acc.: %.4f | Duration of run: %.4f seconds'
% (int(len(str(self.max_iterations))), epoch,
int(len(str(self.max_iterations))),
self.max_iterations, cost, train_acc))
start = time.time()
if self.save_model is True:
model_directory = os.path.dirname(self.model_name)
file_utils.mkdir_p(model_directory)
self.model.save(self.session,
self.model_name,
global_step=epoch)
if epoch == 0:
prev_cost = cost
else:
if math.fabs(cost - prev_cost) < self.tolerance:
converged = False
epoch += 1
def fit(self,
data,
labels,
classes,
test_data=None,
test_labels=None,
test_classes=None):
if self.device == 'cpu':
print('Using CPU')
config = tf.ConfigProto(
log_device_placement=True,
allow_soft_placement=True,
#allow_growth=True,
#device_count={'CPU': 0}
)
else:
print('Using GPU')
config = tf.ConfigProto(
log_device_placement=True,
allow_soft_placement=True,
#allow_growth=True,
#device_count={'GPU': 0}
)
if self.session_type == 'default':
self.session = tf.Session(config=config)
if self.session_type == 'interactive':
self.session = tf.InteractiveSession(config=config)
print('Session: ' + str(self.session))
self.session.run(self.init_var)
if self.tensorboard_logs is True:
file_utils.mkdir_p(self.tensorboard_log_dir)
self.merged_summary_op = tf.summary.merge_all()
if self.restore is False:
file_utils.delete_all_files_in_dir(self.tensorboard_log_dir)
if self.separate_writer is False:
self.summary_writer = tf.summary.FileWriter(
self.tensorboard_log_dir, graph=self.session.graph)
else:
self.train_writer = tf.summary.FileWriter(
self.tensorboard_log_dir + 'train',
graph=self.session.graph)
if self.train_validate_split is not None:
self.validate_writer = tf.summary.FileWriter(
self.tensorboard_log_dir + 'validate',
graph=self.session.graph)
if self.test_log is True:
self.test_writer = tf.summary.FileWriter(
self.tensorboard_log_dir + 'test',
graph=self.session.graph)
if self.save_model is True:
self.model = tf.train.Saver(max_to_keep=5)
if self.train_validate_split is not None:
train_data, validate_data, train_labels, validate_labels, train_classes, validate_classes = \
train_test_split(data, labels, classes, train_size=self.train_validate_split)
if self.verbose is True:
print('Data shape: ' + str(data.shape))
print('Labels shape: ' + str(labels.shape))
print('Classes shape: ' + str(classes.shape))
print('Train Data shape: ' + str(train_data.shape))
print('Train Labels shape: ' + str(train_labels.shape))
print('Train Classes shape: ' + str(train_classes.shape))
print('Validate Data shape: ' + str(validate_data.shape))
print('Validate Labels shape: ' + str(validate_labels.shape))
print('Validate Classes shape: ' + str(validate_classes.shape))
if self.test_log is False:
self.optimize(train_data,
train_labels,
train_classes,
validate_data=validate_data,
validate_labels=validate_labels,
validate_classes=validate_classes)
else:
self.optimize(train_data,
train_labels,
train_classes,
validate_data=validate_data,
validate_labels=validate_labels,
validate_classes=validate_classes,
test_data=test_data,
test_labels=test_labels,
test_classes=test_classes)
else:
if self.test_log is False:
self.optimize(data, labels, classes)
else:
self.optimize(data,
labels,
classes,
test_data=test_data,
test_labels=test_labels,
test_classes=test_classes)
def learn(self,
train_data,
train_labels,
train_classes,
validate_data=None,
validate_labels=None,
validate_classes=None,
test_data=None,
test_labels=None,
test_classes=None):
start = time.time()
feed_dict_validate = {
self.predict_params['input']: validate_data,
self.predict_params['true_one_hot']: validate_labels,
self.predict_params['true_class']: validate_classes
}
epoch = self.session.run(self.global_step)
self.out_params['list_train_loss'] = self.session.run(
self.params['var_train_loss']).tolist()
self.out_params['list_train_acc'] = self.session.run(
self.params['var_train_acc']).tolist()
print('Length of train loss : %d' %
len(self.out_params['list_train_loss']))
print('Length of train accuracy : %d' %
len(self.out_params['list_train_acc']))
self.out_params['list_val_loss'] = self.session.run(
self.params['var_val_loss']).tolist()
self.out_params['list_val_acc'] = self.session.run(
self.params['var_val_acc']).tolist()
print('Length of validate loss : %d' %
len(self.out_params['list_val_loss']))
print('Length of validate accuracy : %d' %
len(self.out_params['list_val_acc']))
if self.test_log is True:
self.out_params['list_test_acc'] = \
self.session.run(self.params['var_test_acc']).tolist()
print('Length of test accuracy : %d' %
len(self.out_params['list_test_acc']))
print('Restoring training from epoch :', epoch)
converged = False
prev_cost = 0
while (epoch != self.max_iterations) and converged is False:
start = time.time()
num_batches = int(math.ceil(train_data.shape[0] / self.batch_size))
print('Training using original train data using batch size of %d '
'and total batches of %d' % (self.batch_size, num_batches))
start_batch_index = 0
for batch in range(num_batches):
end_batch_index = start_batch_index + self.batch_size
if end_batch_index < train_data.shape[0]:
train_batch_data = train_data[
start_batch_index:end_batch_index, :]
train_batch_labels = train_labels[
start_batch_index:end_batch_index, :]
train_batch_classes = train_classes[
start_batch_index:end_batch_index]
else:
train_batch_data = train_data[start_batch_index:, :]
train_batch_labels = train_labels[start_batch_index:, :]
train_batch_classes = train_classes[start_batch_index:]
feed_dict_train = {
self.predict_params['input']: train_batch_data,
self.predict_params['true_one_hot']: train_batch_labels,
self.predict_params['true_class']: train_batch_classes
}
_, train_loss, train_loss_summary, \
train_acc, train_acc_summary, curr_epoch \
= self.session.run([self.params['optimizer'],
self.model_params['train_loss'], self.summary_params['train_loss'],
self.model_params['train_acc'], self.summary_params['train_acc'],
self.last_epoch],
feed_dict=feed_dict_train)
start_batch_index += self.batch_size
if self.augmentation is not None:
aug_data, aug_labels, aug_classes = \
self.make_data_augmentation(train_data, train_labels, train_classes)
num_batches = int(
math.ceil(aug_data.shape[0] / self.batch_size))
print(
'Training using augmented train data using batch size of %d '
'and total batches of %d' % (self.batch_size, num_batches))
start_batch_index = 0
for batch in range(num_batches):
end_batch_index = start_batch_index + self.batch_size
if end_batch_index < aug_data.shape[0]:
aug_batch_data = aug_data[
start_batch_index:end_batch_index, :]
aug_batch_labels = aug_labels[
start_batch_index:end_batch_index, :]
aug_batch_classes = aug_classes[
start_batch_index:end_batch_index]
else:
aug_batch_data = aug_data[start_batch_index:, :]
aug_batch_labels = aug_labels[start_batch_index:, :]
aug_batch_classes = aug_classes[start_batch_index:]
feed_dict_train = {
self.predict_params['input']: aug_batch_data,
self.predict_params['true_one_hot']: aug_batch_labels,
self.predict_params['true_class']: aug_batch_classes
}
_, train_loss, train_loss_summary, \
train_acc, train_acc_summary, curr_epoch \
= self.session.run([self.params['optimizer'],
self.model_params['train_loss'], self.summary_params['train_loss'],
self.model_params['train_acc'], self.summary_params['train_acc'],
self.last_epoch],
feed_dict=feed_dict_train)
start_batch_index += self.batch_size
val_loss, val_loss_summary, val_acc, val_acc_summary = \
self.session.run([self.model_params['val_loss'], self.summary_params['val_loss'],
self.model_params['val_acc'], self.summary_params['val_acc']],
feed_dict=feed_dict_validate)
learn_rate_summary = self.session.run(
self.summary_params['learn_rate'])
self.out_params['list_train_loss'].append(train_loss)
self.params['update_train_loss'] = tf.assign(
self.params['var_train_loss'],
self.out_params['list_train_loss'],
validate_shape=False).eval()
self.out_params['list_train_acc'].append(train_acc)
self.params['update_train_acc'] = \
tf.assign(self.params['var_train_acc'], self.out_params['list_train_acc'],
validate_shape=False).eval()
self.out_params['list_learn_rate'].append(self.current_learn_rate)
self.params['update_learn_rate'] = \
tf.assign(self.params['var_learn_rate'], self.out_params['list_learn_rate'],
validate_shape=False).eval()
self.out_params['list_val_loss'].append(val_loss)
self.params['update_val_loss'] = \
tf.assign(self.params['var_val_loss'], self.out_params['list_val_loss'],
validate_shape=False).eval()
self.out_params['list_val_acc'].append(val_acc)
self.params['update_val_acc'] = \
tf.assign(self.params['var_val_acc'], self.out_params['list_val_acc'],
validate_shape=False).eval()
self.summary_writer.add_summary(train_loss_summary, epoch)
self.summary_writer.add_summary(train_acc_summary, epoch)
self.summary_writer.add_summary(val_loss_summary, epoch)
self.summary_writer.add_summary(val_acc_summary, epoch)
self.summary_writer.add_summary(learn_rate_summary, epoch)
if self.test_log is True:
feed_dict_test = {
self.predict_params['input']: test_data,
self.predict_params['true_one_hot']: test_labels,
self.predict_params['true_class']: test_classes
}
test_acc, test_acc_summary = \
self.session.run([self.model_params['test_acc'],
self.summary_params['test_acc']], feed_dict=feed_dict_test)
self.out_params['list_test_acc'].append(test_acc)
self.params['update_test_acc'] = tf.assign(
self.params['var_test_acc'],
self.out_params['list_test_acc'],
validate_shape=False).eval()
self.summary_writer.add_summary(test_acc_summary, epoch)
if epoch % self.display_step == 0:
duration = time.time() - start
if self.test_log is False:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print(
'train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'Time: %.4f s' %
(train_loss, train_acc, val_loss, val_acc, duration))
else:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print(
'train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'test_acc: %.4f | Time: %.4f s' %
(train_loss, train_acc, val_loss, val_acc, test_acc,
duration))
if self.save_checkpoint is True:
model_directory = os.path.dirname(self.checkpoint_filename)
file_utils.mkdir_p(model_directory)
self.model.save(self.session,
self.checkpoint_filename,
global_step=epoch)
if epoch == 0:
prev_cost = train_loss
else:
if math.fabs(train_loss - prev_cost) < self.err_tolerance:
converged = False
epoch += 1
end = time.time()
print('Fit completed in %.4f seconds' % (end - start))
if self.save_model is True:
print('Saving the graph to %s' %
(self.logging_dir + self.model_name.split('/')[-1]))
self.freeze_graph(self.logging_dir)
def learn(self, train_data, train_labels, train_classes,
validate_data=None, validate_labels=None, validate_classes=None,
test_data=None, test_labels=None, test_classes=None):
start = time.time()
feed_dict_train = {self.predict_params['input']: train_data,
self.predict_params['true_one_hot']: train_labels,
self.predict_params['true_class']: train_classes}
feed_dict_validate = {self.predict_params['input']: validate_data,
self.predict_params['true_one_hot']: validate_labels,
self.predict_params['true_class']: validate_classes}
epoch = self.session.run(self.global_step)
self.out_params['list_train_loss'] = self.session.run(self.params['var_train_loss']).tolist()
self.out_params['list_train_acc'] = self.session.run(self.params['var_train_acc']).tolist()
print('Length of train loss : %d' % len(self.out_params['list_train_loss']))
print('Length of train accuracy : %d' % len(self.out_params['list_train_acc']))
self.out_params['list_val_loss'] = self.session.run(self.params['var_val_loss']).tolist()
self.out_params['list_val_acc'] = self.session.run(self.params['var_val_acc']).tolist()
print('Length of validate loss : %d' % len(self.out_params['list_val_loss']))
print('Length of validate accuracy : %d' % len(self.out_params['list_val_acc']))
if self.test_log is True:
self.out_params['list_test_acc'] = \
self.session.run(self.params['var_test_acc']).tolist()
print('Length of test accuracy : %d' % len(self.out_params['list_test_acc']))
print('Restoring training from epoch :', epoch)
converged = False
prev_cost = 0
while (epoch != self.max_iterations) and converged is False:
start = time.time()
_, train_loss, train_loss_summary, \
train_acc, train_acc_summary, curr_epoch \
= self.session.run([self.params['optimizer'],
self.model_params['train_loss'], self.summary_params['train_loss'],
self.model_params['train_acc'], self.summary_params['train_acc'],
self.last_epoch],
feed_dict=feed_dict_train)
val_loss, val_loss_summary, val_acc, val_acc_summary = \
self.session.run([self.model_params['val_loss'], self.summary_params['val_loss'],
self.model_params['val_acc'], self.summary_params['val_acc']],
feed_dict=feed_dict_validate)
learn_rate_summary = self.session.run(self.summary_params['learn_rate'])
self.out_params['list_train_loss'].append(train_loss)
self.params['update_train_loss'] = tf.assign(self.params['var_train_loss'],
self.out_params['list_train_loss'],
validate_shape=False).eval()
self.out_params['list_train_acc'].append(train_acc)
self.params['update_train_acc'] = \
tf.assign(self.params['var_train_acc'], self.out_params['list_train_acc'],
validate_shape=False).eval()
self.out_params['list_learn_rate'].append(self.current_learn_rate)
self.params['update_learn_rate'] = \
tf.assign(self.params['var_learn_rate'], self.out_params['list_learn_rate'],
validate_shape=False).eval()
self.out_params['list_val_loss'].append(val_loss)
self.params['update_val_loss'] = \
tf.assign(self.params['var_val_loss'], self.out_params['list_val_loss'],
validate_shape=False).eval()
self.out_params['list_val_acc'].append(val_acc)
self.params['update_val_acc'] = \
tf.assign(self.params['var_val_acc'], self.out_params['list_val_acc'],
validate_shape=False).eval()
self.summary_writer.add_summary(train_loss_summary, epoch)
self.summary_writer.add_summary(train_acc_summary, epoch)
self.summary_writer.add_summary(val_loss_summary, epoch)
self.summary_writer.add_summary(val_acc_summary, epoch)
self.summary_writer.add_summary(learn_rate_summary, epoch)
if self.test_log is True:
feed_dict_test = {self.predict_params['input']: test_data,
self.predict_params['true_one_hot']: test_labels,
self.predict_params['true_class']: test_classes}
test_acc, test_acc_summary = \
self.session.run([self.model_params['test_acc'],
self.summary_params['test_acc']], feed_dict=feed_dict_test)
self.out_params['list_test_acc'].append(test_acc)
self.params['update_test_acc'] = tf.assign(self.params['var_test_acc'],
self.out_params['list_test_acc'],
validate_shape=False).eval()
self.summary_writer.add_summary(test_acc_summary, epoch)
if epoch % self.display_step == 0:
duration = time.time() - start
if self.test_log is False:
print('>>> Epoch [%*d/%*d]'
% (int(len(str(self.max_iterations))), epoch,
int(len(str(self.max_iterations))), self.max_iterations))
print('train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'Time: %.4f s' % (train_loss, train_acc, val_loss, val_acc, duration))
else:
print('>>> Epoch [%*d/%*d]'
% (int(len(str(self.max_iterations))), epoch,
int(len(str(self.max_iterations))), self.max_iterations))
print('train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'test_acc: %.4f | Time: %.4f s'
% (train_loss, train_acc, val_loss, val_acc, test_acc, duration))
if self.save_checkpoint is True:
model_directory = os.path.dirname(self.checkpoint_filename)
file_utils.mkdir_p(model_directory)
self.model.save(self.session, self.checkpoint_filename, global_step=epoch)
if epoch == 0:
prev_cost = train_loss
else:
if math.fabs(train_loss - prev_cost) < self.err_tolerance:
converged = False
epoch += 1
end = time.time()
print('Fit completed in %.4f seconds' % (end-start))
if self.save_model is True:
print('Saving the graph to %s' % (self.logging_dir+self.model_name.split('/')[-1]))
self.freeze_graph(self.logging_dir)
def run(self,
train_data,
train_labels,
train_classes,
validate_data=None,
validate_labels=None,
validate_classes=None,
test_data=None,
test_labels=None,
test_classes=None):
feed_dict_train = {
self.x: train_data,
self.y_true: train_labels,
self.y_true_cls: train_classes
}
if self.train_validate_split is not None:
feed_dict_validate = {
self.x: validate_data,
self.y_true: validate_labels,
self.y_true_cls: validate_classes
}
if self.test_log is True:
feed_dict_test = {
self.x: test_data,
self.y_true: test_labels,
self.y_true_cls: test_classes
}
epoch = self.session.run(self.global_step)
self.list_train_loss = self.session.run(self.var_train_loss)
self.list_train_loss = self.list_train_loss.tolist()
self.list_train_acc = self.session.run(self.var_train_acc)
self.list_train_acc = self.list_train_acc.tolist()
print('Length of train loss : %d' % len(self.list_train_loss))
print('Length of train accuracy : %d' % len(self.list_train_acc))
if self.train_validate_split is not None:
self.list_validate_loss = self.session.run(self.var_validate_loss)
self.list_validate_loss = self.list_validate_loss.tolist()
self.list_validate_acc = self.session.run(self.var_validate_acc)
self.list_validate_acc = self.list_validate_acc.tolist()
print('Length of validate loss : %d' %
len(self.list_validate_loss))
print('Length of validate accuracy : %d' %
len(self.list_validate_acc))
if self.test_log is True:
self.list_test_acc = self.session.run(self.test_var_acc)
self.list_test_acc = self.list_test_acc.tolist()
print('Length of test accuracy : %d' % len(self.list_test_acc))
print('Restoring training from epoch :', epoch)
converged = False
prev_cost = 0
num_batches = int(train_data.shape[0] / self.batch_size)
while (epoch != self.max_iterations) and converged is False:
start = time.time()
start_batch_index = 0
for batch in range(num_batches):
# print('Training on batch %d' %batch)
end_batch_index = start_batch_index + self.batch_size
if end_batch_index < train_data.shape[0]:
train_batch_data = train_data[
start_batch_index:end_batch_index, :]
train_batch_labels = train_labels[
start_batch_index:end_batch_index, :]
train_batch_classes = train_classes[
start_batch_index:end_batch_index]
else:
train_batch_data = train_data[start_batch_index:, :]
train_batch_labels = train_labels[start_batch_index:, :]
train_batch_classes = train_classes[start_batch_index:]
feed_dict_train = {
self.x: train_batch_data,
self.y_true: train_batch_labels,
self.y_true_cls: train_batch_classes
}
_, train_loss, train_acc \
= self.session.run([self.optimizer, self.train_loss, self.train_accuracy], feed_dict=feed_dict_train)
train_loss_summary = self.session.run(
self.train_loss_summary, feed_dict=feed_dict_train)
train_acc_summary = self.session.run(self.train_acc_summary,
feed_dict=feed_dict_train)
start_batch_index += self.batch_size
curr_epoch = self.session.run(self.last_epoch)
validate_loss_summary = self.session.run(
self.validate_loss_summary, feed_dict=feed_dict_validate)
learning_rate_summary = self.session.run(
self.learning_rate_summary)
self.list_train_loss.append(train_loss)
self.update_train_loss = tf.assign(self.var_train_loss,
self.list_train_loss,
validate_shape=False)
self.update_train_loss.eval()
self.list_train_acc.append(train_acc)
self.update_train_acc = tf.assign(self.var_train_acc,
self.list_train_acc,
validate_shape=False)
self.update_train_acc.eval()
self.list_learning_rate.append(self.current_learning_rate)
self.update_learning_rate = tf.assign(self.var_learning_rate,
self.list_learning_rate,
validate_shape=False)
self.update_learning_rate.eval()
# w_hist = self.session.run(self.w_hist, feed_dict=feed_dict_train)
# self.summary_writer.add_summary(w_hist, epoch)
# w_im = self.session.run(self.w_im, feed_dict=feed_dict_train)
# self.summary_writer.add_summary(w_im, epoch)
if self.train_validate_split is not None:
validate_loss, validate_acc, validate_acc_summary = \
self.session.run([self.validate_loss, self.validate_accuracy, self.validate_acc_summary],
feed_dict=feed_dict_validate)
self.list_validate_loss.append(validate_loss)
self.update_validate_loss = tf.assign(self.var_validate_loss,
self.list_validate_loss,
validate_shape=False)
self.update_validate_loss.eval()
self.list_validate_acc.append(validate_acc)
self.update_validate_acc = tf.assign(self.var_validate_acc,
self.list_validate_acc,
validate_shape=False)
self.update_validate_acc.eval()
if self.test_log is True:
test_acc, test_acc_summary = \
self.session.run([self.test_accuracy, self.test_acc_summary], feed_dict=feed_dict_test)
self.list_test_acc.append(test_acc)
self.update_test_acc = tf.assign(self.test_var_acc,
self.list_test_acc,
validate_shape=False)
self.update_test_acc.eval()
if self.separate_writer is False:
self.summary_writer.add_summary(train_loss_summary, epoch)
self.summary_writer.add_summary(train_acc_summary, epoch)
self.summary_writer.add_summary(validate_loss_summary, epoch)
self.summary_writer.add_summary(validate_acc_summary, epoch)
self.summary_writer.add_summary(test_acc_summary, epoch)
self.summary_writer.add_summary(learning_rate_summary, epoch)
else:
self.train_writer.add_summary(train_loss_summary, epoch)
self.train_writer.add_summary(train_acc_summary, epoch)
if self.train_validate_split is not None:
self.validate_writer.add_summary(validate_loss_summary,
epoch)
self.validate_writer.add_summary(validate_acc_summary,
epoch)
if self.test_log is True:
self.test_writer.add_summary(test_acc_summary, epoch)
if epoch % self.display_step == 0:
duration = time.time() - start
if self.train_validate_split is not None and self.test_log is False:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print(
'train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'Time: %.4f s' % (train_loss, train_acc, validate_loss,
validate_acc, duration))
elif self.train_validate_split is not None and self.test_log is True:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print(
'train_loss: %.4f | train_acc: %.4f | val_loss: %.4f | val_acc: %.4f | '
'test_acc: %.4f | Time: %.4f s' %
(train_loss, train_acc, validate_loss, validate_acc,
test_acc, duration))
elif self.train_validate_split is None and self.test_log is True:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print(
'train_loss: %.4f | train_acc: %.4f | test_acc: %.4f | Time: %.4f s'
% (train_loss, train_acc, test_acc, duration))
else:
print('>>> Epoch [%*d/%*d]' %
(int(len(str(self.max_iterations))), epoch,
int(len(str(
self.max_iterations))), self.max_iterations))
print('train_loss: %.4f | train_acc: %.4f | Time: %.4f s' %
(train_loss, train_acc, duration))
if self.save_model is True:
model_directory = os.path.dirname(self.model_name)
file_utils.mkdir_p(model_directory)
self.model.save(self.session,
self.model_name,
global_step=epoch)
if epoch == 0:
prev_cost = train_loss
else:
if math.fabs(train_loss - prev_cost) < self.tolerance:
converged = False
epoch += 1
