def build_prediction_graph(self, serialized_examples):
# video_id, model_input_raw, labels_batch, num_frames = (
### Newly
video_id, model_input_raw, labels_batch, coarse_labels_batch, num_frames = (
###
self.reader.prepare_serialized_examples(serialized_examples))
feature_dim = len(model_input_raw.get_shape()) - 1
model_input = tf.nn.l2_normalize(model_input_raw, feature_dim)
with tf.variable_scope("tower"):
result = self.model.create_model(
model_input,
num_frames=num_frames,
vocab_size=self.reader.num_classes,
labels=labels_batch,
### Newly
coarse_labels=coarse_labels_batch,
###
is_training=False)
for variable in slim.get_model_variables():
tf.summary.histogram(variable.op.name, variable)
predictions = result["predictions"]
top_predictions, top_indices = tf.nn.top_k(predictions,
_TOP_PREDICTIONS_IN_OUTPUT)
return video_id, top_indices, top_predictions
python类get_model_variables()的实例源码
def forward(self, input):
self.depth = input
conv = self._conv
upconv = self._upconv
maxpool = self._maxpool
with tf.variable_scope('feature'):
with tf.variable_scope('encoder'):
conv1 = conv(self.depth, 96, 11, 4) # H/4
pool1 = maxpool(conv1, 3) # H/8
conv2 = conv(pool1, 256, 5, 1) # H/8
pool2 = maxpool(conv2, 3) # H/16
conv3 = conv(pool2, 384, 3, 1) # H/16
conv4 = conv(conv3, 384, 3, 1) # H/16
conv5 = conv(conv4, 256, 3, 1) # H/16
pool5 = maxpool(conv5, 3) # H/32
conv6 = conv(pool5, 4096, 1, 1) # H/32
conv7 = conv(conv6, 4096, 1, 1) # H/32
with tf.variable_scope('skips'):
skip1 = conv1
skip2 = conv2
skip3 = conv5
with tf.variable_scope('decoder'):
upconv5 = upconv(conv7, 256, 3, 2) #H/16
concat5 = tf.concat([upconv5, skip3], 3)
iconv5 = conv(concat5, 256, 3, 1)
upconv4 = upconv(iconv5, 256, 3, 2) #H/8
concat4 = tf.concat([upconv4, skip2], 3)
iconv4 = conv(concat4, 256, 3, 1)
upconv3 = upconv(iconv4, 96, 3, 2) #H/4
concat3 = tf.concat([upconv3, skip1], 3)
iconv3 = conv(concat3, 96, 3, 1)
upconv2 = upconv(iconv3, 48, 3, 2) #H/2
upconv1 = upconv(upconv2, 16, 3, 2) #H/1
self.feature = upconv1
self.feat_vars = slim.get_model_variables()
return self.feature
def train(self):
img_size = [self.image_height, self.image_width, self.image_depth]
train_batch = tf.train.shuffle_batch([read_tfrecord(self.train_file, img_size)],
batch_size = self.train_batch_size,
capacity = 3000,
num_threads = 2,
min_after_dequeue = 1000)
test_batch = tf.train.shuffle_batch([read_tfrecord(self.test_file, img_size)],
batch_size = self.test_batch_size,
capacity = 500,
num_threads = 2,
min_after_dequeue = 300)
init = tf.global_variables_initializer()
init_fn = slim.assign_from_checkpoint_fn("resnet_v2_50.ckpt", slim.get_model_variables('resnet_v2'))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
init_fn(sess)
train_writer = tf.summary.FileWriter(self.log_dir + "/train", sess.graph)
test_writer = tf.summary.FileWriter(self.log_dir + "/test", sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
inputs_test, outputs_gt_test = build_img_pair(sess.run(test_batch))
for iter in range(self.max_iteration):
inputs_train, outputs_gt_train = build_img_pair(sess.run(train_batch))
# train with dynamic learning rate
if iter <= 500:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-3, self.batch_size:self.train_batch_size})
elif iter <= self.max_iteration - 1000:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:0.5e-3, self.batch_size:self.train_batch_size})
else:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-4, self.batch_size:self.train_batch_size})
# print training loss and test loss
if iter%10 == 0:
summary_train = sess.run(self.summary, {self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
train_writer.add_summary(summary_train, iter)
train_writer.flush()
summary_test = sess.run(self.summary, {self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
test_writer.add_summary(summary_test, iter)
test_writer.flush()
# record training loss and test loss
if iter%10 == 0:
train_loss = self.cross_entropy.eval({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
test_loss = self.cross_entropy.eval({self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
print("iter step %d trainning batch loss %f"%(iter, train_loss))
print("iter step %d test loss %f\n"%(iter, test_loss))
# record model
if iter%100 == 0:
saver.save(sess, self.log_dir + "/model.ckpt", global_step=iter)
coord.request_stop()
coord.join(threads)
def train(self):
img_size = [self.image_height, self.image_width, self.image_depth]
train_batch = tf.train.shuffle_batch([read_tfrecord(self.train_file, img_size)],
batch_size = self.train_batch_size,
capacity = 3000,
num_threads = 2,
min_after_dequeue = 1000)
test_batch = tf.train.shuffle_batch([read_tfrecord(self.test_file, img_size)],
batch_size = self.test_batch_size,
capacity = 500,
num_threads = 2,
min_after_dequeue = 300)
init = tf.global_variables_initializer()
init_fn = slim.assign_from_checkpoint_fn("resnet_v2_50.ckpt", slim.get_model_variables('resnet_v2'))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
init_fn(sess)
train_writer = tf.summary.FileWriter(self.log_dir + "/train", sess.graph)
test_writer = tf.summary.FileWriter(self.log_dir + "/test", sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
inputs_test, outputs_gt_test = build_img_pair(sess.run(test_batch))
for iter in range(self.max_iteration):
inputs_train, outputs_gt_train = build_img_pair(sess.run(train_batch))
# train with dynamic learning rate
if iter <= 500:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-3, self.batch_size:self.train_batch_size})
elif iter <= self.max_iteration - 1000:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:0.5e-3, self.batch_size:self.train_batch_size})
else:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-4, self.batch_size:self.train_batch_size})
# print training loss and test loss
if iter%10 == 0:
summary_train = sess.run(self.summary, {self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
train_writer.add_summary(summary_train, iter)
train_writer.flush()
summary_test = sess.run(self.summary, {self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
test_writer.add_summary(summary_test, iter)
test_writer.flush()
# record training loss and test loss
if iter%10 == 0:
train_loss = self.cross_entropy.eval({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
test_loss = self.cross_entropy.eval({self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
print("iter step %d trainning batch loss %f"%(iter, train_loss))
print("iter step %d test loss %f\n"%(iter, test_loss))
# record model
if iter%100 == 0:
saver.save(sess, self.log_dir + "/model.ckpt", global_step=iter)
coord.request_stop()
coord.join(threads)
def train(self):
img_size = [self.image_height, self.image_width, self.image_depth]
train_batch = tf.train.shuffle_batch([read_tfrecord(self.train_file, img_size)],
batch_size = self.train_batch_size,
capacity = 2000,
num_threads = 2,
min_after_dequeue = 1000)
test_batch = tf.train.shuffle_batch([read_tfrecord(self.test_file, img_size)],
batch_size = self.test_batch_size,
capacity = 500,
num_threads = 2,
min_after_dequeue = 300)
init = tf.global_variables_initializer()
init_fn = slim.assign_from_checkpoint_fn("vgg_16.ckpt", slim.get_model_variables('vgg_16'))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
init_fn(sess)
train_writer = tf.summary.FileWriter(self.log_dir + "/train", sess.graph)
test_writer = tf.summary.FileWriter(self.log_dir + "/test", sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
inputs_test, outputs_gt_test = build_img_pair(sess.run(test_batch))
for iter in range(self.max_iteration):
inputs_train, outputs_gt_train = build_img_pair(sess.run(train_batch))
# train with dynamic learning rate
if iter <= 500:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train, self.is_training:True,
self.learning_rate:1e-4})
elif iter <= self.max_iteration - 1000:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train, self.is_training:True,
self.learning_rate:0.5e-4})
else:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train, self.is_training:True,
self.learning_rate:1e-5})
# print training loss and test loss
if iter%10 == 0:
summary_train = sess.run(self.summary, {self.input_data:inputs_train, self.gt:outputs_gt_train,
self.is_training:False})
train_writer.add_summary(summary_train, iter)
train_writer.flush()
summary_test = sess.run(self.summary, {self.input_data:inputs_test, self.gt:outputs_gt_test,
self.is_training:False})
test_writer.add_summary(summary_test, iter)
test_writer.flush()
# record training loss and test loss
if iter%10 == 0:
train_loss = self.cross_entropy.eval({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.is_training:False})
test_loss = self.cross_entropy.eval({self.input_data:inputs_test, self.gt:outputs_gt_test,
self.is_training:False})
print("iter step %d trainning batch loss %f"%(iter, train_loss))
print("iter step %d test loss %f\n"%(iter, test_loss))
# record model
if iter%100 == 0:
saver.save(sess, self.log_dir + "/model.ckpt", global_step=iter)
coord.request_stop()
coord.join(threads)
def train(self):
img_size = [self.image_height, self.image_width, self.image_depth]
train_batch = tf.train.shuffle_batch([read_tfrecord(self.train_file, img_size)],
batch_size = self.train_batch_size,
capacity = 3000,
num_threads = 2,
min_after_dequeue = 1000)
test_batch = tf.train.shuffle_batch([read_tfrecord(self.test_file, img_size)],
batch_size = self.test_batch_size,
capacity = 500,
num_threads = 2,
min_after_dequeue = 300)
init = tf.global_variables_initializer()
init_fn = slim.assign_from_checkpoint_fn("resnet_v2_50.ckpt", slim.get_model_variables('resnet_v2'))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
init_fn(sess)
train_writer = tf.summary.FileWriter(self.log_dir + "/train", sess.graph)
test_writer = tf.summary.FileWriter(self.log_dir + "/test", sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
inputs_test, outputs_gt_test = build_img_pair(sess.run(test_batch))
for iter in range(self.max_iteration):
inputs_train, outputs_gt_train = build_img_pair(sess.run(train_batch))
# train with dynamic learning rate
if iter <= 500:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-3, self.batch_size:self.train_batch_size})
elif iter <= self.max_iteration - 1000:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:0.5e-3, self.batch_size:self.train_batch_size})
else:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-4, self.batch_size:self.train_batch_size})
# print training loss and test loss
if iter%10 == 0:
summary_train = sess.run(self.summary, {self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
train_writer.add_summary(summary_train, iter)
train_writer.flush()
summary_test = sess.run(self.summary, {self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
test_writer.add_summary(summary_test, iter)
test_writer.flush()
# record training loss and test loss
if iter%10 == 0:
train_loss = self.cross_entropy.eval({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.batch_size:self.train_batch_size})
test_loss = self.cross_entropy.eval({self.input_data:inputs_test, self.gt:outputs_gt_test,
self.batch_size:self.test_batch_size})
print("iter step %d trainning batch loss %f"%(iter, train_loss))
print("iter step %d test loss %f\n"%(iter, test_loss))
# record model
if iter%100 == 0:
saver.save(sess, self.log_dir + "/model.ckpt", global_step=iter)
coord.request_stop()
coord.join(threads)
def train(self):
img_size = [self.image_height, self.image_width, self.image_depth]
train_batch = tf.train.shuffle_batch([read_tfrecord(self.train_file, img_size)],
batch_size = self.train_batch_size,
capacity = 2000,
num_threads = 2,
min_after_dequeue = 1000)
test_batch = tf.train.shuffle_batch([read_tfrecord(self.test_file, img_size)],
batch_size = self.test_batch_size,
capacity = 500,
num_threads = 2,
min_after_dequeue = 300)
init = tf.global_variables_initializer()
init_fn = slim.assign_from_checkpoint_fn("vgg_16.ckpt", slim.get_model_variables('vgg_16'))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
init_fn(sess)
train_writer = tf.summary.FileWriter(self.log_dir + "/train", sess.graph)
test_writer = tf.summary.FileWriter(self.log_dir + "/test", sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
inputs_test, outputs_gt_test = build_img_pair(sess.run(test_batch))
for iter in range(self.max_iteration):
inputs_train, outputs_gt_train = build_img_pair(sess.run(train_batch))
# train with dynamic learning rate
if iter <= 500:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-4, self.is_training:True})
elif iter <= self.max_iteration - 1000:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:0.5e-4, self.is_training:True})
else:
self.train_step.run({self.input_data:inputs_train, self.gt:outputs_gt_train,
self.learning_rate:1e-5, self.is_training:True})
# print training loss and test loss
if iter%10 == 0:
summary_train = sess.run(self.summary, {self.input_data:inputs_train, self.gt:outputs_gt_train, self.is_training:False})
train_writer.add_summary(summary_train, iter)
train_writer.flush()
summary_test = sess.run(self.summary, {self.input_data:inputs_test, self.gt:outputs_gt_test, self.is_training:False})
test_writer.add_summary(summary_test, iter)
test_writer.flush()
# record training loss and test loss
if iter%10 == 0:
train_loss = self.cross_entropy.eval({self.input_data:inputs_train, self.gt:outputs_gt_train, self.is_training:False})
test_loss = self.cross_entropy.eval({self.input_data:inputs_test, self.gt:outputs_gt_test, self.is_training:False})
print("iter step %d trainning batch loss %f"%(iter, train_loss))
print("iter step %d test loss %f\n"%(iter, test_loss))
# record model
if iter%100 == 0:
saver.save(sess, self.log_dir + "/model.ckpt", global_step=iter)
coord.request_stop()
coord.join(threads)
def __get_init_fn(self):
"""Returns a function run by the chief worker to warm-start the training.
Note that the init_fn is only run when initializing the model during the very
first global step.
Returns:
An init function run by the supervisor.
"""
if self.checkpoint_path is None:
return None
# Warn the user if a checkpoint exists in the train_dir. Then we'll be
# ignoring the checkpoint anyway.
if tf.train.latest_checkpoint(self.train_dir):
tf.logging.info(
'Ignoring --checkpoint_path because a checkpoint already exists in %s'
% self.train_dir)
return None
exclusions = []
if self.checkpoint_exclude_scopes:
exclusions = [scope.strip()
for scope in self.checkpoint_exclude_scopes.split(',')]
# TODO(sguada) variables.filter_variables()
variables_to_restore = []
for var in slim.get_model_variables():
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
break
if not excluded:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(self.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(self.checkpoint_path)
else:
checkpoint_path = self.checkpoint_path
tf.logging.info('Fine-tuning from %s' % checkpoint_path)
return slim.assign_from_checkpoint_fn(
checkpoint_path,
variables_to_restore,
ignore_missing_vars=self.ignore_missing_vars)
def _get_init_fn():
"""Returns a function run by the chief worker to warm-start the training.
Note that the init_fn is only run when initializing the model during the very
first global step.
Returns:
An init function run by the supervisor.
"""
if FLAGS.checkpoint_path is None:
return None
# Warn the user if a checkpoint exists in the train_dir. Then we'll be
# ignoring the checkpoint anyway.
if tf.train.latest_checkpoint(FLAGS.train_dir):
tf.logging.info(
'Ignoring --checkpoint_path because a checkpoint already exists in %s'
% FLAGS.train_dir)
return None
exclusions = []
if FLAGS.checkpoint_exclude_scopes:
exclusions = [scope.strip()
for scope in FLAGS.checkpoint_exclude_scopes.split(',')]
# TODO(sguada) variables.filter_variables()
variables_to_restore = []
for var in slim.get_model_variables():
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
break
if not excluded:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(FLAGS.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
else:
checkpoint_path = FLAGS.checkpoint_path
tf.logging.info('Fine-tuning from %s' % checkpoint_path)
return slim.assign_from_checkpoint_fn(
checkpoint_path,
variables_to_restore,
ignore_missing_vars=FLAGS.ignore_missing_vars)
def __get_init_fn(self):
"""Returns a function run by the chief worker to warm-start the training.
Note that the init_fn is only run when initializing the model during the very
first global step.
Returns:
An init function run by the supervisor.
"""
if self.checkpoint_path is None:
return None
# Warn the user if a checkpoint exists in the train_dir. Then we'll be
# ignoring the checkpoint anyway.
if tf.train.latest_checkpoint(self.train_dir):
tf.logging.info(
'Ignoring --checkpoint_path because a checkpoint already exists in %s'
% self.train_dir)
return None
exclusions = []
if self.checkpoint_exclude_scopes:
exclusions = [scope.strip()
for scope in self.checkpoint_exclude_scopes.split(',')]
# TODO(sguada) variables.filter_variables()
variables_to_restore = []
all_variables = slim.get_model_variables()
if self.fine_tune_vgg16:
global_step = slim.get_or_create_global_step()
all_variables.append(global_step)
for var in all_variables:
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
break
if not excluded:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(self.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(self.checkpoint_path)
else:
checkpoint_path = self.checkpoint_path
tf.logging.info('Fine-tuning from %s' % checkpoint_path)
return slim.assign_from_checkpoint_fn(
checkpoint_path,
variables_to_restore,
ignore_missing_vars=self.ignore_missing_vars)
def _get_init_fn():
"""Returns a function run by the chief worker to warm-start the training.
Note that the init_fn is only run when initializing the model during the very
first global step.
Returns:
An init function run by the supervisor.
"""
if FLAGS.checkpoint_path is None:
return None
# Warn the user if a checkpoint exists in the train_dir. Then we'll be
# ignoring the checkpoint anyway.
if tf.train.latest_checkpoint(FLAGS.train_dir):
tf.logging.info(
'Ignoring --checkpoint_path because a checkpoint already exists in %s'
% FLAGS.train_dir)
return None
exclusions = []
if FLAGS.checkpoint_exclude_scopes:
exclusions = [scope.strip()
for scope in FLAGS.checkpoint_exclude_scopes.split(',')]
# TODO(sguada) variables.filter_variables()
variables_to_restore = []
for var in slim.get_model_variables():
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
break
if not excluded:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(FLAGS.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
else:
checkpoint_path = FLAGS.checkpoint_path
tf.logging.info('Fine-tuning from %s' % checkpoint_path)
return slim.assign_from_checkpoint_fn(
checkpoint_path,
variables_to_restore,
ignore_missing_vars=FLAGS.ignore_missing_vars)
def export_to_h5(checkpoint_dir, export_path, images, end_points, num_samples,
batch_size, sact):
"""Exports ponder cost maps and other useful info to an HDF5 file."""
output_file = h5py.File(export_path, 'w')
output_file.attrs['block_scopes'] = end_points['block_scopes']
keys_to_tensors = {}
for block_scope in end_points['block_scopes']:
for k in ('{}/ponder_cost'.format(block_scope),
'{}/num_units'.format(block_scope),
'{}/halting_distribution'.format(block_scope),
'{}/flops'.format(block_scope)):
keys_to_tensors[k] = end_points[k]
keys_to_tensors['images'] = images
keys_to_tensors['flops'] = end_points['flops']
if sact:
keys_to_tensors['ponder_cost_map'] = sact_map(end_points, 'ponder_cost')
keys_to_tensors['num_units_map'] = sact_map(end_points, 'num_units')
keys_to_datasets = {}
for key, tensor in keys_to_tensors.iteritems():
sh = tensor.get_shape().as_list()
sh[0] = num_samples
print(key, sh)
keys_to_datasets[key] = output_file.create_dataset(
key, sh, compression='lzf')
variables_to_restore = slim.get_model_variables()
checkpoint_path = tf.train.latest_checkpoint(checkpoint_dir)
assert checkpoint_path is not None
init_fn = slim.assign_from_checkpoint_fn(checkpoint_path,
variables_to_restore)
sv = tf.train.Supervisor(
graph=tf.get_default_graph(),
logdir=None,
summary_op=None,
summary_writer=None,
global_step=None,
saver=None)
assert num_samples % batch_size == 0
num_batches = num_samples // batch_size
with sv.managed_session('', start_standard_services=False) as sess:
init_fn(sess)
sv.start_queue_runners(sess)
for i in range(num_batches):
tf.logging.info('Evaluating batch %d/%d', i + 1, num_batches)
end_points_out = sess.run(keys_to_tensors)
for key, dataset in keys_to_datasets.iteritems():
dataset[i * batch_size:(i + 1) * batch_size, ...] = end_points_out[key]
