def set_key_vars(self, restore_scope_exclude, train_scopes):
"""Set critical variables for relevant tasks.
Set vars_to_train and vars_to_restore.
Called after build_model.
Args:
restore_scope_exclude: variable scopes to exclude for restoring.
train_scopes: variable scopes to train.
"""
# self.dm_model.use_graph()
self.vars_to_restore = slim.get_variables_to_restore(
exclude=restore_scope_exclude)
self.vars_to_train = []
if train_scopes is not None:
for scope in train_scopes:
variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)
self.vars_to_train.extend(variables)
if not self.vars_to_train:
print "[set_key_vars: info] No variables to train were defined." \
" Will train ALL variables."
self.vars_to_train = None
#base_model.print_variable_names(self.vars_to_train)
python类get_variables_to_restore()的实例源码
def test_model_restore(self):
model_path = model_common.get_builtin_net_weights_fn(
commons.ModelType.INCEPTION_V4)
reader = pywrap_tensorflow.NewCheckpointReader(model_path)
inputs = tf.placeholder(tf.float32, shape=(None, 299, 299, 3))
model_common.create_builtin_net(commons.ModelType.INCEPTION_V3, inputs, 78)
vars_to_restore = slim.get_variables_to_restore(
exclude=["InceptionV3/Logits"])
if isinstance(vars_to_restore, (tuple, list)):
vars_to_restore = {var.op.name: var for var in vars_to_restore}
for checkpoint_var_name in vars_to_restore:
var = vars_to_restore[checkpoint_var_name]
if not reader.has_tensor(checkpoint_var_name):
raise ValueError('Checkpoint is missing variable [%s]' %
checkpoint_var_name)
var_value = reader.get_tensor(checkpoint_var_name)
print "tensor {} has shape {}, and its value has shape {}".format(
checkpoint_var_name, var.get_shape(), var_value.shape)
new_value = var_value.reshape(var.get_shape())
def train(self):
s = tf.Session()
init_fn = slim.assign_from_checkpoint_fn("./vgg_19.ckpt", slim.get_variables_to_restore(exclude = ['generate_image']))
#optimizer = tf.train.AdamOptimizer(learning_rate = 1e-1, beta1 = 0.5, beta2 = 0.5).minimize(self.loss, var_list = [self.target])
optimizer = tf.contrib.opt.ScipyOptimizerInterface(self.loss, options={'maxiter': 1000}, var_list = [self.target])
s.run(tf.global_variables_initializer())
init_fn(s)
#for i in range(10000):
# _, loss_out = s.run([optimizer, self.loss])
# print("Current loss is: %.3f" %loss_out, end="\r")
#print("")
optimizer.minimize(s)
loss_out = s.run(self.loss)
print("Final loss: %.3f" %loss_out)
plt.imshow(np.clip(s.run(self.target), 0, 255).astype(np.uint8))
plt.show()
def _create_image_encoder(preprocess_fn, factory_fn, image_shape, batch_size=32,
session=None, checkpoint_path=None,
loss_mode="cosine"):
image_var = tf.placeholder(tf.uint8, (None, ) + image_shape)
preprocessed_image_var = tf.map_fn(
lambda x: preprocess_fn(x, is_training=False),
tf.cast(image_var, tf.float32))
l2_normalize = loss_mode == "cosine"
feature_var, _ = factory_fn(
preprocessed_image_var, l2_normalize=l2_normalize, reuse=None)
feature_dim = feature_var.get_shape().as_list()[-1]
if session is None:
session = tf.Session()
if checkpoint_path is not None:
slim.get_or_create_global_step()
init_assign_op, init_feed_dict = slim.assign_from_checkpoint(
checkpoint_path, slim.get_variables_to_restore())
session.run(init_assign_op, feed_dict=init_feed_dict)
def encoder(data_x):
out = np.zeros((len(data_x), feature_dim), np.float32)
_run_in_batches(
lambda x: session.run(feature_var, feed_dict=x),
{image_var: data_x}, out, batch_size)
return out
return encoder
def __init__(self, content, style, content_names, style_names):
"""
Suppose the content and style is a numpy array,
"""
self.content_names = content_names
self.style_names = style_names
self.VGG_MEAN = [123.68, 116.78, 103.94]
tf.reset_default_graph()
content = tf.constant(content) - tf.reshape(tf.constant(self.VGG_MEAN), [1, 1, 3])
_, self.content_layers = nets.vgg.vgg_19(tf.expand_dims(content, axis = 0), is_training = False, spatial_squeeze = False)
layer_name, layer_value = zip(*filter(lambda x: x[0] in content_names, self.content_layers.items()))
init_fn = slim.assign_from_checkpoint_fn("./vgg_19.ckpt", slim.get_variables_to_restore())
with tf.Session() as s, tf.device("/device:XLA_CPU:0"):
init_fn(s)
layer_value = s.run(layer_value)
self.content_map = dict(zip(layer_name, layer_value))
#print(content_map)
tf.reset_default_graph()
style = tf.constant(style) - tf.reshape(tf.constant(self.VGG_MEAN), [1, 1, 3])
_, self.style_layers = nets.vgg.vgg_19(tf.expand_dims(style, axis = 0), is_training = False, spatial_squeeze = False)
layer_name, layer_value = zip(*filter(lambda x: x[0] in style_names, self.style_layers.items()))
init_fn = slim.assign_from_checkpoint_fn("./vgg_19.ckpt", slim.get_variables_to_restore())
with tf.Session() as s, tf.device("/device:XLA_CPU:0"):
init_fn(s)
layer_value = s.run(layer_value)
self.style_map = dict(zip(layer_name, layer_value))
#print(content_map)
tf.reset_default_graph()
self.target = tf.Variable(np.random.randint(0, 256, content.shape), dtype = tf.float32, name = "generate_image")
self._build_graph()
def use_fined_model(self):
image_size = inception.inception_v4.default_image_size
batch_size = 3
flowers_data_dir = "../../data/flower"
train_dir = '/tmp/inception_finetuned/'
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO)
dataset = flowers.get_split('train', flowers_data_dir)
images, images_raw, labels = self.load_batch(dataset, height=image_size, width=image_size)
# Create the model, use the default arg scope to configure the batch norm parameters.
with slim.arg_scope(inception.inception_v4_arg_scope()):
logits, _ = inception.inception_v4(images, num_classes=dataset.num_classes, is_training=True)
probabilities = tf.nn.softmax(logits)
checkpoint_path = tf.train.latest_checkpoint(train_dir)
init_fn = slim.assign_from_checkpoint_fn(
checkpoint_path,
slim.get_variables_to_restore())
with tf.Session() as sess:
with slim.queues.QueueRunners(sess):
sess.run(tf.initialize_local_variables())
init_fn(sess)
np_probabilities, np_images_raw, np_labels = sess.run([probabilities, images_raw, labels])
for i in range(batch_size):
image = np_images_raw[i, :, :, :]
true_label = np_labels[i]
predicted_label = np.argmax(np_probabilities[i, :])
predicted_name = dataset.labels_to_names[predicted_label]
true_name = dataset.labels_to_names[true_label]
plt.figure()
plt.imshow(image.astype(np.uint8))
plt.title('Ground Truth: [%s], Prediction [%s]' % (true_name, predicted_name))
plt.axis('off')
plt.show()
return
def main():
model = config.get('config', 'model')
cachedir = utils.get_cachedir(config)
with open(os.path.join(cachedir, 'names'), 'r') as f:
names = [line.strip() for line in f]
width = config.getint(model, 'width')
height = config.getint(model, 'height')
yolo = importlib.import_module('model.' + model)
cell_width, cell_height = utils.calc_cell_width_height(config, width, height)
tf.logging.info('(width, height)=(%d, %d), (cell_width, cell_height)=(%d, %d)' % (width, height, cell_width, cell_height))
with tf.Session() as sess:
paths = [os.path.join(cachedir, profile + '.tfrecord') for profile in args.profile]
num_examples = sum(sum(1 for _ in tf.python_io.tf_record_iterator(path)) for path in paths)
tf.logging.warn('num_examples=%d' % num_examples)
image_rgb, labels = utils.data.load_image_labels(paths, len(names), width, height, cell_width, cell_height, config)
image_std = tf.image.per_image_standardization(image_rgb)
image_rgb = tf.cast(image_rgb, tf.uint8)
ph_image = tf.placeholder(image_std.dtype, [1] + image_std.get_shape().as_list(), name='ph_image')
global_step = tf.contrib.framework.get_or_create_global_step()
builder = yolo.Builder(args, config)
builder(ph_image)
variables_to_restore = slim.get_variables_to_restore()
ph_labels = [tf.placeholder(l.dtype, [1] + l.get_shape().as_list(), name='ph_' + l.op.name) for l in labels]
with tf.name_scope('total_loss') as name:
builder.create_objectives(ph_labels)
total_loss = tf.losses.get_total_loss(name=name)
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess, coord)
_image_rgb, _image_std, _labels = sess.run([image_rgb, image_std, labels])
coord.request_stop()
coord.join(threads)
feed_dict = dict([(ph, np.expand_dims(d, 0)) for ph, d in zip(ph_labels, _labels)])
feed_dict[ph_image] = np.expand_dims(_image_std, 0)
logdir = utils.get_logdir(config)
assert os.path.exists(logdir)
model_path = tf.train.latest_checkpoint(logdir)
tf.logging.info('load ' + model_path)
slim.assign_from_checkpoint_fn(model_path, variables_to_restore)(sess)
tf.logging.info('global_step=%d' % sess.run(global_step))
tf.logging.info('total_loss=%f' % sess.run(total_loss, feed_dict))
_ = Drawer(sess, names, builder.model.cell_width, builder.model.cell_height, _image_rgb, _labels, builder.model, feed_dict)
plt.show()