def tfslim_vgg16():
import tensorflow as tf
from nets import nets_factory
from preprocessing import vgg_preprocessing
from model_tools.activations.tensorflow import load_resize_image
tf.reset_default_graph()
image_size = 224
placeholder = tf.placeholder(dtype=tf.string, shape=[64])
preprocess_image = lambda image: vgg_preprocessing.preprocess_image(
image, image_size, image_size, resize_side_min=image_size)
preprocess = lambda image_path: preprocess_image(
load_resize_image(image_path, image_size))
preprocess = tf.map_fn(preprocess, placeholder, dtype=tf.float32)
model_ctr = nets_factory.get_network_fn('vgg_16',
num_classes=1001,
is_training=False)
logits, endpoints = model_ctr(preprocess)
session = tf.Session()
session.run(tf.initialize_all_variables())
return TensorflowSlimWrapper(identifier='tf-vgg16',
labels_offset=1,
endpoints=endpoints,
inputs=placeholder,
session=session)
def tfslim_custom():
from model_tools.activations.tensorflow import load_resize_image
import tensorflow as tf
slim = tf.contrib.slim
tf.compat.v1.reset_default_graph()
image_size = 224
placeholder = tf.compat.v1.placeholder(dtype=tf.string, shape=[64])
preprocess = lambda image_path: load_resize_image(image_path, image_size)
preprocess = tf.map_fn(preprocess, placeholder, dtype=tf.float32)
with tf.compat.v1.variable_scope('my_model', values=[preprocess]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
outputs_collections=[end_points_collection]):
net = slim.conv2d(preprocess, 64, [11, 11], 4, padding='VALID', scope='conv1')
net = slim.max_pool2d(net, [5, 5], 5, scope='pool1')
net = slim.max_pool2d(net, [3, 3], 2, scope='pool2')
net = slim.flatten(net, scope='flatten')
net = slim.fully_connected(net, 1000, scope='logits')
endpoints = slim.utils.convert_collection_to_dict(end_points_collection)
session = tf.compat.v1.Session()
session.run(tf.compat.v1.initialize_all_variables())
return TensorflowSlimWrapper(identifier='tf-custom', labels_offset=0,
endpoints=endpoints, inputs=placeholder, session=session)
def build_model_ending_points(img_paths):
image_size = 224
_load_func = lambda image_path: load_resize_image(image_path, image_size)
imgs = tf.map_fn(_load_func, img_paths, dtype=tf.float32)
setting_name = 'cate_res18_exp1'
ending_points, _ = get_network_outputs(
{'images': imgs},
prep_type='mean_std',
model_type='vm_model',
setting_name=setting_name,
#module_name=['encode', 'category'],
module_name=['encode'],
)
for key in ending_points:
ending_points[key] = tf.transpose(ending_points[key], [0, 3, 1, 2])
return ending_points
def _get_imgs_from_paths(self, img_paths):
_load_func = lambda image_path: load_resize_image(image_path, 224)
imgs = tf.map_fn(_load_func, img_paths, dtype=tf.float32)
return imgs