def deeplabv3_plus(image):
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
net, end_points = resnet_v1.resnet_v1_50(image,
num_classes=None,
is_training=None,
global_pool=False,
output_stride=16,
spatial_squeeze=False)
# ASPP
aspp = atrous_spp16(net)
with tf.variable_scope('decoder'):
# Low level
low_level_features = end_points[
'resnet_v1_50/block1/unit_2/bottleneck_v1']
low_level_features = slim.conv2d(
low_level_features,
48, [1, 1],
scope='low_level_feature_conv_1x1')
low_level_features_shape = low_level_features.get_shape(
).as_list()[1:3]
# Upsample
net = tf.image.resize_images(aspp, low_level_features_shape)
net = tf.concat([net, low_level_features], axis=3)
net = slim.conv2d(net, 256, [3, 3], scope='conv_3x3_1')
net = slim.conv2d(net, 256, [3, 3], scope='conv_3x3_2')
return net
def resnet_50_network(img_ph,
background=None,
is_training=False,
reuse=False,
scope=None):
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
_, end_points = resnet_v1.resnet_v1_50(img_ph,
reuse=reuse,
is_training=is_training)
x = end_points['resnet_v1_50/block3/unit_5/bottleneck_v1']
return x
def model(images, weight_decay=1e-5, is_training=True):
# 图片减去均值
images = mean_image_subtraction(images)
# 101层残差
with slim.arg_scope(resnet_v1.resnet_arg_scope(weight_decay=weight_decay)):
logits, end_points = resnet_v1.resnet_v1_101(images,
is_training=is_training,
scope='resnet_v1_101')
with tf.variable_scope('feature_fusion', values=[end_points.values]):
batch_norm_params = {
'decay': 0.997,
'epsilon': 1e-5,
'scale': True,
'is_training': is_training
}
with slim.arg_scope(
[slim.conv2d],
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=slim.l2_regularizer(weight_decay)):
f = [
end_points['pool5'], end_points['pool4'], end_points['pool3'],
end_points['pool2']
]
for i in range(4):
print('Shape of f_{} {}'.format(i, f[i].shape))
g = [None, None, None, None]
h = [None, None, None, None]
for i in range(4):
h[i] = slim.conv2d(f[i], 256, 1)
for i in range(4):
print('Shape of h_{} {}'.format(i, h[i].shape))
g[0] = RefineBlock(h[0])
g[1] = RefineBlock(g[0], h[1])
g[2] = RefineBlock(g[1], h[2])
g[3] = RefineBlock(g[2], h[3])
g[3] = unpool(g[3], scale=4)
f_score = slim.conv2d(g[3],
21,
1,
activation_fn=tf.nn.relu,
normalizer_fn=None)
return f_score
def resnet(self, image):
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits, end_points = resnet_v1.resnet_v1_50(image, is_training=True, scope="resnet_v1_50")
with tf.variable_scope('feature_fusion', values=[end_points.values]):
batch_norm_params = {
'decay': 0.997,
'epsilon': 1e-5,
'scale': True,
'is_training': True
}
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=slim.l2_regularizer(1e-4)):
f = [end_points['pool5'], end_points['pool4'],
end_points['pool3'], end_points['pool2'],
end_points['pool1'], end_points['pool0']]
global_pool = tf.reduce_mean(f[0], [1, 2], name='pool5', keep_dims=True)
for i in range(6):
print('Shape of f_{} {}'.format(i, f[i].shape))
g = [None, None, None, None, None, None]
h = [None, None, None, None, None, None]
num_outputs = [128, 128, 64, 32, 32, 16]
for i in range(4):
if i == 0:
global_pool = self.unpool(global_pool, size=[tf.shape(f[i])[1], tf.shape(f[i])[2]])
h[i] = tf.concat([f[i], global_pool],
axis=-1)
h[i] = slim.conv2d(h[i], num_outputs[i], 3)
# h[i] = f[i]
else:
c1_1 = slim.conv2d(tf.concat([g[i - 1], f[i]], axis=-1), num_outputs[i], 1)
h[i] = slim.conv2d(c1_1, num_outputs[i], [3, 3])
if i < 3:
g[i] = self.unpool(h[i], size=[tf.shape(f[i + 1])[1], tf.shape(f[i + 1])[2]])
if i == 3:
# c1_1 = slim.conv2d(tf.concat([g[i], end_points['pool1']], axis=-1), num_outputs[i], 1)
g[i] = self.unpool(h[i])
g[i] = slim.conv2d(g[i], num_outputs[i], 3, padding="SAME")
print('Shape of h_{} {}, g_{} {}'.format(i, h[i].shape, i, g[i].shape))
mask = self.unpool(g[3])
mask = slim.conv2d(mask, num_outputs[3], 3, padding="SAME")
output_map = slim.conv2d(mask, 2, 1, activation_fn=None, normalizer_fn=None)
output_contour = slim.conv2d(mask, 2, 1, activation_fn=None, normalizer_fn=None)
return output_map, output_contour
def resnet_18_network(img_ph,
background=None,
is_training=False,
reuse=False,
scope=None):
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
_, end_points = resnet_v1.resnet_v1_18(img_ph,
reuse=reuse,
is_training=is_training)
if scope is None:
x = end_points['resnet_v1_18/block3/unit_1/bottleneck_v1']
else:
x = end_points['%s/resnet_v1_18/block3/unit_1/bottleneck_v1' % scope]
return x
import tensorflow as tf
slim = tf.contrib.slim
from resnet import resnet_v1
import numpy as np
from scipy.misc import imread, imresize
checkpoint_file = './data/resnet-50/resnet_v1_50.ckpt'
sample_images = ['dog.jpg', 'panda.jpg']
#Load the model
sess = tf.Session()
inputs = tf.placeholder(tf.float32, [None, 160, 160, 3])
background = tf.placeholder(tf.float32, [None, 160, 160, 3])
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
net, end_points = resnet_v1.resnet_v1_50(inputs,
background,
1000,
is_training=False)
saver = tf.train.Saver()
saver.restore(sess, checkpoint_file)
# import pdb; pdb.set_trace()
# def get_candidate(result, topn=5):
# with open('./imagenet.txt') as f:
# dictionary = eval(f.read())
# idxes = np.argsort(result[0])[-topn:][::-1]
# results = []
# for idx in idxes:
# results.append(dictionary[idx])
# return results
# imgs = ['./cat.jpg', './car.jpeg', './house.jpeg']