def _create_conv(self, input_im):
VGG_MEAN = [103.939, 116.779, 123.68]
red, green, blue = tf.split(axis=3,
num_or_size_splits=3,
value=input_im)
input_bgr = tf.concat(axis=3,
values=[
blue - VGG_MEAN[0],
green - VGG_MEAN[1],
red - VGG_MEAN[2],
])
data_dict = {}
if self._is_load:
data_dict = np.load(self._pre_train_path, encoding='latin1').item()
arg_scope = tf.contrib.framework.arg_scope
with arg_scope([conv],
nl=tf.nn.relu,
trainable=False,
data_dict=data_dict):
conv1_1 = conv(input_bgr, 3, 64, 'conv1_1')
conv1_2 = conv(conv1_1, 3, 64, 'conv1_2')
pool1 = max_pool(conv1_2, 'pool1', padding='SAME')
conv2_1 = conv(pool1, 3, 128, 'conv2_1')
conv2_2 = conv(conv2_1, 3, 128, 'conv2_2')
pool2 = max_pool(conv2_2, 'pool2', padding='SAME')
conv3_1 = conv(pool2, 3, 256, 'conv3_1')
conv3_2 = conv(conv3_1, 3, 256, 'conv3_2')
conv3_3 = conv(conv3_2, 3, 256, 'conv3_3')
conv3_4 = conv(conv3_3, 3, 256, 'conv3_4')
pool3 = max_pool(conv3_4, 'pool3', padding='SAME')
conv4_1 = conv(pool3, 3, 512, 'conv4_1')
conv4_2 = conv(conv4_1, 3, 512, 'conv4_2')
conv4_3 = conv(conv4_2, 3, 512, 'conv4_3')
conv4_4 = conv(conv4_3, 3, 512, 'conv4_4')
pool4 = max_pool(conv4_4, 'pool4', padding='SAME')
conv5_1 = conv(pool4, 3, 512, 'conv5_1')
conv5_2 = conv(conv5_1, 3, 512, 'conv5_2')
conv5_3 = conv(conv5_2, 3, 512, 'conv5_3')
conv5_4 = conv(conv5_3, 3, 512, 'conv5_4')
return conv5_4
def inception_layer(inputs,
conv_11_size,
conv_33_reduce_size, conv_33_size,
conv_55_reduce_size, conv_55_size,
pool_size,
data_dict={},
trainable=False,
name='inception'):
arg_scope = tf.contrib.framework.arg_scope
with arg_scope([conv], nl=tf.nn.relu, trainable=trainable,
data_dict=data_dict):
conv_11 = conv(inputs, 1, conv_11_size, '{}_1x1'.format(name))
conv_33_reduce = conv(inputs, 1, conv_33_reduce_size,
'{}_3x3_reduce'.format(name))
conv_33 = conv(conv_33_reduce, 3, conv_33_size, '{}_3x3'.format(name))
conv_55_reduce = conv(inputs, 1, conv_55_reduce_size,
'{}_5x5_reduce'.format(name))
conv_55 = conv(conv_55_reduce, 5, conv_55_size, '{}_5x5'.format(name))
pool = max_pool(inputs, '{}_pool'.format(name), stride=1,
padding='SAME', filter_size=3)
convpool = conv(pool, 1, pool_size, '{}_pool_proj'.format(name))
return tf.concat([conv_11, conv_33, conv_55, convpool],
3, name='{}_concat'.format(name))
def _create_conv(self, inputs):
self.dropout = tf.placeholder(tf.float32, name='dropout')
mlpconv_1 = mlpconv(inputs,
filter_size=8,
hidden_size=[96, 96],
name='mlpconv_1',
wd=self._wd)
# mlpconv_1 = mlpconv(
# inputs,
# filter_size=5,
# hidden_size=[192, 160, 96],
# name='mlpconv_1',
# wd=self._wd)
mlpconv_1 = max_pool(mlpconv_1, 'pool1', padding='SAME')
mlpconv_1 = dropout(mlpconv_1, self.dropout, self._is_traing)
mlpconv_1 = batch_norm(mlpconv_1, train=self._is_traing, name='bn_1')
mlpconv_2 = mlpconv(mlpconv_1,
filter_size=8,
hidden_size=[192, 192],
name='mlpconv_2',
wd=self._wd)
# mlpconv_2 = mlpconv(
# mlpconv_1,
# filter_size=5,
# hidden_size=[192, 192, 192],
# name='mlpconv_2',
# wd=self._wd)
mlpconv_2 = max_pool(mlpconv_2, 'pool2', padding='SAME')
mlpconv_2 = dropout(mlpconv_2, self.dropout, self._is_traing)
mlpconv_2 = batch_norm(mlpconv_2, train=self._is_traing, name='bn_2')
mlpconv_3 = mlpconv(mlpconv_2,
filter_size=5,
hidden_size=[192, self._n_class],
name='mlpconv_3',
wd=self._wd)
# mlpconv_3 = mlpconv(
# mlpconv_2,
# filter_size=3,
# hidden_size=[192, 192, self._n_class],
# name='mlpconv_3',
# wd=self._wd)
# mlpconv_3 = max_pool(mlpconv_3, 'pool3', padding='SAME')
# mlpconv_3 = dropout(pool3, 0.5, self._is_traing)
return mlpconv_3
def _create_conv(self, inputs, data_dict):
arg_scope = tf.contrib.framework.arg_scope
with arg_scope([conv],
trainable=self._trainable,
data_dict=data_dict,
nl=tf.nn.relu):
conv1 = conv(inputs, 7, 64, name='conv1_7x7_s2', stride=2)
padding1 = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]])
conv1_pad = tf.pad(conv1, padding1, 'CONSTANT')
pool1 = max_pool(conv1_pad,
'pool1',
padding='VALID',
filter_size=3,
stride=2)
pool1_lrn = tf.nn.local_response_normalization(pool1,
depth_radius=2,
alpha=2e-05,
beta=0.75,
name='pool1_lrn')
conv2_reduce = conv(pool1_lrn, 1, 64, name='conv2_3x3_reduce')
conv2 = conv(conv2_reduce, 3, 192, name='conv2_3x3')
padding2 = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]])
conv2_pad = tf.pad(conv2, padding1, 'CONSTANT')
pool2 = max_pool(conv2_pad,
'pool2',
padding='VALID',
filter_size=3,
stride=2)
pool2_lrn = tf.nn.local_response_normalization(pool2,
depth_radius=2,
alpha=2e-05,
beta=0.75,
name='pool2_lrn')
with arg_scope([inception_layer],
trainable=self._trainable,
data_dict=data_dict):
inception3a = inception_layer(pool2_lrn,
64,
96,
128,
16,
32,
32,
name='inception_3a')
inception3b = inception_layer(inception3a,
128,
128,
192,
32,
96,
64,
name='inception_3b')
pool3 = max_pool(inception3b,
'pool3',
padding='SAME',
filter_size=3,
stride=2)
inception4a = inception_layer(pool3,
192,
96,
208,
16,
48,
64,
name='inception_4a')
inception4b = inception_layer(inception4a,
160,
112,
224,
24,
64,
64,
name='inception_4b')
inception4c = inception_layer(inception4b,
128,
128,
256,
24,
64,
64,
name='inception_4c')
inception4d = inception_layer(inception4c,
112,
144,
288,
32,
64,
64,
name='inception_4d')
inception4e = inception_layer(inception4d,
256,
160,
320,
32,
128,
128,
name='inception_4e')
pool4 = max_pool(inception4e,
'pool4',
padding='SAME',
filter_size=3,
stride=2)
inception5a = inception_layer(pool4,
256,
160,
320,
32,
128,
128,
name='inception_5a')
inception5b = inception_layer(inception5a,
384,
192,
384,
48,
128,
128,
name='inception_5b')
return inception5b