def define_graph(input_image, batch_size):
print("string define the graph...")
# 卷积层1
conv2d_layer_1 = convolution2d(
input_image,
num_outputs = 32, # filter 个数
kernel_size = (5,5), # filter 的宽和高
activation_fn = tf.nn.relu,
weights_initializer = tf.random_normal_initializer,
stride = [2, 2],
trainable = True)
pool_layer_1 = tf.nn.max_pool(conv2d_layer_1, ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
# 卷积层2
conv2d_layer_2 = convolution2d(
pool_layer_1,
num_outputs = 64,
kernel_size = (5, 5),
activation_fn = tf.nn.relu,
weights_initializer = tf.random_normal_initializer,
stride = (1, 1),
trainable = True)
pool_layer_2 = tf.nn.max_pool(conv2d_layer_2, ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
flattened_layer_2 = tf.reshape(pool_layer_2, [ batch_size, -1 ])
# weight_init 参数也可以接收一个可调用参数，这里使用的了一个lambda 表达式返回了一个截断的正态分布，并指定了标准差
hidden_layer_3 = fully_connected(
flattened_layer_2,
512,
weights_initializer = lambda i, dtype, partition_info: tf.truncated_normal([38912, 512], stddev = 0.1),
activation_fn = tf.nn.relu
)
# 对一些神经元进行dropout，削减他们在模型中的重要性
hidden_layer_3 = tf.nn.dropout(hidden_layer_3, 0.1)
# 输出是前面的层与训练中可用的120个不同狗品种的全连接
final_fully_connected = fully_connected(
hidden_layer_3,
120, # 120 种狗
weights_initializer = lambda i, dtype, partition_info: tf.truncated_normal([512, 120], stddev = 0.1)
)
print("graph is ready")
return final_fully_connected
def encoder(self,x, keep_prob):
e_conv1 = convolution2d(x, 32, [5,5], [2,2], padding='SAME')
#print e_conv1
e_conv2 = convolution2d(e_conv1, 16, [5,5], [2,2], padding='SAME')
#print e_conv2
e_flat1 = flatten(e_conv2)
#print e_flat1
#e_flat1 = tf.nn.dropout(e_flat1 keep_prob)
e_fc1 = fully_connected(e_flat1, self.params['final_layer'], None)
e_fc1 = tf.nn.dropout(e_fc1, keep_prob)
return e_fc1
def create_shallownet(images, scope=None, net=None, dropout=True):
"""
Args:
images: a tensor of shape [B x H x W x C]
net: An optional dict object
scope: The variable scope for the subgraph, defaults to ShallowNet
Returns:
saliency_output: a tensor of shape [B x 48 x 48]
"""
assert len(images.get_shape()) == 4 # [B, H, W, C]
if net is None: net = {}
else: assert isinstance(net, dict)
net['dropout_keep_prob'] = tf.placeholder(tf.float32,
name='dropout_keep_prob')
with tf.variable_scope(scope or 'ShallowNet'):
# CONV
net['conv1'] = convolution2d(
images,
32,
kernel_size=(5, 5),
stride=(1, 1),
padding='VALID',
activation_fn=None, #tf.nn.relu,
weights_initializer=initializers.xavier_initializer_conv2d(
uniform=True),
biases_initializer=tf.constant_initializer(0.0),
variables_collections=['MODEL_VARS'],
scope='conv1')
#net['conv1'] = tflearn.layers.batch_normalization(net['conv1'])
net['conv1'] = tf.nn.relu(net['conv1'])
net['pool1'] = tf.nn.max_pool(net['conv1'],
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
log.info('Conv1 size : %s', net['conv1'].get_shape().as_list())
log.info('Pool1 size : %s', net['pool1'].get_shape().as_list())
net['conv2'] = convolution2d(
net['pool1'],
64,
kernel_size=(3, 3),
stride=(1, 1),
padding='VALID',
activation_fn=None, #tf.nn.relu,
weights_initializer=initializers.xavier_initializer_conv2d(
uniform=True),
biases_initializer=tf.constant_initializer(0.0),
variables_collections=['MODEL_VARS'],
scope='conv2')
#net['conv2'] = tflearn.layers.batch_normalization(net['conv2'])
net['conv2'] = tf.nn.relu(net['conv2'])
net['pool2'] = tf.nn.max_pool(net['conv2'],
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
log.info('Conv2 size : %s', net['conv2'].get_shape().as_list())
log.info('Pool2 size : %s', net['pool2'].get_shape().as_list())
net['conv3'] = convolution2d(
net['pool2'],
32,
kernel_size=(3, 3),
stride=(1, 1),
padding='VALID',
activation_fn=None, #tf.nn.relu,
weights_initializer=initializers.xavier_initializer_conv2d(
uniform=True),
biases_initializer=tf.constant_initializer(0.0),
variables_collections=['MODEL_VARS'],
scope='conv3')
#net['conv3'] = tflearn.layers.batch_normalization(net['conv3'])
net['conv3'] = tf.nn.relu(net['conv3'])
net['pool3'] = tf.nn.max_pool(net['conv3'],
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
log.info('Conv3 size : %s', net['conv3'].get_shape().as_list())
log.info('Pool3 size : %s', net['pool3'].get_shape().as_list())
# FC layer
n_inputs = int(np.prod(net['pool3'].get_shape().as_list()[1:]))
pool3_flat = tf.reshape(net['pool3'], [-1, n_inputs])
net['fc1'] = fully_connected(
pool3_flat,
4802,
activation_fn=None, #tf.nn.relu,
weights_initializer=initializers.xavier_initializer(
uniform=True),
biases_initializer=tf.constant_initializer(0.0),
variables_collections=['MODEL_VARS'],
scope='fc1')
log.info('fc1 size : %s', net['fc1'].get_shape().as_list())
#net['fc1'] = tflearn.layers.batch_normalization(net['fc1'])
net['fc1'] = tf.nn.relu(net['fc1'])
if dropout:
net['fc1'] = tf.nn.dropout(net['fc1'],
net['dropout_keep_prob'])
fc1_slice1, fc1_slice2 = tf.split(
net['fc1'], num_or_size_splits=2, axis=1, name='fc1_slice'
) #syntax probably wrong here for newer tensorflow version
net['max_out'] = tf.maximum(fc1_slice1,
fc1_slice2,
name='fc1_maxout')
log.info('maxout size : %s', net['max_out'].get_shape().as_list())
net['fc2'] = fully_connected(
net['max_out'],
4802,
activation_fn=None, # no relu here
weights_initializer=initializers.xavier_initializer(
uniform=True),
biases_initializer=tf.constant_initializer(0.0),
variables_collections=['MODEL_VARS'],
scope='fc2')
#net['fc2'] = tflearn.layers.batch_normalization(net['fc2'])
net['fc2'] = tf.nn.relu(net['fc2'])
#if dropout:
# net['fc2'] = tf.nn.dropout( net['fc2'], net['dropout_keep_prob'] )
log.info('fc2 size : %s', net['fc2'].get_shape().as_list())
fc2_slice1, fc2_slice2 = tf.split(net['fc2'],
num_or_size_splits=2,
axis=1,
name='fc2_slice')
net['max_out2'] = tf.maximum(fc2_slice1,
fc2_slice2,
name='fc2_maxout')
'''
net['fc3'] = fully_connected(net['max_out2'], 4802,
activation_fn=None, # no relu here
weights_initXializer=initializers.xavier_initializer(uniform=True),
biases_initializer=tf.constant_initializer(0.0),
weight_collections=['MODEL_VARS'], bias_collections=['MODEL_VARS'],
name='fc3')
#net['fc3'] = tflearn.layers.batch_normalization(net['fc3'])
net['fc3'] = tf.nn.relu(net['fc3'])
fc3_slice1, fc3_slice2 = tf.split(1, 2, net['fc3'], name='fc3_slice')
net['max_out3'] = tf.maximum(fc3_slice1, fc3_slice2, name='fc3_maxout')
net['max_out3'] = tflearn.layers.batch_normalization(net['max_out3'])
'''
#net['fc2'] = tf.nn.dropout( net['fc2'], net['dropout_keep_prob'] )
#log.info('fc3 size : %s', net['fc3'].get_shape().as_list())
# debug and summary
#net['fc1'].get_shape().assert_is_compatible_with([None, 4802])
#net['fc2'].get_shape().assert_is_compatible_with([None, 4802])
#net['fc3'].get_shape().assert_is_compatible_with([None, 4802])
#for t in [self.conv1, self.conv2, self.conv3,
# self.pool1, self.pool2, self.pool3,
# self.fc1, self.max_out, self.fc2]:
# _add_activation_histogram_summary(t)
net['saliency'] = tf.reshape(net['max_out2'], [-1, 49, 49],
name='saliency')
return net['saliency']
conv1=tf.layers.conv2d(
tf.image.convert_image_dtype(x_img,dtype=tf.float32),
filters=32, # 输出通道由1->32
kernel_size=(3,3), # 3x3卷积核
activation=tf.nn.relu,
padding='SAME',
kernel_initializer=tf.random_uniform_initializer,
bias_initializer=tf.random_normal_initializer
)
"""
conv1=convolution2d(
# tf.image.convert_image_dtype(x_img,dtype=tf.float32),
x,
num_outputs=64,
kernel_size=(3,3),
activation_fn=tf.nn.relu,
normalizer_fn=tf.layers.batch_normalization,
weights_initializer=tf.random_uniform_initializer,
biases_initializer=tf.random_normal_initializer,
trainable=True
)
conv1=tf.nn.max_pool(conv1,[1,2,2,1],[1,2,2,1],padding="SAME") # [n,16,16,64]
conv1 = tf.nn.dropout(conv1, keep)
conv1 = tf.nn.lrn(conv1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75,
name='norm1')
# convolution2
"""
conv2=tf.layers.conv2d(
conv1,
