获取某个节点的张量
graph = tf.get_default_graph()
inputs = graph.get_tensor_by_name('placeholder:0')
获得所有操作节点
tensor_name_list = [tensor.name for tensor in graph.as_graph_def().node]
import tensorflow as tf
import numpy as np
import os
def vgg16_layer(image_batch):
""" vgg16_conv5_3 layer """
# Gray to color if necessary.
if image_batch.shape[3] == 1:
node_x = tf.nn.conv2d(image_batch, np.ones((1, 1, 1, 3)),
np.ones(4).tolist(), 'VALID')
else:
node_x = image_batch
# Subtract trained average image.
average_rgb = tf.get_variable('average_rgb1', 3, dtype=image_batch.dtype)
node_x = node_x - average_rgb
# VGG16-convolution FUNC
def vgg_conv(inputs, layer_name, out_dim, with_relu):
if with_relu:
activation = tf.nn.relu
else:
activation = None
result = tf.layers.conv2d(inputs, out_dim, [3, 3], 1, padding='same',
activation=activation,
name='conv1_%s' % layer_name)
return result
# VGG16-Pooling FUNC
def vgg_pool(inputs):
return tf.layers.max_pooling2d(inputs, 2, 2)
node_x = vgg_conv(node_x, '1_1', 64, True)
node_x = vgg_conv(node_x, '1_2', 64, False)
node_x = vgg_pool(node_x)
node_x = tf.nn.relu(node_x)
node_x = vgg_conv(node_x, '2_1', 128, True)
node_x = vgg_conv(node_x, '2_2', 128, False)
node_x = vgg_pool(node_x)
node_x = tf.nn.relu(node_x)
#
# node_x = vgg_conv(node_x, '3_1', 256, True)
# node_x = vgg_conv(node_x, '3_2', 256, True)
# node_x = vgg_conv(node_x, '3_3', 256, False)
# node_x = vgg_pool(node_x)
# node_x = tf.nn.relu(node_x)
#
# node_x = vgg_conv(node_x, '4_1', 512, True)
# node_x = vgg_conv(node_x, '4_2', 512, True)
# node_x = vgg_conv(node_x, '4_3', 512, False)
# node_x = vgg_pool(node_x)
# node_x = tf.nn.relu(node_x)
#
# node_x = vgg_conv(node_x, '5_1', 512, True)
# node_x = vgg_conv(node_x, '5_2', 512, True)
# node_x = vgg_conv(node_x, '5_3', 512, False)
return node_x
if __name__ == '__main__':
input = tf.placeholder(tf.float32,(None,640,480,3))
output = vgg16_layer(input)
with tf.Session() as sess:
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.save(sess,'./model/model1.ckpt')
import tensorflow as tf
import numpy as np
import os
def vgg16_layers(image_batch):
""" vgg16_conv5_3 layer """
# Gray to color if necessary.
# if image_batch.shape[3] == 1:
# node_x = tf.nn.conv2d(image_batch, np.ones((1, 1, 1, 3)),
# np.ones(4).tolist(), 'VALID')
# else:
node_x = image_batch
# Subtract trained average image.
# average_rgb = tf.get_variable('average_rgb2', 3, dtype=image_batch.dtype)
# node_x = node_x - average_rgb
# VGG16-convolution FUNC
def vgg_conv(inputs, layer_name, out_dim, with_relu):
if with_relu:
activation = tf.nn.relu
else:
activation = None
result = tf.layers.conv2d(inputs, out_dim, [3, 3], 1, padding='same',
activation=activation,
name='conv2_%s' % layer_name)
return result
# VGG16-Pooling FUNC
def vgg_pool(inputs):
return tf.layers.max_pooling2d(inputs, 2, 2)
node_x = vgg_conv(node_x, '3_1', 256, True)
node_x = vgg_conv(node_x, '3_2', 256, True)
node_x = vgg_conv(node_x, '3_3', 256, False)
node_x = vgg_pool(node_x)
node_x = tf.nn.relu(node_x)
node_x = vgg_conv(node_x, '4_1', 512, True)
node_x = vgg_conv(node_x, '4_2', 512, True)
node_x = vgg_conv(node_x, '4_3', 512, False)
node_x = vgg_pool(node_x)
node_x = tf.nn.relu(node_x)
node_x = vgg_conv(node_x, '5_1', 512, True)
node_x = vgg_conv(node_x, '5_2', 512, True)
node_x = vgg_conv(node_x, '5_3', 512, False)
return node_x
if __name__ == '__main__':
input = tf.placeholder(tf.float32,(None,160,120,128))
output = vgg16_layers(input)
with tf.Session() as sess:
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.save(sess,'./model/model2.ckpt')
import tensorflow as tf
import os
import model1,model2
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
inputs = tf.placeholder(tf.float32,(None,640,480,3))
output = model1.vgg16_layer(inputs)
output = model2.vgg16_layers(output)
var_list = tf.global_variables()
saver = tf.train.Saver(var_list[:9])
with tf.Session() as sess:
sess.run(tf.variables_initializer(var_list[:9]))
saver.restore(sess,'./model/model1.ckpt')
saver = tf.train.Saver(var_list[9:])
sess.run(tf.variables_initializer(var_list[9:]))
saver.restore(sess,'./model/model2.ckpt')
saver = tf.train.Saver(tf.global_variables())
saver.save(sess,'./model/merge.ckpt')
print('done')
# graph1 = tf.train.import_meta_graph('./model/model1.ckpt.meta')
# var_list = tf.global_variables()
#
# with tf.Session() as sess:
# saver = tf.train.Saver()
# sess.run(tf.variables_initializer(var_list))
# saver.restore(sess,'./model/model1.ckpt')
#
# graph2 = tf.train.import_meta_graph('./model/model2.ckpt.meta')
# var_list = tf.global_variables()[9:]
# with tf.Session() as sess:
# sess.run(tf.variables_initializer(var_list))
# saver.restore(sess, './model/model2.ckpt')
# print('done')
# saver.save(sess, './model/merge.ckpt')
