从张量流模型中获取权重

Question

您好，我想从 tensorflow 微调 VGG 模型。我有两个问题。

如何从网络中获取权重？我的 trainable_variables returns 空列表。

我使用了这里的现有模型：https://github.com/ry/tensorflow-vgg16。我发现 post 关于获取权重，但这对我不起作用，因为 import_graph_def。

import tensorflow as tf
import PIL.Image
import numpy as np

with open("../vgg16.tfmodel", mode='rb') as f:
  fileContent = f.read()

graph_def = tf.GraphDef()
graph_def.ParseFromString(fileContent)

images = tf.placeholder("float", [None, 224, 224, 3])

tf.import_graph_def(graph_def, input_map={ "images": images })
print("graph loaded from disk")

graph = tf.get_default_graph()

cat = np.asarray(PIL.Image.open('../cat224.jpg'))
print(cat.shape)
init = tf.initialize_all_variables()

with tf.Session(graph=graph) as sess:
  print(tf.trainable_variables() )
  sess.run(init)

Answer 1

这 pretrained VGG-16 model encodes all of the model parameters as tf.constant() ops. (See, for example, the calls to tf.constant() here.) As a result, the model parameters would not appear in tf.trainable_variables(), and the model is not mutable without substantial surgery: you would need to replace the constant nodes with tf.Variable 个以相同值开始的对象，以便继续训练。

一般来说，在导入图形进行再训练时，tf.train.import_meta_graph() function should be used, as this function loads additional metadata (including the collections of variables). The tf.import_graph_def() 函数级别较低，不会填充这些集合。

从张量流模型中获取权重

Get weights from tensorflow model

python

neural-network

conv-neural-network

tensorflow