TensorFlow实现打印每一层的输出

技术  /  管理员 发布于 7年前   216

在test.py中可以通过如下代码直接生成带weight的pb文件，也可以通过tf官方的freeze_graph.py将ckpt转为pb文件。

constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,['net_loss/inference/encode/conv_output/conv_output'])with tf.gfile.FastGFile('net_model.pb', mode='wb') as f:  f.write(constant_graph.SerializeToString())

tf1.0中通过带weight的pb文件与get_tensor_by_name函数可以获取每一层的输出

import osimport os.path as opsimport argparseimport timeimport math import tensorflow as tfimport globimport numpy as npimport matplotlib.pyplot as pltimport cv2 os.environ["CUDA_VISIBLE_DEVICES"] = "-1" gragh_path = './model.pb'image_path = './lvds1901.JPG'inputtensorname = 'input_tensor:0'tensorname = 'loss/inference/encode/resize_images/ResizeBilinear'filepath='./net_output.txt'HEIGHT=256WIDTH=256VGG_MEAN = [103.939, 116.779, 123.68] with tf.Graph().as_default():  graph_def = tf.GraphDef()  with tf.gfile.GFile(gragh_path, 'rb') as fid:    serialized_graph = fid.read()    graph_def.ParseFromString(serialized_graph)     tf.import_graph_def(graph_def, name='')     image = cv2.imread(image_path)    image = cv2.resize(image, (WIDTH, HEIGHT), interpolation=cv2.INTER_CUBIC)    image_np = np.array(image)    image_np = image_np - VGG_MEAN    image_np_expanded = np.expand_dims(image_np, axis=0)     with tf.Session() as sess:      ops = tf.get_default_graph().get_operations()      tensor_name = tensorname + ':0'      tensor_dict = tf.get_default_graph().get_tensor_by_name(tensor_name)      image_tensor = tf.get_default_graph().get_tensor_by_name(inputtensorname)      output = sess.run(tensor_dict, feed_dict={image_tensor: image_np_expanded})ftxt = open(filepath,'w')      transform = output.transpose(0, 3, 1, 2)      transform = transform.flatten()      weight_count = 0      for i in transform:        if weight_count % 10 == 0 and weight_count != 0:          ftxt.write('\n')        ftxt.write(str(i) + ',')        weight_count += 1      ftxt.close()

以上这篇TensorFlow实现打印每一层的输出就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持。