keras实现theano和tensorflow训练的模型相互转换

我就废话不多说了，大家还是直接看代码吧~

</pre><pre code_snippet_id="1947416" snippet_file_name="blog_20161025_1_3331239" name="code" class="python">

# coding:utf-8
"""
If you want to load pre-trained weights that include convolutions (layers Convolution2D or Convolution1D),
be mindful of this: Theano and TensorFlow implement convolution in different ways (TensorFlow actually implements correlation, much like Caffe),
and thus, convolution kernels trained with Theano (resp. TensorFlow) need to be converted before being with TensorFlow (resp. Theano).
"""
from keras import backend as K
from keras.utils.np_utils import convert_kernel
from text_classifier import keras_text_classifier
import sys

def th2tf( model):
 import tensorflow as tf
 ops = []
 for layer in model.layers:
   if layer.__class__.__name__ in ['Convolution1D', 'Convolution2D']:
     original_w = K.get_value(layer.W)
     converted_w = convert_kernel(original_w)
     ops.append(tf.assign(layer.W, converted_w).op)
 K.get_session().run(ops)
 return model

def tf2th(model):
 for layer in model.layers:
   if layer.__class__.__name__ in ['Convolution1D', 'Convolution2D']:
     original_w = K.get_value(layer.W)
     converted_w = convert_kernel(original_w)
     K.set_value(layer.W, converted_w)
 return model

def conv_layer_converted(tf_weights, th_weights, m = 0):
 """
 :param tf_weights:
 :param th_weights:
 :param m: 0-tf2th, 1-th2tf
 :return:
 """
 if m == 0: # tf2th
   tc = keras_text_classifier(weights_path=tf_weights)
   model = tc.loadmodel()
   model = tf2th(model)
   model.save_weights(th_weights)
 elif m == 1: # th2tf
   tc = keras_text_classifier(weights_path=th_weights)
   model = tc.loadmodel()
   model = th2tf(model)
   model.save_weights(tf_weights)
 else:
   print("0-tf2th, 1-th2tf")
   return
if __name__ == '__main__':
 if len(sys.argv) < 4:
   print("python tf_weights th_weights <0|1>\n0-tensorflow to theano\n1-theano to tensorflow")
   sys.exit(0)
 tf_weights = sys.argv[1]
 th_weights = sys.argv[2]
 m = int(sys.argv[3])
 conv_layer_converted(tf_weights, th_weights, m)

补充知识：keras学习之修改底层为TensorFlow还是theano

我们知道，keras的底层是TensorFlow或者theano

要知道我们是用的哪个为底层，只需要import keras即可显示

修改方法：

打开

修改

来源：https://blog.csdn.net/cdj0311/article/details/52918687