Keras模型在Theano与TensorFlow后端间的权重转换方法

模型权重转换代码实现

# 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)