/keras.json file to tensorflow.
Official document: https://keras.io/backend/ The operation will go wrong because the convolution in tensorflow is actually related, and the convolution in theano is the real convolution! ! ! ! Therefore, when switching backend, the convolution kernel needs to be flipped. See : https://github.com/fchollet/keras/wiki/Converting-convolution-kernels-from-Theano-to-TensorFlow-and-vice-versa general conversion The code is as follows (theano and tensorflow are converted to each other):
from keras import backend as K
from keras.utils.np_utils import convert_kernel
model = model_from_json(open(os.path.join('.', 'model.json')).read())
model.load_weights(os.path.join('.', 'model_weights.h5'))
for layer in model.layers:
if layer.__class__.__name__ in ['Convolution1D', 'Convolution2D','Convolution3D', 'AtrousConvolution2D']:
original_w = K.get_value(layer.W)
converted_w = convert_kernel(original_w)
K.set_value(layer.W, converted_w)
print('running')
K.get_session().run(ops)
print('saving')
model.save_weights('model_weights_anotherBackend.h5')
The tensoflow dedicated conversion code is as follows:
from keras import backend as K
from keras.utils.np_utils import convert_kernel
import tensorflow as tf
model = model_from_json(open(os.path.join('.', 'model.json')).read())
model.load_weights(os.path.join('.', 'model_weights.h5'))
ops = []
for layer in model.layers:
if layer.__class__.__name__ in ['Convolution1D', 'Convolution2D', 'Convolution3D', 'AtrousConvolution2D']:
original_w = K.get_value(layer.W)
print(layer.W.name)
print('\t',end='')
print(layer.W.get_shape().to_list())
converted_w = convert_kernel(original_w)
ops.append(tf.assign(layer.W, converted_w).op)
print('running')
K.get_session().run(ops)
print('saving')
model.save_weights('model_weights_tensorflow.h5')