Keras vgg转移学习保存/加载自定义完全连接图层

Question

在转移学习中，只能在训练期间保存完全连接的图层，并在以后加载以便继续训练？如果可能的话，该怎么做？

Answer 1

这可能不是最好的解决方案，但我认为这可以解决您的问题。

import pickle 

#set a name for all fully connected layers. 
model.add(Dense(...,name='fc1')) 
model.add(Dense(...,name='fc2')) 
model.add(Dense(...,name='fc3')) 


layers_to_save = ['fc1','fc2','fc3'] # add here any layer you want to save 

# Save weights to a dictionary 
weights_dict = dict([(layer.name, layer.get_weights()) for layer in model.layers if layer.name in layers_to_save]) 

with open('filename.pickle', 'wb') as handle: 
    pickle.dump(weights_dict, handle, protocol=pickle.HIGHEST_PROTOCOL) 




# Load weights 
with open('filename.pickle', 'rb') as handle: 
    weights_dict = pickle.load(handle) 

for name in layers_to_save: 
    model.get_layer(name).set_weights(weights_dict[name]) 

您还可以检查此keras blog post下半年的另一种方法。