-
Notifications
You must be signed in to change notification settings - Fork 0
/
reformat_model.py
35 lines (32 loc) · 1.21 KB
/
reformat_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
import os, sys
from keras.models import load_model, Sequential
from keras.layers import Activation, Layer
import numpy as np
import imageio
from keras.datasets import cifar10
from keras import optimizers
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
if __name__ == '__main__':
modelname = 'model_that_does_not_seperat_activation_from_conv/dense'
modelpath = './models/'+modelname+'.h5'
model = load_model(modelpath)
model.summary()
layers = model.layers
model2 = Sequential()
for layer in layers:
# print( layer.get_config()['activation'] )
if 'activation' in layer.get_config().keys():
act_name = layer.get_config()['activation']
config = layer.get_config()
config.pop('activation')
model2.add(layer.__class__.from_config(config))
model2.add(Activation(act_name))
else:
model2.add(layer)
model2.compile(loss='categorical_crossentropy', optimizer=optimizers.SGD(), metrics=['accuracy'])
model2.summary()
model2.load_weights(modelpath)
model2.save('./models/'+modelname+'_2.h5')
model_yaml = model2.to_yaml()
with open(modelname+".yaml", "w") as yaml_file:
yaml_file.write(model_yaml)