|
1
|
|
|
""" |
|
2
|
|
|
Summary: |
|
3
|
|
|
Functions to save and store a model. The current keras |
|
4
|
|
|
function to do this does not work in python3. Therefore, we |
|
5
|
|
|
implemented our own functions until the keras functionality has matured. |
|
6
|
|
|
Example function calls in 'Tutorial mcfly on PAMAP2.ipynb' |
|
7
|
|
|
""" |
|
8
|
|
|
from keras.models import model_from_json |
|
9
|
|
|
import json |
|
10
|
|
|
import numpy as np |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
def savemodel(model, filepath, modelname): |
|
14
|
|
|
""" Save model to json file and weights to npy file |
|
15
|
|
|
|
|
16
|
|
|
Parameters |
|
17
|
|
|
---------- |
|
18
|
|
|
model : Keras object |
|
19
|
|
|
model to save |
|
20
|
|
|
filepath : str |
|
21
|
|
|
directory where the data will be stored |
|
22
|
|
|
modelname : str |
|
23
|
|
|
name of the model to be used in the filename |
|
24
|
|
|
|
|
25
|
|
|
Returns |
|
26
|
|
|
---------- |
|
27
|
|
|
json_path : str |
|
28
|
|
|
Path to json file with architecture |
|
29
|
|
|
numpy_path : str |
|
30
|
|
|
Path to npy file with weights |
|
31
|
|
|
""" |
|
32
|
|
|
json_string = model.to_json() # save architecture to json string |
|
33
|
|
|
json_path = filepath + modelname + '_architecture.json' |
|
34
|
|
|
with open(json_path, 'w') as outfile: |
|
35
|
|
|
json.dump(json_string, outfile, sort_keys=True, indent=4, |
|
36
|
|
|
ensure_ascii=False) |
|
37
|
|
|
wweights = model.get_weights() # get weight from model |
|
38
|
|
|
numpy_path = filepath + modelname + '_weights' |
|
39
|
|
|
np.save(numpy_path, |
|
40
|
|
|
wweights) # save weights in npy file |
|
41
|
|
|
return json_path, numpy_path |
|
42
|
|
|
|
|
43
|
|
|
|
|
44
|
|
|
def loadmodel(filepath, modelname): |
|
45
|
|
|
""" Load model + weights from json + npy file, respectively |
|
46
|
|
|
|
|
47
|
|
|
Parameters |
|
48
|
|
|
---------- |
|
49
|
|
|
filepath : str |
|
50
|
|
|
directory where the data will be stored |
|
51
|
|
|
modelname : str |
|
52
|
|
|
name of the model to be used in the filename |
|
53
|
|
|
|
|
54
|
|
|
Returns |
|
55
|
|
|
---------- |
|
56
|
|
|
model_repro : Keras object |
|
57
|
|
|
reproduced model |
|
58
|
|
|
""" |
|
59
|
|
|
with open(filepath + modelname + '_architecture.json', 'r') as outfile: |
|
60
|
|
|
json_string_loaded = json.load(outfile) |
|
61
|
|
|
model_repro = model_from_json(json_string_loaded) |
|
62
|
|
|
# wweights2 = model_repro.get_weights() |
|
63
|
|
|
# extracting the weights would give us the untrained/default weights |
|
64
|
|
|
wweights_recovered = np.load( |
|
65
|
|
|
filepath + modelname + '_weights.npy') # load the original weights |
|
66
|
|
|
model_repro.set_weights(wweights_recovered) # now set the weights |
|
67
|
|
|
return model_repro |
|
68
|
|
|
|
|
69
|
|
|
# If we would use standard Keras function, which stores model and weights |
|
70
|
|
|
# in HDF5 format it would look like code below. However, we did not use this |
|
71
|
|
|
# because |
|
72
|
|
|
# https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model |
|
73
|
|
|
# it is not compatible with default Keras version in python3. |
|
74
|
|
|
# from keras.models import load_model |
|
75
|
|
|
# import h5py |
|
76
|
|
|
# modelh5=models[0] |
|
77
|
|
|
# modelh5.save(resultpath+'mymodel.h5') |
|
78
|
|
|
# del modelh5 |
|
79
|
|
|
# modelh5 = load_model(resultpath+'mymodel.h5') |
|
80
|
|
|
|
NLeSC /
mcfly
