|
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
|
1 |
|
from keras.models import model_from_json |
|
9
|
1 |
|
import keras |
|
10
|
|
|
|
|
11
|
1 |
|
import json |
|
12
|
1 |
|
import numpy as np |
|
13
|
1 |
|
import os |
|
14
|
1 |
|
import uuid |
|
15
|
1 |
|
from collections import namedtuple |
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
1 |
|
TrainedModel = namedtuple( |
|
19
|
|
|
'TrainedModel', ['history', 'model']) |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
1 |
|
def savemodel(model, filepath, modelname): |
|
23
|
|
|
""" Save model to json file and weights to npy file |
|
24
|
|
|
|
|
25
|
|
|
Parameters |
|
26
|
|
|
---------- |
|
27
|
|
|
model : Keras object |
|
28
|
|
|
model to save |
|
29
|
|
|
filepath : str |
|
30
|
|
|
directory where the data will be stored |
|
31
|
|
|
modelname : str |
|
32
|
|
|
name of the model to be used in the filename |
|
33
|
|
|
|
|
34
|
|
|
Returns |
|
35
|
|
|
---------- |
|
36
|
|
|
json_path : str |
|
37
|
|
|
Path to json file with architecture |
|
38
|
|
|
numpy_path : str |
|
39
|
|
|
Path to npy file with weights |
|
40
|
|
|
""" |
|
41
|
1 |
|
json_string = model.to_json() # save architecture to json string |
|
42
|
1 |
|
json_path = os.path.join(filepath, modelname + '_architecture.json') |
|
43
|
1 |
|
with open(json_path, 'w') as outfile: |
|
44
|
1 |
|
json.dump(json_string, outfile, sort_keys=True, indent=4, |
|
45
|
|
|
ensure_ascii=False) |
|
46
|
1 |
|
wweights = model.get_weights() # get weight from model |
|
47
|
1 |
|
numpy_path = os.path.join(filepath, modelname + '_weights') |
|
48
|
1 |
|
np.save(numpy_path, |
|
49
|
|
|
wweights) # save weights in npy file |
|
50
|
1 |
|
return json_path, numpy_path |
|
51
|
|
|
|
|
52
|
|
|
|
|
53
|
1 |
|
def loadmodel(filepath, modelname): |
|
54
|
|
|
""" Load model + weights from json + npy file, respectively |
|
55
|
|
|
|
|
56
|
|
|
Parameters |
|
57
|
|
|
---------- |
|
58
|
|
|
filepath : str |
|
59
|
|
|
directory where the data will be stored |
|
60
|
|
|
modelname : str |
|
61
|
|
|
name of the model to be used in the filename |
|
62
|
|
|
|
|
63
|
|
|
Returns |
|
64
|
|
|
---------- |
|
65
|
|
|
model_repro : Keras object |
|
66
|
|
|
reproduced model |
|
67
|
|
|
""" |
|
68
|
|
|
with open(os.path.join(filepath, modelname + '_architecture.json'), 'r') as outfile: |
|
|
|
|
|
|
69
|
|
|
json_string_loaded = json.load(outfile) |
|
70
|
|
|
model_repro = model_from_json(json_string_loaded) |
|
71
|
|
|
# wweights2 = model_repro.get_weights() |
|
72
|
|
|
# extracting the weights would give us the untrained/default weights |
|
73
|
|
|
wweights_recovered = np.load( |
|
74
|
|
|
os.path.join(filepath, modelname + '_weights.npy')) # load the original weights |
|
|
|
|
|
|
75
|
|
|
model_repro.set_weights(wweights_recovered) # now set the weights |
|
76
|
|
|
return model_repro |
|
77
|
|
|
|
|
78
|
|
|
# If we would use standard Keras function, which stores model and weights |
|
79
|
|
|
# in HDF5 format it would look like code below. However, we did not use this |
|
80
|
|
|
# because |
|
81
|
|
|
# https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model |
|
82
|
|
|
# it is not compatible with default Keras version in python3. |
|
83
|
|
|
# from keras.models import load_model |
|
84
|
|
|
# import h5py |
|
85
|
|
|
# modelh5=models[0] |
|
86
|
|
|
# modelh5.save(resultpath+'mymodel.h5') |
|
87
|
|
|
# del modelh5 |
|
88
|
|
|
# modelh5 = load_model(resultpath+'mymodel.h5') |
|
89
|
|
|
|
|
90
|
|
|
|
|
91
|
1 |
|
try: |
|
92
|
1 |
|
import noodles |
|
93
|
|
|
from noodles.serial.numpy import arrays_to_string |
|
94
|
|
|
from noodles.serial.namedtuple import SerNamedTuple |
|
95
|
|
|
from pathlib import Path |
|
96
|
|
|
|
|
97
|
|
|
|
|
98
|
|
|
class SerModel(noodles.serial.Serialiser): |
|
99
|
|
|
def __init__(self, tmpdir): |
|
100
|
|
|
self.tmpdir = Path(tmpdir) |
|
101
|
|
|
super(SerModel, self).__init__(keras.models.Model) |
|
102
|
|
|
|
|
103
|
|
|
def encode(self, obj, make_rec): |
|
104
|
|
|
random_filename = self.tmpdir / (str(uuid.uuid4()) + '.hdf5') |
|
105
|
|
|
obj.save(str(random_filename)) |
|
106
|
|
|
return make_rec({'filename': str(random_filename)}, |
|
107
|
|
|
files=[str(random_filename)], ref=True) |
|
108
|
|
|
|
|
109
|
|
|
def decode(self, cls, data): |
|
110
|
|
|
return keras.models.load_model(data['filename']) |
|
111
|
|
|
|
|
112
|
|
|
|
|
113
|
|
|
def serial_registry(tmpdir='.'): |
|
114
|
|
|
return noodles.serial.Registry( |
|
115
|
|
|
# parent=noodles.serial.pickle() + |
|
116
|
|
|
parent=noodles.serial.base() + arrays_to_string(), |
|
117
|
|
|
types={ |
|
118
|
|
|
keras.models.Model: SerModel(tmpdir), |
|
119
|
|
|
TrainedModel: SerNamedTuple(TrainedModel) |
|
120
|
|
|
} |
|
121
|
|
|
) |
|
122
|
|
|
|
|
123
|
1 |
|
except ImportError: |
|
124
|
|
|
pass |
|
125
|
|
|
|
This check looks for lines that are too long. You can specify the maximum line length.