|
1
|
|
|
""" |
|
2
|
|
|
Summary: |
|
3
|
|
|
This module provides the main functionality of mcfly: searching for an |
|
4
|
|
|
optimal model architecture. The work flow is as follows: |
|
5
|
|
|
Function generate_models from modelgen.py generates and compiles models. |
|
6
|
|
|
Function train_models_on_samples trains those models. |
|
7
|
|
|
Function plotTrainingProcess plots the training process. |
|
8
|
|
|
Function find_best_architecture is wrapper function that combines |
|
9
|
|
|
these steps. |
|
10
|
|
|
Example function calls can be found in the tutorial notebook |
|
11
|
|
|
'EvaluateDifferentModels.ipynb'. |
|
12
|
|
|
""" |
|
13
|
1 |
|
import numpy as np |
|
14
|
1 |
|
from . import modelgen |
|
15
|
1 |
|
from sklearn import neighbors, metrics |
|
16
|
1 |
|
import warnings |
|
17
|
1 |
|
import json |
|
18
|
1 |
|
import os |
|
19
|
1 |
|
from keras.callbacks import EarlyStopping |
|
20
|
|
|
|
|
21
|
1 |
|
def train_models_on_samples(X_train, y_train, X_val, y_val, models, |
|
22
|
|
|
nr_epochs=5, subset_size=100, verbose=True, |
|
23
|
|
|
outputfile=None, early_stopping=False): |
|
24
|
|
|
""" |
|
25
|
|
|
Given a list of compiled models, this function trains |
|
26
|
|
|
them all on a subset of the train data. If the given size of the subset is |
|
27
|
|
|
smaller then the size of the data, the complete data set is used. |
|
28
|
|
|
|
|
29
|
|
|
Parameters |
|
30
|
|
|
---------- |
|
31
|
|
|
X_train : numpy array of shape (num_samples, num_timesteps, num_channels) |
|
32
|
|
|
The input dataset for training |
|
33
|
|
|
y_train : numpy array of shape (num_samples, num_classes) |
|
34
|
|
|
The output classes for the train data, in binary format |
|
35
|
|
|
X_val : numpy array of shape (num_samples_val, num_timesteps, num_channels) |
|
36
|
|
|
The input dataset for validation |
|
37
|
|
|
y_val : numpy array of shape (num_samples_val, num_classes) |
|
38
|
|
|
The output classes for the validation data, in binary format |
|
39
|
|
|
models : list of model, params, modeltypes |
|
40
|
|
|
List of keras models to train |
|
41
|
|
|
nr_epochs : int, optional |
|
42
|
|
|
nr of epochs to use for training one model |
|
43
|
|
|
subset_size : |
|
44
|
|
|
The number of samples used from the complete train set |
|
45
|
|
|
verbose : bool, optional |
|
46
|
|
|
flag for displaying verbose output |
|
47
|
|
|
outputfile : str, optional |
|
48
|
|
|
File location to store the model results |
|
49
|
|
|
early_stopping: bool |
|
50
|
|
|
Stop when validation loss does not decrease |
|
51
|
|
|
|
|
52
|
|
|
Returns |
|
53
|
|
|
---------- |
|
54
|
|
|
histories : list of Keras History objects |
|
55
|
|
|
train histories for all models |
|
56
|
|
|
val_accuracies : list of floats |
|
57
|
|
|
validation accuraracies of the models |
|
58
|
|
|
val_losses : list of floats |
|
59
|
|
|
validation losses of the models |
|
60
|
|
|
""" |
|
61
|
|
|
# if subset_size is smaller then X_train, this will work fine |
|
62
|
1 |
|
X_train_sub = X_train[:subset_size, :, :] |
|
63
|
1 |
|
y_train_sub = y_train[:subset_size, :] |
|
64
|
|
|
|
|
65
|
1 |
|
histories = [] |
|
66
|
1 |
|
val_accuracies = [] |
|
67
|
1 |
|
val_losses = [] |
|
68
|
1 |
|
for i, (model, params, model_types) in enumerate(models): |
|
69
|
1 |
|
if verbose: |
|
70
|
|
|
print('Training model %d' % i, model_types) |
|
71
|
1 |
|
if early_stopping: |
|
72
|
|
|
callbacks = [EarlyStopping(monitor='val_loss', patience=0, verbose=verbose, mode='auto')] |
|
|
|
|
|
|
73
|
|
|
else: |
|
74
|
1 |
|
callbacks = [] |
|
75
|
1 |
|
history = model.fit(X_train_sub, y_train_sub, |
|
76
|
|
|
epochs=nr_epochs, batch_size=20, |
|
77
|
|
|
# see comment on subsize_set |
|
78
|
|
|
validation_data=(X_val, y_val), |
|
79
|
|
|
verbose=verbose, |
|
80
|
|
|
callbacks=callbacks) |
|
81
|
1 |
|
histories.append(history) |
|
82
|
1 |
|
val_accuracies.append(history.history['val_acc'][-1]) |
|
83
|
1 |
|
val_losses.append(history.history['val_loss'][-1]) |
|
84
|
1 |
|
if outputfile is not None: |
|
85
|
|
|
store_train_hist_as_json(params, model_types, |
|
86
|
|
|
history.history, outputfile) |
|
87
|
1 |
|
return histories, val_accuracies, val_losses |
|
88
|
|
|
|
|
89
|
|
|
|
|
90
|
1 |
|
def store_train_hist_as_json(params, model_type, history, outputfile): |
|
91
|
|
|
""" |
|
92
|
|
|
This function stores the model parameters, the loss and accuracy history |
|
93
|
|
|
of one model in a JSON file. It appends the model information to the |
|
94
|
|
|
existing models in the file. |
|
95
|
|
|
|
|
96
|
|
|
Parameters |
|
97
|
|
|
---------- |
|
98
|
|
|
params : dict |
|
99
|
|
|
parameters for one model |
|
100
|
|
|
model_type : Keras model object |
|
101
|
|
|
Keras model object for one model |
|
102
|
|
|
history : dict |
|
103
|
|
|
training history from one model |
|
104
|
|
|
outputfile : str |
|
105
|
|
|
path where the json file needs to be stored |
|
106
|
|
|
""" |
|
107
|
1 |
|
jsondata = params.copy() |
|
108
|
1 |
|
for k in jsondata.keys(): |
|
109
|
1 |
|
if isinstance(jsondata[k], np.ndarray): |
|
110
|
1 |
|
jsondata[k] = jsondata[k].tolist() |
|
111
|
1 |
|
jsondata['train_acc'] = history['acc'] |
|
112
|
1 |
|
jsondata['train_loss'] = history['loss'] |
|
113
|
1 |
|
jsondata['val_acc'] = history['val_acc'] |
|
114
|
1 |
|
jsondata['val_loss'] = history['val_loss'] |
|
115
|
1 |
|
jsondata['modeltype'] = model_type |
|
116
|
1 |
|
jsondata['modeltype'] = model_type |
|
117
|
1 |
|
if os.path.isfile(outputfile): |
|
118
|
|
|
with open(outputfile, 'r') as outfile: |
|
119
|
|
|
previousdata = json.load(outfile) |
|
120
|
|
|
else: |
|
121
|
1 |
|
previousdata = [] |
|
122
|
1 |
|
previousdata.append(jsondata) |
|
123
|
1 |
|
with open(outputfile, 'w') as outfile: |
|
124
|
1 |
|
json.dump(previousdata, outfile, sort_keys=True, |
|
125
|
|
|
indent=4, ensure_ascii=False) |
|
126
|
|
|
|
|
127
|
|
|
|
|
128
|
1 |
|
def find_best_architecture(X_train, y_train, X_val, y_val, verbose=True, |
|
129
|
|
|
number_of_models=5, nr_epochs=5, subset_size=100, |
|
130
|
|
|
outputpath=None, **kwargs |
|
131
|
|
|
): |
|
132
|
|
|
""" |
|
133
|
|
|
Tries out a number of models on a subsample of the data, |
|
134
|
|
|
and outputs the best found architecture and hyperparameters. |
|
135
|
|
|
|
|
136
|
|
|
Parameters |
|
137
|
|
|
---------- |
|
138
|
|
|
X_train : numpy array |
|
139
|
|
|
The input dataset for training of shape |
|
140
|
|
|
(num_samples, num_timesteps, num_channels) |
|
141
|
|
|
y_train : numpy array |
|
142
|
|
|
The output classes for the train data, in binary format of shape |
|
143
|
|
|
(num_samples, num_classes) |
|
144
|
|
|
X_val : numpy array |
|
145
|
|
|
The input dataset for validation of shape |
|
146
|
|
|
(num_samples_val, num_timesteps, num_channels) |
|
147
|
|
|
y_val : numpy array |
|
148
|
|
|
The output classes for the validation data, in binary format of shape |
|
149
|
|
|
(num_samples_val, num_classes) |
|
150
|
|
|
verbose : bool, optional |
|
151
|
|
|
flag for displaying verbose output |
|
152
|
|
|
number_of_models : int, optiona |
|
153
|
|
|
The number of models to generate and test |
|
154
|
|
|
nr_epochs : int, optional |
|
155
|
|
|
The number of epochs that each model is trained |
|
156
|
|
|
subset_size : int, optional |
|
157
|
|
|
The size of the subset of the data that is used for finding |
|
158
|
|
|
the optimal architecture |
|
159
|
|
|
outputpath : str, optional |
|
160
|
|
|
File location to store the model results |
|
161
|
|
|
**kwargs: key-value parameters |
|
162
|
|
|
parameters for generating the models |
|
163
|
|
|
(see docstring for modelgen.generate_models) |
|
164
|
|
|
|
|
165
|
|
|
Returns |
|
166
|
|
|
---------- |
|
167
|
|
|
best_model : Keras model |
|
168
|
|
|
Best performing model, already trained on a small sample data set. |
|
169
|
|
|
best_params : dict |
|
170
|
|
|
Dictionary containing the hyperparameters for the best model |
|
171
|
|
|
best_model_type : str |
|
172
|
|
|
Type of the best model |
|
173
|
|
|
knn_acc : float |
|
174
|
|
|
accuaracy for kNN prediction on validation set |
|
175
|
|
|
""" |
|
176
|
1 |
|
models = modelgen.generate_models(X_train.shape, y_train.shape[1], |
|
177
|
|
|
number_of_models=number_of_models, |
|
178
|
|
|
**kwargs) |
|
179
|
1 |
|
histories, val_accuracies, val_losses = train_models_on_samples(X_train, |
|
180
|
|
|
y_train, |
|
181
|
|
|
X_val, |
|
182
|
|
|
y_val, |
|
183
|
|
|
models, |
|
184
|
|
|
nr_epochs, |
|
185
|
|
|
subset_size=subset_size, |
|
|
|
|
|
|
186
|
|
|
verbose=verbose, |
|
|
|
|
|
|
187
|
|
|
outputfile=outputpath) |
|
|
|
|
|
|
188
|
1 |
|
best_model_index = np.argmax(val_accuracies) |
|
189
|
1 |
|
best_model, best_params, best_model_type = models[best_model_index] |
|
190
|
1 |
|
knn_acc = kNN_accuracy( |
|
191
|
|
|
X_train[:subset_size, :, :], y_train[:subset_size, :], X_val, y_val) |
|
192
|
1 |
|
if verbose: |
|
193
|
|
|
print('Best model: model ', best_model_index) |
|
194
|
|
|
print('Model type: ', best_model_type) |
|
195
|
|
|
print('Hyperparameters: ', best_params) |
|
196
|
|
|
print('Accuracy on validation set: ', val_accuracies[best_model_index]) |
|
197
|
|
|
print('Accuracy of kNN on validation set', knn_acc) |
|
198
|
|
|
|
|
199
|
1 |
|
if val_accuracies[best_model_index] < knn_acc: |
|
200
|
|
|
warnings.warn('Best model not better than kNN: ' + |
|
201
|
|
|
str(val_accuracies[best_model_index]) + ' vs ' + |
|
202
|
|
|
str(knn_acc) |
|
203
|
|
|
) |
|
204
|
1 |
|
return best_model, best_params, best_model_type, knn_acc |
|
205
|
|
|
|
|
206
|
|
|
|
|
207
|
1 |
|
def kNN_accuracy(X_train, y_train, X_val, y_val, k=1): |
|
208
|
|
|
""" |
|
209
|
|
|
Performs k-Neigherst Neighbors and returns the accuracy score. |
|
210
|
|
|
|
|
211
|
|
|
Parameters |
|
212
|
|
|
---------- |
|
213
|
|
|
X_train : numpy array |
|
214
|
|
|
Train set of shape (num_samples, num_timesteps, num_channels) |
|
215
|
|
|
y_train : numpy array |
|
216
|
|
|
Class labels for train set |
|
217
|
|
|
X_val : numpy array |
|
218
|
|
|
Validation set of shape (num_samples, num_timesteps, num_channels) |
|
219
|
|
|
y_val : numpy array |
|
220
|
|
|
Class labels for validation set |
|
221
|
|
|
k : int |
|
222
|
|
|
number of neighbors to use for classifying |
|
223
|
|
|
|
|
224
|
|
|
Returns |
|
225
|
|
|
------- |
|
226
|
|
|
accuracy: float |
|
227
|
|
|
accuracy score on the validation set |
|
228
|
|
|
""" |
|
229
|
1 |
|
num_samples, num_timesteps, num_channels = X_train.shape |
|
230
|
1 |
|
clf = neighbors.KNeighborsClassifier(k) |
|
231
|
1 |
|
clf.fit( |
|
232
|
|
|
X_train.reshape( |
|
233
|
|
|
num_samples, |
|
234
|
|
|
num_timesteps * |
|
235
|
|
|
num_channels), |
|
236
|
|
|
y_train) |
|
237
|
1 |
|
num_samples, num_timesteps, num_channels = X_val.shape |
|
238
|
1 |
|
val_predict = clf.predict( |
|
239
|
|
|
X_val.reshape(num_samples, |
|
240
|
|
|
num_timesteps * num_channels)) |
|
241
|
|
|
return metrics.accuracy_score(val_predict, y_val) |
|
242
|
|
|
|
NLeSC /
mcfly
Loading history...
This check looks for lines that are too long. You can specify the maximum line length.