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 as sklearnmetrics |
16
|
1 |
|
import warnings |
17
|
1 |
|
import json |
18
|
1 |
|
import os |
19
|
1 |
|
from keras.callbacks import EarlyStopping |
20
|
1 |
|
from keras import metrics |
21
|
|
|
|
22
|
|
|
|
23
|
1 |
|
def train_models_on_samples(X_train, y_train, X_val, y_val, models, |
24
|
|
|
nr_epochs=5, subset_size=100, verbose=True, outputfile=None, |
|
|
|
|
25
|
|
|
model_path=None, early_stopping=False, |
26
|
|
|
batch_size=20, metric='accuracy'): |
27
|
|
|
""" |
28
|
|
|
Given a list of compiled models, this function trains |
29
|
|
|
them all on a subset of the train data. If the given size of the subset is |
30
|
|
|
smaller then the size of the data, the complete data set is used. |
31
|
|
|
|
32
|
|
|
Parameters |
33
|
|
|
---------- |
34
|
|
|
X_train : numpy array of shape (num_samples, num_timesteps, num_channels) |
35
|
|
|
The input dataset for training |
36
|
|
|
y_train : numpy array of shape (num_samples, num_classes) |
37
|
|
|
The output classes for the train data, in binary format |
38
|
|
|
X_val : numpy array of shape (num_samples_val, num_timesteps, num_channels) |
39
|
|
|
The input dataset for validation |
40
|
|
|
y_val : numpy array of shape (num_samples_val, num_classes) |
41
|
|
|
The output classes for the validation data, in binary format |
42
|
|
|
models : list of model, params, modeltypes |
43
|
|
|
List of keras models to train |
44
|
|
|
nr_epochs : int, optional |
45
|
|
|
nr of epochs to use for training one model |
46
|
|
|
subset_size : |
47
|
|
|
The number of samples used from the complete train set |
48
|
|
|
verbose : bool, optional |
49
|
|
|
flag for displaying verbose output |
50
|
|
|
outputfile: str, optional |
51
|
|
|
Filename to store the model training results |
52
|
|
|
model_path : str, optional |
53
|
|
|
Directory to store the models as HDF5 files |
54
|
|
|
early_stopping: bool |
55
|
|
|
Stop when validation loss does not decrease |
56
|
|
|
batch_size : int |
57
|
|
|
nr of samples per batch |
58
|
|
|
metric : str |
59
|
|
|
metric to store in the history object |
60
|
|
|
|
61
|
|
|
Returns |
62
|
|
|
---------- |
63
|
|
|
histories : list of Keras History objects |
64
|
|
|
train histories for all models |
65
|
|
|
val_metrics : list of floats |
66
|
|
|
validation accuraracies of the models |
67
|
|
|
val_losses : list of floats |
68
|
|
|
validation losses of the models |
69
|
|
|
""" |
70
|
|
|
# if subset_size is smaller then X_train, this will work fine |
71
|
1 |
|
X_train_sub = X_train[:subset_size, :, :] |
72
|
1 |
|
y_train_sub = y_train[:subset_size, :] |
73
|
|
|
|
74
|
1 |
|
metric_name = get_metric_name(metric) |
75
|
|
|
|
76
|
1 |
|
histories = [] |
77
|
1 |
|
val_metrics = [] |
78
|
1 |
|
val_losses = [] |
79
|
1 |
|
for i, (model, params, model_types) in enumerate(models): |
80
|
1 |
|
if verbose: |
81
|
|
|
print('Training model %d' % i, model_types) |
82
|
1 |
|
model_metrics = [get_metric_name(name) for name in model.metrics] |
83
|
1 |
|
if metric_name not in model_metrics: |
84
|
|
|
raise ValueError( |
85
|
|
|
'Invalid metric. The model was not compiled with {} as metric'.format(metric_name)) |
|
|
|
|
86
|
1 |
|
if early_stopping: |
87
|
|
|
callbacks = [ |
88
|
|
|
EarlyStopping(monitor='val_loss', patience=0, verbose=verbose, mode='auto')] |
|
|
|
|
89
|
|
|
else: |
90
|
1 |
|
callbacks = [] |
91
|
1 |
|
history = model.fit(X_train_sub, y_train_sub, |
92
|
|
|
epochs=nr_epochs, batch_size=batch_size, |
93
|
|
|
# see comment on subsize_set |
94
|
|
|
validation_data=(X_val, y_val), |
95
|
|
|
verbose=verbose, |
96
|
|
|
callbacks=callbacks) |
97
|
1 |
|
histories.append(history) |
98
|
|
|
|
99
|
1 |
|
val_metrics.append(history.history['val_' + metric_name][-1]) |
100
|
1 |
|
val_losses.append(history.history['val_loss'][-1]) |
101
|
1 |
|
if outputfile is not None: |
102
|
|
|
store_train_hist_as_json(params, model_types, |
103
|
|
|
history.history, outputfile) |
104
|
1 |
|
if model_path is not None: |
105
|
|
|
model.save(os.path.join(model_path, 'model_{}.h5'.format(i))) |
|
|
|
|
106
|
|
|
|
107
|
1 |
|
return histories, val_metrics, val_losses |
108
|
|
|
|
109
|
|
|
|
110
|
1 |
|
def store_train_hist_as_json(params, model_type, history, outputfile, metric_name='acc'): |
|
|
|
|
111
|
|
|
""" |
112
|
|
|
This function stores the model parameters, the loss and accuracy history |
113
|
|
|
of one model in a JSON file. It appends the model information to the |
114
|
|
|
existing models in the file. |
115
|
|
|
|
116
|
|
|
Parameters |
117
|
|
|
---------- |
118
|
|
|
params : dict |
119
|
|
|
parameters for one model |
120
|
|
|
model_type : Keras model object |
121
|
|
|
Keras model object for one model |
122
|
|
|
history : dict |
123
|
|
|
training history from one model |
124
|
|
|
outputfile : str |
125
|
|
|
path where the json file needs to be stored |
126
|
|
|
metric_name : str, optional |
127
|
|
|
name of metric from history to store |
128
|
|
|
""" |
129
|
1 |
|
jsondata = params.copy() |
130
|
1 |
|
for k in jsondata.keys(): |
131
|
1 |
|
if isinstance(jsondata[k], np.ndarray): |
132
|
1 |
|
jsondata[k] = jsondata[k].tolist() |
133
|
1 |
|
jsondata['train_metric'] = history[metric_name] |
134
|
1 |
|
jsondata['train_loss'] = history['loss'] |
135
|
1 |
|
jsondata['val_metric'] = history['val_' + metric_name] |
136
|
1 |
|
jsondata['val_loss'] = history['val_loss'] |
137
|
1 |
|
jsondata['modeltype'] = model_type |
138
|
1 |
|
jsondata['metric'] = metric_name |
139
|
1 |
|
if os.path.isfile(outputfile): |
140
|
|
|
with open(outputfile, 'r') as outfile: |
141
|
|
|
previousdata = json.load(outfile) |
142
|
|
|
else: |
143
|
1 |
|
previousdata = [] |
144
|
1 |
|
previousdata.append(jsondata) |
145
|
1 |
|
with open(outputfile, 'w') as outfile: |
146
|
1 |
|
json.dump(previousdata, outfile, sort_keys=True, |
147
|
|
|
indent=4, ensure_ascii=False) |
148
|
|
|
|
149
|
|
|
|
150
|
1 |
|
def find_best_architecture(X_train, y_train, X_val, y_val, verbose=True, |
151
|
|
|
number_of_models=5, nr_epochs=5, subset_size=100, |
152
|
|
|
outputpath=None, model_path=None, metric='accuracy', |
153
|
|
|
**kwargs): |
154
|
|
|
""" |
155
|
|
|
Tries out a number of models on a subsample of the data, |
156
|
|
|
and outputs the best found architecture and hyperparameters. |
157
|
|
|
|
158
|
|
|
Parameters |
159
|
|
|
---------- |
160
|
|
|
X_train : numpy array |
161
|
|
|
The input dataset for training of shape |
162
|
|
|
(num_samples, num_timesteps, num_channels) |
163
|
|
|
y_train : numpy array |
164
|
|
|
The output classes for the train data, in binary format of shape |
165
|
|
|
(num_samples, num_classes) |
166
|
|
|
X_val : numpy array |
167
|
|
|
The input dataset for validation of shape |
168
|
|
|
(num_samples_val, num_timesteps, num_channels) |
169
|
|
|
y_val : numpy array |
170
|
|
|
The output classes for the validation data, in binary format of shape |
171
|
|
|
(num_samples_val, num_classes) |
172
|
|
|
verbose : bool, optional |
173
|
|
|
flag for displaying verbose output |
174
|
|
|
number_of_models : int, optiona |
175
|
|
|
The number of models to generate and test |
176
|
|
|
nr_epochs : int, optional |
177
|
|
|
The number of epochs that each model is trained |
178
|
|
|
subset_size : int, optional |
179
|
|
|
The size of the subset of the data that is used for finding |
180
|
|
|
the optimal architecture |
181
|
|
|
outputpath : str, optional |
182
|
|
|
File location to store the model results |
183
|
|
|
model_path: str, optional |
184
|
|
|
Directory to save the models as HDF5 files |
185
|
|
|
metric: str, optional |
186
|
|
|
metric that is used to evaluate the model on the validation set. |
187
|
|
|
See https://keras.io/metrics/ for possible metrics |
188
|
|
|
**kwargs: key-value parameters |
189
|
|
|
parameters for generating the models |
190
|
|
|
(see docstring for modelgen.generate_models) |
191
|
|
|
|
192
|
|
|
Returns |
193
|
|
|
---------- |
194
|
|
|
best_model : Keras model |
195
|
|
|
Best performing model, already trained on a small sample data set. |
196
|
|
|
best_params : dict |
197
|
|
|
Dictionary containing the hyperparameters for the best model |
198
|
|
|
best_model_type : str |
199
|
|
|
Type of the best model |
200
|
|
|
knn_acc : float |
201
|
|
|
accuaracy for kNN prediction on validation set |
202
|
|
|
""" |
203
|
1 |
|
models = modelgen.generate_models(X_train.shape, y_train.shape[1], |
204
|
|
|
number_of_models=number_of_models, |
205
|
|
|
metrics=[metric], |
206
|
|
|
**kwargs) |
207
|
1 |
|
histories, val_accuracies, val_losses = train_models_on_samples(X_train, |
208
|
|
|
y_train, |
209
|
|
|
X_val, |
210
|
|
|
y_val, |
211
|
|
|
models, |
212
|
|
|
nr_epochs, |
213
|
|
|
subset_size=subset_size, |
|
|
|
|
214
|
|
|
verbose=verbose, |
|
|
|
|
215
|
|
|
outputfile=outputpath, |
|
|
|
|
216
|
|
|
model_path=model_path, |
|
|
|
|
217
|
|
|
metric=metric) |
|
|
|
|
218
|
1 |
|
best_model_index = np.argmax(val_accuracies) |
219
|
1 |
|
best_model, best_params, best_model_type = models[best_model_index] |
220
|
1 |
|
knn_acc = kNN_accuracy( |
221
|
|
|
X_train[:subset_size, :, :], y_train[:subset_size, :], X_val, y_val) |
222
|
1 |
|
if verbose: |
223
|
|
|
print('Best model: model ', best_model_index) |
224
|
|
|
print('Model type: ', best_model_type) |
225
|
|
|
print('Hyperparameters: ', best_params) |
226
|
|
|
print(str(metric) + ' on validation set: ', |
227
|
|
|
val_accuracies[best_model_index]) |
228
|
|
|
print('Accuracy of kNN on validation set', knn_acc) |
229
|
|
|
|
230
|
1 |
|
if val_accuracies[best_model_index] < knn_acc: |
231
|
|
|
warnings.warn('Best model not better than kNN: ' + |
232
|
|
|
str(val_accuracies[best_model_index]) + ' vs ' + |
233
|
|
|
str(knn_acc) |
234
|
|
|
) |
235
|
1 |
|
return best_model, best_params, best_model_type, knn_acc |
236
|
|
|
|
237
|
|
|
|
238
|
1 |
|
def get_metric_name(name): |
239
|
|
|
""" |
240
|
|
|
Gives the keras name for a metric |
241
|
|
|
|
242
|
|
|
Parameters |
243
|
|
|
---------- |
244
|
|
|
name : str |
245
|
|
|
original name of the metric |
246
|
|
|
Returns |
247
|
|
|
------- |
248
|
|
|
|
249
|
|
|
""" |
250
|
1 |
|
if name == 'acc' or name == 'accuracy': |
251
|
1 |
|
return 'acc' |
252
|
1 |
|
try: |
253
|
1 |
|
metric_fn = metrics.get(name) |
254
|
1 |
|
return metric_fn.__name__ |
255
|
|
|
except: |
256
|
|
|
pass |
257
|
|
|
return name |
258
|
|
|
|
259
|
|
|
|
260
|
1 |
|
def kNN_accuracy(X_train, y_train, X_val, y_val, k=1): |
261
|
|
|
""" |
262
|
|
|
Performs k-Neigherst Neighbors and returns the accuracy score. |
263
|
|
|
|
264
|
|
|
Parameters |
265
|
|
|
---------- |
266
|
|
|
X_train : numpy array |
267
|
|
|
Train set of shape (num_samples, num_timesteps, num_channels) |
268
|
|
|
y_train : numpy array |
269
|
|
|
Class labels for train set |
270
|
|
|
X_val : numpy array |
271
|
|
|
Validation set of shape (num_samples, num_timesteps, num_channels) |
272
|
|
|
y_val : numpy array |
273
|
|
|
Class labels for validation set |
274
|
|
|
k : int |
275
|
|
|
number of neighbors to use for classifying |
276
|
|
|
|
277
|
|
|
Returns |
278
|
|
|
------- |
279
|
|
|
accuracy: float |
280
|
|
|
accuracy score on the validation set |
281
|
|
|
""" |
282
|
1 |
|
num_samples, num_timesteps, num_channels = X_train.shape |
283
|
1 |
|
clf = neighbors.KNeighborsClassifier(k) |
284
|
1 |
|
clf.fit( |
285
|
|
|
X_train.reshape( |
286
|
|
|
num_samples, |
287
|
|
|
num_timesteps * |
288
|
|
|
num_channels), |
289
|
|
|
y_train) |
290
|
1 |
|
num_samples, num_timesteps, num_channels = X_val.shape |
291
|
1 |
|
val_predict = clf.predict( |
292
|
|
|
X_val.reshape(num_samples, |
293
|
|
|
num_timesteps * num_channels)) |
294
|
|
|
return sklearnmetrics.accuracy_score(val_predict, y_val) |
295
|
|
|
|
This check looks for lines that are too long. You can specify the maximum line length.