|
1
|
|
|
''' |
|
2
|
|
|
Summary: |
|
3
|
|
|
Function generate_models from modelgen.py generates and compiles models |
|
4
|
|
|
Function train_models_on_samples trains those models |
|
5
|
|
|
Function plotTrainingProcess plots the training process |
|
6
|
|
|
Function find_best_architecture is wrapper function that combines |
|
7
|
|
|
these steps |
|
8
|
|
|
Example function calls in 'EvaluateDifferentModels.ipynb' |
|
9
|
|
|
''' |
|
10
|
1 |
|
import numpy as np |
|
11
|
1 |
|
from matplotlib import pyplot as plt |
|
12
|
1 |
|
from . import modelgen |
|
13
|
1 |
|
from sklearn import neighbors, metrics |
|
14
|
1 |
|
import warnings |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
1 |
|
def train_models_on_samples(X_train, y_train, X_val, y_val, models, |
|
|
|
|
|
|
18
|
|
|
nr_epochs=5, subsize_set=100, verbose=True): |
|
19
|
|
|
''' |
|
20
|
|
|
Given a list of compiled models, this function trains |
|
21
|
|
|
them all on a subset of the train data. If the given size of the subset is |
|
22
|
|
|
smaller then the size of the data, the complete data set is used. |
|
23
|
|
|
Parameters |
|
24
|
|
|
---------- |
|
25
|
|
|
X_train : numpy array of shape (num_samples, num_timesteps, num_channels) |
|
26
|
|
|
The input dataset for training |
|
27
|
|
|
y_train : numpy array of shape (num_samples, num_classes) |
|
28
|
|
|
The output classes for the train data, in binary format |
|
29
|
|
|
X_val : numpy array of shape (num_samples_val, num_timesteps, num_channels) |
|
30
|
|
|
The input dataset for validation |
|
31
|
|
|
y_val : numpy array of shape (num_samples_val, num_classes) |
|
32
|
|
|
The output classes for the validation data, in binary format |
|
33
|
|
|
models : list of model, params, modeltypes |
|
34
|
|
|
List of keras models to train |
|
35
|
|
|
nr_epochs : int, optional |
|
36
|
|
|
nr of epochs to use for training one model |
|
37
|
|
|
subsize_set : int, optional |
|
38
|
|
|
number of samples to use from the training set for training these models |
|
|
|
|
|
|
39
|
|
|
verbose : bool, optional |
|
40
|
|
|
flag for displaying verbose output |
|
41
|
|
|
|
|
42
|
|
|
Returns |
|
43
|
|
|
---------- |
|
44
|
|
|
histories : list of Keras History objects |
|
45
|
|
|
train histories for all models |
|
46
|
|
|
val_accuracies : list of floats |
|
47
|
|
|
validation accuraracies of the models |
|
48
|
|
|
val_losses : list of floats |
|
49
|
|
|
validation losses of the models |
|
50
|
|
|
''' |
|
51
|
|
|
# if subset_size is smaller then X_train, this will work fine |
|
52
|
1 |
|
X_train_sub = X_train[:subsize_set, :, :] |
|
|
|
|
|
|
53
|
1 |
|
y_train_sub = y_train[:subsize_set, :] |
|
54
|
|
|
|
|
55
|
1 |
|
histories = [] |
|
56
|
1 |
|
val_accuracies = [] |
|
57
|
1 |
|
val_losses = [] |
|
58
|
1 |
|
for model, params, model_types in models: |
|
59
|
1 |
|
history = model.fit(X_train_sub, y_train_sub, |
|
60
|
|
|
nb_epoch=nr_epochs, batch_size=20, # see comment on subsize_set |
|
|
|
|
|
|
61
|
|
|
validation_data=(X_val, y_val), |
|
62
|
|
|
verbose=verbose) |
|
63
|
1 |
|
histories.append(history) |
|
64
|
1 |
|
val_accuracies.append(history.history['val_acc'][-1]) |
|
65
|
1 |
|
val_losses.append(history.history['val_loss'][-1]) |
|
66
|
|
|
|
|
67
|
1 |
|
return histories, val_accuracies, val_losses |
|
68
|
|
|
|
|
69
|
|
|
|
|
70
|
1 |
|
def plotTrainingProcess(history, name='Model', ax=None): |
|
|
|
|
|
|
71
|
|
|
''' |
|
72
|
|
|
This function plots the loss and accuracy on the train and validation set, |
|
73
|
|
|
for each epoch in the history of one model. |
|
74
|
|
|
|
|
75
|
|
|
Parameters |
|
76
|
|
|
---------- |
|
77
|
|
|
history : keras History object for one model |
|
78
|
|
|
The history object of the training process corresponding to one model |
|
79
|
|
|
Returns |
|
80
|
|
|
---------- |
|
81
|
|
|
|
|
82
|
|
|
''' |
|
83
|
|
|
if ax is None: |
|
84
|
|
|
fig, ax = plt.subplots() |
|
85
|
|
|
ax2 = ax.twinx() |
|
86
|
|
|
LN = len(history.history['val_loss']) |
|
|
|
|
|
|
87
|
|
|
val_loss, = ax.plot(range(LN), history.history['val_loss'], 'g--', |
|
88
|
|
|
label='validation loss') |
|
89
|
|
|
train_loss, = ax.plot(range(LN), history.history['loss'], 'g-', |
|
90
|
|
|
label='train loss') |
|
91
|
|
|
val_acc, = ax2.plot(range(LN), history.history['val_acc'], 'b--', |
|
92
|
|
|
label='validation accuracy') |
|
93
|
|
|
train_acc, = ax2.plot(range(LN), history.history['acc'], 'b-', |
|
94
|
|
|
label='train accuracy') |
|
95
|
|
|
ax.set_xlabel('epoch') |
|
96
|
|
|
ax.set_ylabel('loss', color='g') |
|
97
|
|
|
ax2.set_ylabel('accuracy', color='b') |
|
98
|
|
|
plt.legend(handles=[val_loss, train_loss, val_acc, train_acc], |
|
99
|
|
|
loc=2, bbox_to_anchor=(1.1, 1)) |
|
100
|
|
|
plt.title(name) |
|
101
|
|
|
|
|
102
|
|
|
|
|
103
|
1 |
|
def find_best_architecture(X_train, y_train, X_val, y_val, verbose=True, |
|
|
|
|
|
|
104
|
|
|
number_of_models=5, nr_epochs=5, **kwargs): |
|
105
|
|
|
''' |
|
106
|
|
|
Tries out a number of models on a subsample of the data, |
|
107
|
|
|
and outputs the best found architecture and hyperparameters. |
|
108
|
|
|
|
|
109
|
|
|
Parameters |
|
110
|
|
|
---------- |
|
111
|
|
|
X_train : numpy array of shape (num_samples, num_timesteps, num_channels) |
|
112
|
|
|
The input dataset for training |
|
113
|
|
|
y_train : numpy array of shape (num_samples, num_classes) |
|
114
|
|
|
The output classes for the train data, in binary format |
|
115
|
|
|
X_val : numpy array of shape (num_samples_val, num_timesteps, num_channels) |
|
116
|
|
|
The input dataset for validation |
|
117
|
|
|
y_val : numpy array of shape (num_samples_val, num_classes) |
|
118
|
|
|
The output classes for the validation data, in binary format |
|
119
|
|
|
verbose : bool, optional |
|
120
|
|
|
flag for displaying verbose output |
|
121
|
|
|
**kwargs: key-value parameters |
|
122
|
|
|
parameters for generating the models |
|
123
|
|
|
|
|
124
|
|
|
Returns |
|
125
|
|
|
---------- |
|
126
|
|
|
best_model : Keras model |
|
127
|
|
|
Best performing model, already trained on a small sample data set. |
|
128
|
|
|
best_params : dict |
|
129
|
|
|
Dictionary containing the hyperparameters for the best model |
|
130
|
|
|
best_model_type : str |
|
131
|
|
|
Type of the best model |
|
132
|
|
|
knn_acc : float |
|
133
|
|
|
accuaracy for kNN prediction on validation set |
|
134
|
|
|
''' |
|
135
|
1 |
|
models = modelgen.generate_models(X_train.shape, y_train.shape[1], |
|
136
|
|
|
number_of_models=number_of_models, |
|
137
|
|
|
**kwargs) |
|
138
|
1 |
|
subsize_set = 100 |
|
139
|
1 |
|
histories, val_accuracies, val_losses = train_models_on_samples(X_train, |
|
140
|
|
|
y_train, |
|
141
|
|
|
X_val, |
|
142
|
|
|
y_val, |
|
143
|
|
|
models, |
|
144
|
|
|
nr_epochs, |
|
145
|
|
|
subsize_set=subsize_set, |
|
|
|
|
|
|
146
|
|
|
verbose=verbose) |
|
|
|
|
|
|
147
|
1 |
|
best_model_index = np.argmax(val_accuracies) |
|
148
|
1 |
|
best_model, best_params, best_model_type = models[best_model_index] |
|
149
|
1 |
|
knn_acc = kNN_accuracy(X_train[:subsize_set, :, :], y_train[:subsize_set, :], X_val, y_val) |
|
|
|
|
|
|
150
|
1 |
|
if verbose: |
|
151
|
|
|
for i in range(len(models)): #<= now one plot per model, ultimately we |
|
152
|
|
|
# may want all models in one plot to allow for direct comparison |
|
153
|
|
|
name = str(models[i][1]) |
|
154
|
|
|
plotTrainingProcess(histories[i], name) |
|
155
|
|
|
print('Best model: model ', best_model_index) |
|
156
|
|
|
print('Model type: ', best_model_type) |
|
157
|
|
|
print('Hyperparameters: ', best_params) |
|
158
|
|
|
print('Accuracy on validation set: ', val_accuracies[best_model_index]) |
|
159
|
|
|
print('Accuracy of kNN on validation set', knn_acc) |
|
160
|
|
|
|
|
161
|
1 |
|
if val_accuracies[best_model_index] < knn_acc: |
|
162
|
|
|
warnings.warn('Best model not beter than kNN: ' + |
|
163
|
|
|
str(val_accuracies[best_model_index]) + ' vs ' + |
|
164
|
|
|
str(knn_acc) |
|
165
|
|
|
) |
|
166
|
1 |
|
return best_model, best_params, best_model_type, knn_acc |
|
167
|
|
|
|
|
168
|
|
|
|
|
169
|
1 |
|
def kNN_accuracy(X_train, y_train, X_val, y_val, k=1): |
|
|
|
|
|
|
170
|
1 |
|
num_samples, num_timesteps, num_channels = X_train.shape |
|
171
|
1 |
|
clf = neighbors.KNeighborsClassifier(k) |
|
172
|
1 |
|
clf.fit(X_train.reshape(num_samples, num_timesteps*num_channels), y_train) |
|
173
|
1 |
|
num_samples, num_timesteps, num_channels = X_val.shape |
|
174
|
1 |
|
val_predict = clf.predict(X_val.reshape(num_samples, num_timesteps*num_channels)) |
|
|
|
|
|
|
175
|
|
|
return metrics.accuracy_score(val_predict, y_val) |
|
176
|
|
|
|
NLeSC /
mcfly
Loading history...
This check looks for invalid names for a range of different identifiers.
You can set regular expressions to which the identifiers must conform if the defaults do not match your requirements.
If your project includes a Pylint configuration file, the settings contained in that file take precedence.
To find out more about Pylint, please refer to their site.