-
Notifications
You must be signed in to change notification settings - Fork 269
/
models.py
224 lines (195 loc) · 7.5 KB
/
models.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
from .. import utils as U
from ..imports import *
from ..models import bn_drop_lin
MLP = "mlp"
TABULAR_MODELS = {
MLP: "a configurable multilayer perceptron with categorical variable embeddings [https://arxiv.org/abs/1604.06737]",
}
def print_tabular_classifiers():
for k, v in TABULAR_MODELS.items():
print("%s: %s" % (k, v))
def print_tabular_regression_models():
for k, v in TABULAR_MODELS.items():
print("%s: %s" % (k, v))
def _tabular_model(
name,
train_data,
multilabel=None,
is_regression=False,
metrics=None,
hidden_layers=[1000, 500],
hidden_dropouts=[0.0, 0.5],
bn=False,
verbose=1,
):
"""
```
Build and return a classification or regression model for tabular data
Args:
name (string): currently accepts 'mlp' for multilayer perceptron
train_data (TabularDataset): TabularDataset instance returned from one of the tabular_from_* functions
multilabel (bool): If True, multilabel model will be returned.
If false, binary/multiclass model will be returned.
If None, multilabel will be inferred from data.
is_regression(bool): If True, will build a regression model, else classification model.
metrics(list): List of metrics to use. If None: 'accuracy' is used for binar/multiclass,
'binary_accuracy' is used for multilabel classification, and 'mae' is used for regressio
hidden_layers(list): number of units in each hidden layer of NN
hidden_dropouts(list): Dropout values after each hidden layer of NN
bn(bool): If True, BatchNormalization will be used before each fully-connected layer in NN
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
```
"""
# check arguments
if not U.is_tabular_from_data(train_data):
err = """
Please pass training data in the form of data returned from a ktrain tabular_from* function.
"""
raise Exception(err)
if len(hidden_layers) != len(hidden_dropouts):
raise ValueError("len(hidden_layers) must equal len(hidden_dropouts)")
# reformat dropouts for each of construction
output_dropout = hidden_dropouts[1]
hidden_dropouts[1] = hidden_dropouts[0]
hidden_dropouts[0] = 0.0
# set model configuration values
if is_regression: # regression
if metrics is None or metrics == ["accuracy"]:
metrics = ["mae"]
num_classes = 1
multilabel = False
loss_func = "mse"
activation = "linear"
else: # classification
# set metrics
if multilabel and metrics is None:
metrics = ["binary_accuracy"]
elif metrics is None:
metrics = ["accuracy"]
# set number of classes and multilabel flag
num_classes = U.nclasses_from_data(train_data)
# determine multilabel
if multilabel is None:
multilabel = U.is_multilabel(train_data)
U.vprint("Is Multi-Label? %s" % (multilabel), verbose=verbose)
# set loss and activations
loss_func = "categorical_crossentropy"
activation = "softmax"
if multilabel:
loss_func = "binary_crossentropy"
activation = "sigmoid"
# construct model
ilayers = []
n_cat = len(train_data.cat_columns)
n_cont = len(train_data.cont_columns)
if n_cat == 0 and n_cont == 0:
raise ValueError("There are zero continuous and cateorical variables.")
# categorical inputs and embeddings
if n_cat > 0:
emblayers = []
num_uniques = [
max(c.cat.codes.values + 1) + 1
for n, c in train_data.df[train_data.cat_columns].items()
]
for i in range(n_cat):
inp = keras.layers.Input(shape=(1,))
ilayers.append(inp)
emb_size = min(50, (num_uniques[i] // 2) + 1)
# emb_size = min(600, round(1.6 * num_uniques[i]**0.56))
emb = keras.layers.Embedding(num_uniques[i], emb_size, input_length=1)(inp)
emblayers.append(emb)
x = keras.layers.concatenate(emblayers) if len(emblayers) > 1 else emblayers[0]
x = keras.layers.Flatten()(x)
# continuous inputs
if n_cont > 0:
x_cont = keras.layers.Input(shape=(n_cont,))
ilayers.append(x_cont)
x = keras.layers.concatenate([x, x_cont]) if n_cat > 0 else x_cont
# hidden layers
output = x
for i, n_out in enumerate(hidden_layers):
output = bn_drop_lin(output, n_out, bn=bn, p=hidden_dropouts[i], actn="relu")
# output layer
output = bn_drop_lin(output, num_classes, bn=bn, p=output_dropout, actn=activation)
# construct and compile model
model = keras.Model(inputs=ilayers, outputs=output)
model.compile(optimizer=U.DEFAULT_OPT, loss=loss_func, metrics=metrics)
U.vprint("done.", verbose=verbose)
return model
def tabular_classifier(
name,
train_data,
multilabel=None,
metrics=None,
hidden_layers=[1000, 500],
hidden_dropouts=[0.0, 0.5],
bn=False,
verbose=1,
):
"""
```
Build and return a classification model for tabular data
Args:
name (string): currently accepts 'mlp' for multilayer perceptron
train_data (TabularDataset): TabularDataset instance returned from one of the tabular_from_* functions
multilabel (bool): If True, multilabel model will be returned.
If false, binary/multiclass model will be returned.
If None, multilabel will be inferred from data.
metrics(list): List of metrics to use. If None: 'accuracy' is used for binar/multiclass,
'binary_accuracy' is used for multilabel classification, and 'mae' is used for regressio
hidden_layers(list): number of units in each hidden layer of NN
hidden_dropouts(list): Dropout values after each hidden layer of NN
bn(bool): If True, BatchNormalization will be used before each fully-connected layer in NN
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
```
"""
return _tabular_model(
name,
train_data,
multilabel=multilabel,
metrics=metrics,
hidden_layers=hidden_layers,
hidden_dropouts=hidden_dropouts,
bn=bn,
verbose=verbose,
is_regression=False,
)
def tabular_regression_model(
name,
train_data,
metrics=["mae"],
hidden_layers=[1000, 500],
hidden_dropouts=[0.0, 0.5],
bn=False,
verbose=1,
):
"""
```
Build and return a regression model for tabular data
Args:
name (string): currently accepts 'mlp' for multilayer perceptron
train_data (TabularDataset): TabularDataset instance returned from one of the tabular_from_* functions
metrics(list): list of metrics to use
hidden_layers(list): number of units in each hidden layer of NN
hidden_dropouts(list): Dropout values after each hidden layer of NN
bn(bool): If True, BatchNormalization will be before used each fully-connected layer in NN
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
```
"""
return _tabular_model(
name,
train_data,
multilabel=None,
metrics=metrics,
hidden_layers=hidden_layers,
hidden_dropouts=hidden_dropouts,
bn=bn,
verbose=verbose,
is_regression=True,
)