-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
226 lines (178 loc) · 12.8 KB
/
main.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 sklearn.preprocessing import MinMaxScaler
import argparse
import glob
import os
from random import seed
from datetime import datetime
import numpy as np
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
from keras.utils import to_categorical
# own modules
from config import get_cnn_config, get_svm_config, get_crnn_config, get_lstm_config
from utils import load_data_object, create_folders, to_hot, X_reshape, parameter_str
from measures import f1_all_folds, recall_all_folds, export_results, export_ys
import model_training, network_utils
from data import Data
## global
SEED = 23
seed(SEED)
## interface
parser = argparse.ArgumentParser(description='Prepare feature .pkl and run experiments')
parser.add_argument('-f','--feature_type', type=str, dest='feature_type', action='store', default='compare',
help='specify the type of features you want to use')
parser.add_argument('-l','--label_type', type=str, dest='label_type', action='store', default='point',
help='specify the type of label you want to use')
parser.add_argument('-m', '--model_type', type=str, dest='model_type', action='store', default='svm',
help='name of model type')
parser.add_argument('-n','--experiment_name', type=str, dest='experiment_name', action='store', default='test',
help='name of experiment')
parser.add_argument('-g','--gender', dest='gender', nargs='+', default=['all','m','w'],
help='gender of data: all m w')
parser.add_argument('--verbose', dest='verbose', action='store_true', default=False,
help='prints more output information if true')
args = parser.parse_args()
if __name__ == "__main__":
if args.label_type == 'point':
gender = args.gender
else:
gender = ['gender_pred_only']
for g in gender:
if args.model_type == 'lstm':
config = get_lstm_config(model_type = args.model_type
, experiment_name = args.experiment_name+'_'+g
, feature_type = args.feature_type)
elif 'crnn' in args.model_type:
config = get_crnn_config(model_type = args.model_type
, experiment_name = args.experiment_name+'_'+g
, feature_type = args.feature_type)
elif args.model_type == 'svm':
config = get_svm_config(model_type = args.model_type
, experiment_name = args.experiment_name+'_'+g
, feature_type = args.feature_type)
# Some other architectures we experimented with but did not make it into the paper
elif args.model_type == 'cnn':
config = get_cnn_config(model_type = args.model_type
, experiment_name = args.experiment_name+'_'+g
, feature_type = args.feature_type)
elif args.model_type == 'cnn_end':
config = get_crnn_config(model_type = args.model_type
, experiment_name = args.experiment_name+'_'+g
, feature_type = args.feature_type)
else:
print("No config for model {} found".format(args.model_type))
exit()
create_folders(config)
data_obj = load_data_object(config, g)
y_pred_folds = {}
y_devel_folds = {}
for parameter in config['parameter_list']:
parameter_text = parameter_str(parameter)
print('[run] ', parameter_text)
y_pred_folds[parameter_text] = []
y_devel_folds[parameter_text] = []
start = datetime.now()
# train and evaluate one model for each fold - 4 splits for training / one hold out
for fold_no in range(len(data_obj.Xs_train)):
print(len(data_obj.Xs_train))
X_train, X_devel = data_obj.Xs_train[fold_no], data_obj.Xs_val[fold_no]
######
# ### select model and data "after care" depending on model
if config['model_type'] == 'lstm':
if args.verbose:
print("model specific data preparation")
print(X_train.shape)
input_shape = (X_train.shape[1], X_train.shape[2])
y_train, y_devel = to_categorical(data_obj.ys_train[fold_no], num_classes=config['num_labels']), to_categorical(data_obj.ys_val[fold_no], num_classes=config['num_labels'])
model = network_utils.create_lstm_model(config, parameter['learning_rate'],input_shape,config['num_labels'])
y_devel_folds[parameter_text], y_pred_folds[parameter_text] = model_training.train_network(config, fold_no
, model, parameter
, X_train , y_train
, X_devel, y_devel
, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
elif config['model_type'] == 'cnn':
if args.verbose:
print("model specific data preparation")
y_train, y_devel = to_categorical(data_obj.ys_train[fold_no], num_classes=config['num_labels']), to_categorical(data_obj.ys_val[fold_no], num_classes=config['num_labels'])
X_train, input_shape = X_reshape(X_train, config['num_channels'])
X_devel, _ = X_reshape(X_devel, config['num_channels'])
if args.verbose:
print("X_train: ", X_train.shape, ", X_devel: ", X_devel.shape)
print("Y_train: ", y_train.shape, ", Y_devel ", y_devel.shape)
model = network_utils.create_cnn_model(parameter['learning_rate'],input_shape,config['num_labels'])
y_devel_folds[parameter_text], y_pred_folds[parameter_text] = model_training.train_network(config, fold_no
, model, parameter
, X_train , y_train
, X_devel, y_devel
, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
elif config['model_type'] == 'cnn_end':
if args.verbose:
print("model specific data preparation")
y_train, y_devel = to_categorical(data_obj.ys_train[fold_no], num_classes=config['num_labels']), to_categorical(data_obj.ys_val[fold_no], num_classes=config['num_labels'])
if args.verbose:
print("X_train: ", X_train.shape, ", X_devel: ", X_devel.shape)
print("Y_train: ", y_train.shape, ", Y_devel ", y_devel.shape)
X_train, input_shape = X_reshape(X_train, config['num_channels'])
X_devel, _ = X_reshape(X_devel, config['num_channels'])
model = network_utils.create_cnn_end2end_model(config, parameter['learning_rate'],input_shape,config['num_labels'])
y_devel_folds[parameter_text], y_pred_folds[parameter_text] = model_training.train_network(config, fold_no
, model, parameter
, X_train , y_train
, X_devel, y_devel
, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
elif 'crnn' in config['model_type']:
if args.verbose:
print("model specific data preparation")
y_train, y_devel = to_categorical(data_obj.ys_train[fold_no], num_classes=config['num_labels']), to_categorical(data_obj.ys_val[fold_no], num_classes=config['num_labels'])
if args.verbose:
print("X_train: ", X_train.shape, ", X_devel: ", X_devel.shape)
print("Y_train: ", y_train.shape, ", Y_devel ", y_devel.shape)
if config['model_type'] == 'crnn':
X_train, input_shape = X_reshape(X_train, config['num_channels'])
X_devel, _ = X_reshape(X_devel, config['num_channels'])
model = network_utils.create_crnn_small_model(parameter['learning_rate'],input_shape,config['num_labels'])
elif config['model_type'] == 'crnn_end':
#X_train = X_train_ #.reshape(X_train_.shape)
#X_devel = X_devel_#.reshape(1, X_devel_.shape)
input_shape = (X_train.shape[1], X_train.shape[2])
print(input_shape)
model = network_utils.create_crnn_end2end_model(config, parameter['learning_rate'],input_shape,config['num_labels'])
y_devel_folds[parameter_text], y_pred_folds[parameter_text] = model_training.train_network(config, fold_no
, model, parameter
, X_train , y_train
, X_devel, y_devel
, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
del model
elif config['model_type'] == 'svm':
y_train, y_devel = data_obj.ys_train[fold_no], data_obj.ys_val[fold_no]
if X_train.ndim > 2 and X_train.shape[1] > 1: #features, we have to get rid off the time dimension
if config['svm_seq_agg'] == 'mean':
X_train = np.mean(X_train, axis=1)
X_devel = np.mean(X_devel, axis=1)
elif config['svm_seq_agg'] == 'middle':
X_train = X_train[:,int(X_train.shape[1]/2),:]
X_devel = X_devel[:,int(X_devel.shape[1]/2),:]
elif config['svm_seq_agg'] == 'flatten':
X_train = X_train.reshape(X_train.shape[0], X_train.shape[1] * X_train.shape[2])
X_devel = X_devel.reshape(X_devel.shape[0], X_devel.shape[1] * X_devel.shape[2])
print('SVM X_train input format: ', X_train.shape)
new_X_train = X_train.reshape(X_train.shape[0], X_train.shape[-1])
new_X_devel = X_devel.reshape(X_devel.shape[0], X_devel.shape[-1])
scaler = MinMaxScaler()
X_train = scaler.fit_transform(new_X_train)
X_devel = scaler.transform(new_X_devel)
if args.verbose:
print("X_train: ", X_train.shape, ", X_devel: ", X_devel.shape)
print("Y_train: ", y_train.shape, ", Y_devel ", y_devel.shape)
y_devel_folds[parameter_text], y_pred_folds[parameter_text] = model_training.train_svm(config, fold_no, parameter
, X_train, y_train
, X_devel, y_devel
, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
else:
print("Model {} not defined".format(config['model_type']))
exit()
# average results of all fold experiments
f1_all_folds(y_devel_folds[parameter_text], y_pred_folds[parameter_text])
# average results of all fold experiments
recall_all_folds(y_devel_folds[parameter_text], y_pred_folds[parameter_text])
export_ys(config, parameter, y_devel_folds[parameter_text], y_pred_folds[parameter_text])
export_results(config, config['parameter_list'], y_devel_folds, y_pred_folds)