-
Notifications
You must be signed in to change notification settings - Fork 2
/
feature_extraction.py
218 lines (184 loc) · 9.94 KB
/
feature_extraction.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
from model.base_models import model_class_map
import torch
import pickle
import pandas as pd
from torch.utils.data import DataLoader, Dataset
from transformers import AutoTokenizer
from utils.options import Options
def get_meld_data(split, cls_3=False):
input_dir = 'data/'
if split == 'train':
df = pd.read_csv(input_dir + "meld/processed/train_full.csv")
elif split == 'val':
df = pd.read_csv(input_dir + "meld/processed/val_full.csv")
elif split == 'test':
df = pd.read_csv(input_dir + "meld/processed/test_full.csv")
emotion_label_map = {'anger': 0, 'disgust': 1, 'fear': 2, 'joy': 3, 'neutral': 4, 'sadness': 5, 'surprise': 6}
sentiment_map = {'positive': 0, 'neutral': 1, 'negative': 2}
labels = []
for i, row in df.iterrows():
labels.append(sentiment_map[row['Sentiment']] if cls_3 else emotion_label_map[row['Emotion']])
return df['Utterance'].values, labels, df['Dialogue_ID'].values, df['Utterance_ID'].values
def get_daily_dialogue_dialog(split):
input_dir = 'data/'
emotion_map = {'no emotion': 0, 'anger': 1, 'disgust': 2, 'fear': 3, 'happiness': 4, 'sadness': 5,
'surprise': 6}
utterances = []
labels = []
dialogue_ids = []
utterance_ids = []
df = pd.read_csv(input_dir + f"daily_dialogue/{split}.csv")
for i, row in df.iterrows():
utterances.append(row['Utterance'])
labels.append(emotion_map[row['Emotion']])
dialogue_ids.append(row['Dialogue_ID'])
utterance_ids.append(row['Utterance_ID'])
return utterances, labels, dialogue_ids, utterance_ids
def get_emorynlp_data(split, cls_3=False):
input_dir = 'data/'
if split == 'train':
df = pd.read_csv(input_dir + "emorynlp/processed/train_full.csv")
elif split == 'val':
df = pd.read_csv(input_dir + "emorynlp/processed/val_full.csv")
elif split == 'test':
df = pd.read_csv(input_dir + "emorynlp/processed/test_full.csv")
emotion_label_map = {'Neutral': 0, 'Joyful': 1, 'Peaceful': 2, 'Powerful': 3,
'Scared': 4, 'Mad': 5, 'Sad': 6}
sentiment_map = {'positive': 0, 'neutral': 1, 'negative': 2}
labels = []
for i, row in df.iterrows():
labels.append(sentiment_map[row['Sentiment']] if cls_3 else emotion_label_map[row['Emotion']])
return df['Utterance'].values, labels, df['Dialogue_ID'].values, df['Utterance_ID'].values
class FeatureInferDataset(Dataset):
def __init__(self, utterances, emotions, dialog_ids, utterance_ids):
super().__init__()
self.utterances = utterances
self.emotions = emotions
self.dialog_id = dialog_ids
self.utterance_id = utterance_ids
def __len__(self):
return len(self.emotions)
def __getitem__(self, item):
return {"text": self.utterances[item], "dialog": self.dialog_id[item],
"utterance": self.utterance_id[item]}
class Collator(object):
def __init__(self, cfg):
self.cfg = cfg
self.tokenizer = AutoTokenizer.from_pretrained('roberta-large')
def __call__(self, batch):
inputs = self.tokenizer(
[ex['text'] for ex in batch],
max_length=self.cfg.max_len,
return_tensors='pt',
padding=True,
truncation=True,
return_offsets_mapping=False)
dialogs = torch.tensor([ex['dialog'] for ex in batch], dtype=torch.long)
utterances = torch.tensor([ex['utterance'] for ex in batch], dtype=torch.long)
return inputs, dialogs, utterances
def inference_fn(model, opt, ds, device):
collator = Collator(opt)
dataloader = DataLoader(ds,
batch_size=opt.batch_size,
shuffle=False,
collate_fn=collator,
num_workers=opt.num_workers,
pin_memory=False,
drop_last=False)
model.eval()
model.to(device)
features = []
dialogues = []
utterances = []
with torch.no_grad():
for step, (inputs, dialog, utterance) in enumerate(dataloader):
for k, v in inputs.items():
inputs[k] = v.to(device)
feature = model.feature(inputs)
features.extend(list(feature.detach().cpu().numpy()))
dialogues.extend(dialog.detach().cpu().numpy().tolist())
utterances.extend(utterance.detach().cpu().numpy().tolist())
return features, dialogues, utterances
def get_emb_dict(features, dialogues, utterances):
emb_dict = {}
for feature, dialogue, utterance in zip(features, dialogues, utterances):
emb_dict[str(dialogue)+'_'+str(utterance)] = feature
return emb_dict
def main():
options = Options()
options.add_extractor_options()
opt = options.parse()[0]
if opt.cls_3:
opt.target_size = 3
INPUT_DIR = 'data/ckpts/' + opt.name + '/'
model_state = torch.load(INPUT_DIR + f"{opt.model}.pth")
model = model_class_map[opt.model](opt)
model.load_state_dict(model_state)
print("Loaded model state from "+INPUT_DIR)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
if opt.dataset == 'meld':
utterances, labels, dialogue_ids, utterance_ids = get_meld_data('train', opt.cls_3)
train_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_meld_data('val', opt.cls_3)
val_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_meld_data('test', opt.cls_3)
test_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
OUTPUT_DIR = 'data/meld/processed/'
elif opt.dataset == 'daily_dialogue':
utterances, labels, dialogue_ids, utterance_ids = get_daily_dialogue_dialog('train')
train_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_daily_dialogue_dialog('val')
val_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_daily_dialogue_dialog('test')
test_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
OUTPUT_DIR = 'data/daily_dialogue/processed/'
elif opt.dataset == 'emorynlp':
utterances, labels, dialogue_ids, utterance_ids = get_emorynlp_data('train')
train_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_emorynlp_data('val')
val_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
utterances, labels, dialogue_ids, utterance_ids = get_emorynlp_data('test')
test_ds = FeatureInferDataset(utterances, labels, dialogue_ids, utterance_ids)
OUTPUT_DIR = 'data/emorynlp/processed/'
if opt.dataset == 'meld':
features, dialogues, utterances = inference_fn(model, opt, val_ds, device)
dev_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/meld/processed/dev_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/meld/processed/dev_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(dev_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, train_ds, device)
train_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/meld/processed/train_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/meld/processed/train_{opt.metric}.pkl',
'wb') as handle:
pickle.dump(train_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, test_ds, device)
test_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/meld/processed/test_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/meld/processed/test_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(test_emb_dict, handle)
elif opt.dataset == 'daily_dialogue':
features, dialogues, utterances = inference_fn(model, opt, val_ds, device)
dev_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(OUTPUT_DIR + f'dev_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(dev_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, train_ds, device)
train_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(OUTPUT_DIR + f'train_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(train_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, test_ds, device)
test_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(OUTPUT_DIR + f'test_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(test_emb_dict, handle)
elif opt.dataset == 'emorynlp':
features, dialogues, utterances = inference_fn(model, opt, val_ds, device)
dev_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/emorynlp/processed/dev_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/emorynlp/processed/dev_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(dev_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, train_ds, device)
train_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/emorynlp/processed/train_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/emorynlp/processed/train_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(train_emb_dict, handle)
features, dialogues, utterances = inference_fn(model, opt, test_ds, device)
test_emb_dict = get_emb_dict(features, dialogues, utterances)
with open(f'data/emorynlp/processed/test_cls3_{opt.metric}.pkl' if opt.cls_3 else f'data/emorynlp/processed/test_{opt.metric}.pkl', 'wb') as handle:
pickle.dump(test_emb_dict, handle)
if __name__ == '__main__':
main()