-
Notifications
You must be signed in to change notification settings - Fork 0
/
sentiment_model_trainer_mira.py
29 lines (22 loc) · 1.29 KB
/
sentiment_model_trainer_mira.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
from sentiment_model_trainer_base import SentimentModelTrainerBase
import numpy as np
class SentimentModelTrainerMIRA(SentimentModelTrainerBase):
def calc_constraints_for_update_step_on_batch(self, previous_w: np.ndarray, documents_batch: list,
feature_vectors_batch: list, inferred_labelings_batch: list):
nr_labelings = sum(len(labelings) for labelings in inferred_labelings_batch)
G = np.zeros((nr_labelings, self.features_extractor.nr_features))
L = []
next_G_line_idx = 0
for document, feature_vector_summed, inferred_labelings in zip(documents_batch,
feature_vectors_batch,
inferred_labelings_batch):
y = document.y()
y_fv = feature_vector_summed
for y_tag in inferred_labelings:
y_tag_fv = self.features_extractor.evaluate_document_feature_vector_summed(document, y_tag)
G[next_G_line_idx, :] = (y_fv - y_tag_fv)
next_G_line_idx += 1
y_tag_loss = self.calc_labeling_loss(y, y_tag)
L.append(y_tag_loss)
L = np.array(L).reshape(-1, )
return G, L