def process_input(data):
"""
Input is in the form of tuples of (class:int, sent:string)
"""
return [(prepare_vec_sequence(wordpunct_tokenize(sent),
word_to_vec,
SENTENCE_DIM,
output='tensor'), label)
for sent, label in data]
def predict(model, input_data, k=1):
with torch.no_grad():
result = []
for sentence in input_data:
tokens_in = wordpunct_tokenize(sentence)
sentence_in = prepare_vec_sequence(tokens_in, word_to_vec, SENTENCE_DIM, output='variable')
scores = F.softmax(model(sentence_in), dim=-1)
topk_scores = torch.topk(scores, k)
result.append(topk_scores)
return result
def neg_log_likelihood(self, sentence, tags):
word_embeds = prepare_vec_sequence(sentence,
word_to_vec,
output='variable')
if self.is_cuda:
word_embeds = word_embeds.cuda()
char_embeds = self.word_encoder(sentence).unsqueeze(1)
sentence_in = torch.cat((word_embeds, char_embeds), dim=-1)
sentence_in = self.dropout(sentence_in)
feats = self._get_lstm_features(sentence_in)
forward_score = self._forward_alg(feats)
gold_score = self._score_sentence(feats, tags)
return forward_score - gold_score
def forward(self, sentence): # dont confuse this with _forward_alg above.
word_embeds = prepare_vec_sequence(sentence,
word_to_vec,
output='variable')
if self.is_cuda:
word_embeds = word_embeds.cuda()
char_embeds = self.word_encoder(sentence).unsqueeze(1)
sentence_in = torch.cat((word_embeds, char_embeds), dim=-1)
# Get the emission scores from the BiLSTM
lstm_feats = self._get_lstm_features(sentence_in)
# Find the best path, given the features.
score, tag_seq = self._viterbi_decode(lstm_feats)
return score, tag_seq