fix ncrf model

Ubuntu · Ubuntu · commit 1a9a1f3bf02c · 2018-12-25T11:38:27.000Z
diff --git a/modules/layers/decoders.py b/modules/layers/decoders.py
@@ -589,16 +589,25 @@ def forward(self, inputs, labels_mask):
         self.eval()
         logits, intent_output = self.forward_model(inputs)
         _, preds = self.crf._viterbi_decode_nbest(logits, labels_mask, self.nbest)
+        # print(preds.shape)
+        preds = preds[:, :, 0]
+        """for idx in range(len(preds)):
+            for idx_ in range(len(preds[0])):
+                if preds[idx][idx_] > 0:
+                    preds[idx][idx_] -= 1
+                else:
+                    raise"""
+        # print(preds)
         self.train()
         return preds, intent_output.argmax(-1)
 
     def score(self, inputs, labels_mask, labels, cls_ids):
         logits, intent_output = self.forward_model(inputs)
-        crf_score = self.crf.neg_log_likelihood_loss(logits, labels_mask, labels) / logits.shape[0]
+        crf_score = self.crf.neg_log_likelihood_loss(logits, labels_mask, labels) / logits.size(0)
         return crf_score + self.intent_loss(intent_output, cls_ids)
 
     @classmethod
     def create(cls, label_size, input_dim, intent_size, input_dropout=0.5, key_dim=64,
                val_dim=64, num_heads=3, use_cuda=True, nbest=8):
-        return cls(NCRF(label_size + 2, use_cuda), label_size, input_dim, intent_size, input_dropout,
+        return cls(NCRF(label_size, use_cuda), label_size + 2, input_dim, intent_size, input_dropout,
                    key_dim, val_dim, num_heads, nbest)
diff --git a/modules/layers/ncrf.py b/modules/layers/ncrf.py
@@ -3,6 +3,14 @@
 # @Date:   2017-12-04 23:19:38
 # @Last Modified by:   Jie Yang,     Contact: jieynlp@gmail.com
 # @Last Modified time: 2018-05-16 16:57:39
+# https://github.com/Das-Boot/NCRFpp
+# @inproceedings{yang2018ncrf,  
+#  title={NCRF++: An Open-source Neural Sequence Labeling Toolkit},  
+#  author={Yang, Jie and Zhang, Yue},  
+#  booktitle={Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics},
+#  Url = {https://arxiv.org/pdf/1806.05626.pdf},
+#  year={2018}  
+# }
 from __future__ import print_function
 import torch
 import torch.autograd as autograd
@@ -90,6 +98,7 @@ def _calculate_PZ(self, feats, mask):
             mask_idx = mask[idx, :].view(batch_size, 1).expand(batch_size, tag_size)
 
             ## effective updated partition part, only keep the partition value of mask value = 1
+            mask_idx = mask_idx.byte()
             masked_cur_partition = cur_partition.masked_select(mask_idx)
             ## let mask_idx broadcastable, to disable warning
             mask_idx = mask_idx.contiguous().view(batch_size, tag_size, 1)
@@ -240,6 +249,7 @@ def _score_sentence(self, scores, mask, tags):
         ### need convert tags id to search from 400 positions of scores
         tg_energy = torch.gather(scores.view(seq_len, batch_size, -1), 2, new_tags).view(seq_len, batch_size)  # seq_len * bat_size
         ## mask transpose to (seq_len, batch_size)
+        mask = mask.byte()
         tg_energy = tg_energy.masked_select(mask.transpose(1,0))
 
         # ## calculate the score from START_TAG to first label
diff --git a/modules/models/bert_models.py b/modules/models/bert_models.py
@@ -309,13 +309,15 @@ def create(cls,
                # Global params
                use_cuda=True,
                # Meta
-               meta_dim=None):
+               meta_dim=None,
+               # NCRFpp
+               nbest=8):
         embedder = BertEmbedder.create(
             bert_config_file, init_checkpoint_pt, embedding_dim, use_cuda, bert_mode, freeze)
         if meta_dim is None:
             encoder = BertBiLSTMEncoder.create(embedder, enc_hidden_dim, rnn_layers, use_cuda)
         else:
             encoder = BertMetaBiLSTMEncoder.create(embedder, meta_dim, enc_hidden_dim, rnn_layers, use_cuda)
         decoder = AttnNCRFJointDecoder.create(
-            label_size, encoder.output_dim, intent_size, input_dropout, key_dim, val_dim, num_heads, use_cuda)
+            label_size, encoder.output_dim, intent_size, input_dropout, key_dim, val_dim, num_heads, use_cuda, nbest=nbest)
         return cls(encoder, decoder, use_cuda)
