TD-BERT Implementation

TD-BERT implementation from songyouwei#147

data_utils.py

@@ -205,6 +205,8 @@ def __init__(self, fname, tokenizer):
                 'aspect_boundary': aspect_boundary,
                 'dependency_graph': dependency_graph,
                 'polarity': polarity,
+                'left_context_len': left_context_len,
+                'aspect_len': aspect_len,
             }
 
             all_data.append(data)

models/td_bert.py

@@ -0,0 +1,50 @@
+# -*- coding: utf-8 -*-
+# file: td_bert.py
+# author: xiangpan <[email protected]>
+# Copyright (C) 2020. All Rights Reserved.
+import torch
+import torch.nn as nn
+from layers.attention import Attention
+
+
+class TD_BERT(nn.Module):
+    def __init__(self, bert, opt):
+        super(TD_BERT, self).__init__()
+        self.bert = bert
+        self.dropout = nn.Dropout(opt.dropout)
+        self.opt = opt
+        self.dense = nn.Linear(opt.bert_dim, opt.polarities_dim)
+
+    def forward(self, inputs):
+        text_bert_indices, bert_segments_ids, left_context_len, aspect_len = (
+            inputs[0],
+            inputs[1],
+            inputs[2],
+            inputs[3],
+        )
+
+        encoded_layers, cls_output = self.bert(
+            text_bert_indices, bert_segments_ids
+        )
+
+
+        pooled_list = []
+        for i in range(0, encoded_layers.shape[0]):  # batch_size  i th batch
+            encoded_layers_i = encoded_layers[i]
+            left_context_len_i = left_context_len[i]
+            aspect_len_i = aspect_len[i]
+            e_list = []
+            if (left_context_len_i + 1) == (left_context_len_i + 1 + aspect_len_i):
+                e_list.append(encoded_layers_i[0])
+            for j in range(left_context_len_i + 1, left_context_len_i + 1 + aspect_len_i):
+                e_list.append(encoded_layers_i[j])
+            e = torch.stack(e_list, 0)
+            embed = torch.stack([e], 0)
+            pooled = nn.functional.max_pool2d(embed, (embed.size(1), 1)).squeeze(1)
+            pooled_list.append(pooled)
+        pooled_output = torch.cat(pooled_list)
+        pooled_output = self.dropout(pooled_output)
+
+        logits = self.dense(pooled_output)
+
+        return logits

train.py

@@ -26,6 +26,7 @@
 from models.bert_spc import BERT_SPC
 from models.albert_spc import ALBERT_SPC
 from models.roberta_spc import ROBERTA_SPC
+from models.td_bert import TD_BERT
 
 logger = logging.getLogger()
 logger.setLevel(logging.INFO)
@@ -249,6 +250,8 @@ def main():
         'lcf_bert': LCF_BERT,
         'albert_spc': ALBERT_SPC,
         'roberta_spc': ROBERTA_SPC,
+        'td_bert': TD_BERT,
+
         # default hyper-parameters for LCF-BERT model is as follws:
         # lr: 2e-5
         # l2: 1e-5
@@ -298,6 +301,7 @@ def main():
         'aen_bert': ['text_bert_indices', 'aspect_bert_indices'],
         'aen_bert': ['text_bert_indices', 'aspect_bert_indices'],
         'lcf_bert': ['concat_bert_indices', 'concat_segments_indices', 'text_bert_indices', 'aspect_bert_indices'],
+        'td_bert': ['text_bert_indices', 'bert_segments_ids','left_context_len','aspect_len'],
     }
     initializers = {
         'xavier_uniform_': torch.nn.init.xavier_uniform_,