If I use transformers instead of pytorch_pretrained_bert, there are something error:

# aagcn_bert.py#56
# encoder_layer, pooled_output = self.bert(
#     text_bert_indices, token_type_ids=bert_segments_ids, output_all_encoded_layers=False)
 output = self.bert(
            text_bert_indices, token_type_ids=bert_segments_ids, return_dict=False)
 text_out = output[0]

And the error is:

Traceback (most recent call last):
  File "train_bert.py", line 338, in <module>
    main()
  File "train_bert.py", line 333, in main
    ins.run()
  File "train_bert.py", line 186, in run
    best_model_path = self._train(
  File "train_bert.py", line 106, in _train
    outputs = self.model(inputs)
  File "/home/liuwei1/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/liuwei1/AAGCN-ACSA/models/aagcn_bert.py", line 61, in forward
    x = F.relu(self.gc1(text_out, e_adj))
  File "/home/liuwei1/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/liuwei1/AAGCN-ACSA/models/aagcn_bert.py", line 25, in forward
    hidden = torch.matmul(text, self.weight)
RuntimeError: mat1 dim 1 must match mat2 dim 0

Looking forward to your help, thanks

I am a graduate student ,I am very interested in your work in sentiment analysis,can you provide your dataset for us,Thank you very much, my email is [email protected]

In section 3.2 and 3.3, you mention the Beta distribution and the edge weight of each word pair of the aspect-aware graph. However, I think the paramters in Equation (2) are not clear. Meanwhile, I found there is a PMI function in your code, rather than the beta distibution function. Could you give me more details ? Thanks.