| This model used hfl/chinese-roberta-wwm-ext-large backbone and was trained on SNLI, MNLI, DNLI, KvPI, OCNLI, CMNLI data in Chinese version. | |
| Model structures are as follows: | |
| ```python | |
| class RobertaForSequenceClassification(nn.Module): | |
| def __init__(self, tagset_size): | |
| super(RobertaForSequenceClassification, self).__init__() | |
| self.tagset_size = tagset_size | |
| self.roberta_single= AutoModel.from_pretrained(pretrain_model_dir) | |
| self.single_hidden2tag = RobertaClassificationHead(bert_hidden_dim, tagset_size) | |
| def forward(self, input_ids, input_mask): | |
| outputs_single = self.roberta_single(input_ids, input_mask, None) | |
| hidden_states_single = outputs_single[1]#torch.tanh(self.hidden_layer_2(torch.tanh(self.hidden_layer_1(outputs_single[1])))) #(batch, hidden) | |
| score_single = self.single_hidden2tag(hidden_states_single) #(batch, tag_set) | |
| return score_single | |
| class RobertaClassificationHead(nn.Module): | |
| def __init__(self, bert_hidden_dim, num_labels): | |
| super(RobertaClassificationHead, self).__init__() | |
| self.dense = nn.Linear(bert_hidden_dim, bert_hidden_dim) | |
| self.dropout = nn.Dropout(0.1) | |
| self.out_proj = nn.Linear(bert_hidden_dim, num_labels) | |
| def forward(self, features): | |
| x = features#[:, 0, :] # take <s> token (equiv. to [CLS]) | |
| x = self.dropout(x) | |
| x = self.dense(x) | |
| x = torch.tanh(x) | |
| x = self.dropout(x) | |
| x = self.out_proj(x) | |
| return x | |
| model = RobertaForSequenceClassification(num_labels) | |
| model.load_state_dict(torch.load(args.model_save_path+'Roberta_large_model.pt', map_location=device)) | |
| ``` |