Class ANCE

src/eval/models.py:11–55

# ANCE model
class ANCE(RobertaForSequenceClassification):
    # class Pooler:   # adapt to DPR
    #     def __init__(self, pooler_output):
    #         self.pooler_output = pooler_output

    def __init__(self, config):
        RobertaForSequenceClassification.__init__(self, config)
        self.embeddingHead = nn.Linear(config.hidden_size, 768)
        self.norm = nn.LayerNorm(768)
        self.apply(self._init_weights)
        self.use_mean = False
    
    def _init_weights(self, module):
        """ Initialize the weights """
        if isinstance(module, (nn.Linear, nn.Embedding, nn.Conv1d)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=0.02)

    def query_emb(self, input_ids, attention_mask):
        outputs1 = self.roberta(input_ids=input_ids,
                                attention_mask=attention_mask)
        outputs1 = outputs1.last_hidden_state
        full_emb = self.masked_mean_or_first(outputs1, attention_mask)
        query1 = self.norm(self.embeddingHead(full_emb))
        return query1


    def doc_emb(self, input_ids, attention_mask):
        return self.query_emb(input_ids, attention_mask)
    

    def masked_mean_or_first(self, emb_all, mask):
        if self.use_mean:
            return self.masked_mean(emb_all, mask)
        else:
            return emb_all[:, 0]
    
    def masked_mean(self, t, mask):
        s = torch.sum(t * mask.unsqueeze(-1).float(), axis=1)
        d = mask.sum(axis=1, keepdim=True).float()
        return s / d
    
    def forward(self, input_ids, attention_mask, wrap_pooler=False):
        return self.query_emb(input_ids, attention_mask)
      
      
def load_model(model_type, model_path):