Method forward

src/custom_modules.py:293–312

(self, text)

            param.requires_grad = False

    def forward(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.device)

        indices = tokens == self.modifier_token_id[-1]
        for token_id in self.modifier_token_id:
            indices |= tokens == token_id

        indices = (indices*1).unsqueeze(-1)

        input_shape = tokens.size()
        tokens = tokens.view(-1, input_shape[-1])

        hidden_states = self.transformer.text_model.embeddings(input_ids=tokens)
        hidden_states = (1-indices)*hidden_states.detach() + indices*hidden_states

        z = self.custom_forward(hidden_states, tokens)

        return z

    def encode(self, text):
        return self(text)