How to reproduce VQA performance

[ys-zong]

Hi, I'm trying to reproduce the VQA performance in Table 17 (VQAv2, SQA, POPE, MMB). However, my reproduced performance is quite low, e.g. 43% on ScienceQA compared to the reported 64.9%. I'm using the original LLaVA evaluation scripts and I changed the forward/generate functions in PSALM class to only generate texts (mostly following LLaVA):

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        labels: Optional[torch.LongTensor] = None,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        if inputs_embeds is None:
            input_ids, attention_mask, past_key_values, inputs_embeds, _ = self.mm_conv_prepare_inputs_labels_for_multimodal(
                input_ids, attention_mask, past_key_values, labels, images)

        return super().forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            labels=labels,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )
    @torch.no_grad()
    def generate(
            self,
            inputs: Optional[torch.Tensor] = None,
            images: Optional[torch.Tensor] = None,
            **kwargs,
    ) -> Union[GenerateOutput, torch.LongTensor]:

        position_ids = kwargs.pop("position_ids", None)
        attention_mask = kwargs.pop("attention_mask", None)
        if "inputs_embeds" in kwargs:
            raise NotImplementedError("`inputs_embeds` is not supported")

        if images is not None:
            inputs, attention_mask, past_key_values, inputs_embeds, _ = self.mm_conv_prepare_inputs_labels_for_multimodal(
            inputs, attention_mask, None, None, images)
        else:
            inputs_embeds = self.get_model().embed_tokens(inputs)

        return super().generate(
            position_ids=position_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            **kwargs
        )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None,
                                      inputs_embeds=None, **kwargs):
        images = kwargs.pop("images", None)
        image_sizes = kwargs.pop("image_sizes", None)
        inputs = super().prepare_inputs_for_generation(
            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
        )
        if images is not None:
            inputs['images'] = images
        if image_sizes is not None:
            inputs['image_sizes'] = image_sizes
        return inputs

Do you think this is the correct implementation or can you share your evaluation scripts for these tasks? Thanks a lot!