src/pixparse/task/task_cruller_eval_docvqa.py [61:83]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        device_env: DeviceEnv,
        monitor: Monitor = None,
    ):
        super().__init__(
            cfg=cfg,
            device_env=device_env,
            monitor=monitor,
        )
        self.cfg = cfg
        self.amp_dtype = None
        if cfg.dtype is not None:
            self.amp_dtype = (
                torch.bfloat16 if cfg.dtype in ("bfloat16", "bf16") else torch.float16
            )

        self.task_start_token = "<s_docvqa>"
        self.prompt_end_token = "<s_answer>"
        self.max_position_embeddings = cfg.model.text_decoder.max_length
        self.text_anno_fn = True  # set for image-text dataset experiments
        self.tokenizer = TokenizerHF(cfg.tokenizer)

        self.state_dict = OrderedDict()
        self.resume = False
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



src/pixparse/task/task_cruller_finetune_docvqa.py [71:96]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        device_env: DeviceEnv,
        monitor: Monitor = None,
    ):
        super().__init__(
            cfg=cfg,
            device_env=device_env,
            monitor=monitor,
        )
        self.cfg = cfg
        # NOTE dtype is currently being used as 'amp dtype' only, ie the low precision type,
        #  we may want to differentiate different precision modes such as
        #  amp + dtype, pure float16/bfloat16, custom mixed prec, etc
        self.amp_dtype = None
        if cfg.dtype is not None:
            self.amp_dtype = (
                torch.bfloat16 if cfg.dtype in ("bfloat16", "bf16") else torch.float16
            )

        self.task_start_token = "<s_docvqa>"
        self.prompt_end_token = "<s_answer>" # Slice prompt right before answer content
        self.max_position_embeddings = cfg.model.text_decoder.max_length
        self.text_anno_fn = True  # set for image-text dataset experiments
        self.tokenizer = TokenizerHF(cfg.tokenizer)

        self.state_dict = OrderedDict()
        self.resume = False
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -