Function get_batch

scripts/demo/streamlit_helpers.py:600–654

(keys, value_dict, N: Union[List, ListConfig], device="cuda")

def get_batch(keys, value_dict, N: Union[List, ListConfig], device="cuda"):
    # Hardcoded demo setups; might undergo some changes in the future

    batch = {}
    batch_uc = {}

    for key in keys:
        if key == "txt":
            batch["txt"] = (
                np.repeat([value_dict["prompt"]], repeats=math.prod(N))
                .reshape(N)
                .tolist()
            )
            batch_uc["txt"] = (
                np.repeat([value_dict["negative_prompt"]], repeats=math.prod(N))
                .reshape(N)
                .tolist()
            )
        elif key == "original_size_as_tuple":
            batch["original_size_as_tuple"] = (
                torch.tensor([value_dict["orig_height"], value_dict["orig_width"]])
                .to(device)
                .repeat(*N, 1)
            )
        elif key == "crop_coords_top_left":
            batch["crop_coords_top_left"] = (
                torch.tensor(
                    [value_dict["crop_coords_top"], value_dict["crop_coords_left"]]
                )
                .to(device)
                .repeat(*N, 1)
            )
        elif key == "aesthetic_score":
            batch["aesthetic_score"] = (
                torch.tensor([value_dict["aesthetic_score"]]).to(device).repeat(*N, 1)
            )
            batch_uc["aesthetic_score"] = (
                torch.tensor([value_dict["negative_aesthetic_score"]])
                .to(device)
                .repeat(*N, 1)
            )

        elif key == "target_size_as_tuple":
            batch["target_size_as_tuple"] = (
                torch.tensor([value_dict["target_height"], value_dict["target_width"]])
                .to(device)
                .repeat(*N, 1)
            )
        else:
            batch[key] = value_dict[key]

    for key in batch.keys():
        if key not in batch_uc and isinstance(batch[key], torch.Tensor):
            batch_uc[key] = torch.clone(batch[key])
    return batch, batch_uc


@torch.no_grad()