ldm_baseline/utils.py

import pickle
from typing import Union

import numpy as np
import opensr_test
import torch
from diffusers import LDMSuperResolutionPipeline


def create_stable_diffusion_model(
    device: Union[str, torch.device] = "cuda"
) -> LDMSuperResolutionPipeline:
    """Create the stable diffusion model

    Returns:
        LDMSuperResolutionPipeline: The model to use for
            super resolution.
    """
    model_id = "CompVis/ldm-super-resolution-4x-openimages"
    pipeline = LDMSuperResolutionPipeline.from_pretrained(model_id)
    pipeline = pipeline.to(device)
    return pipeline


def run_diffuser(
    model: LDMSuperResolutionPipeline,
    lr: torch.Tensor,
    hr: torch.Tensor,
    device: Union[str, torch.device] = "cuda",
) -> dict:
    """Run the model on the low resolution image

    Args:
        model (LDMSuperResolutionPipeline): The model to use
        lr (torch.Tensor): The low resolution image
        hr (torch.Tensor): The high resolution image
        device (Union[str, torch.device], optional): The device
            to use. Defaults to "cuda".

    Returns:
        dict: The results of the model
    """

    # move the images to the device
    lr = (torch.from_numpy(lr[[3, 2, 1]]) / 2000).to(device).clamp(0, 1)

    if lr.shape[1] == 121:
        # add padding
        lr = torch.nn.functional.pad(
            lr[None], pad=(3, 4, 3, 4), mode="reflect"
        ).squeeze()

        # run the model
        with torch.no_grad():
            sr = model(lr[None], num_inference_steps=100, eta=1)
            sr = torch.from_numpy(np.array(sr.images[0]) / 255).permute(2, 0, 1).float()

        # remove padding
        sr = sr[:, 3 * 4 : -4 * 4, 3 * 4 : -4 * 4]
        lr = lr[:, 3:-4, 3:-4]
    else:
        # run the model
        with torch.no_grad():
            sr = model(lr[None], num_inference_steps=100, eta=1)
            sr = torch.from_numpy(np.array(sr.images[0]) / 255).permute(2, 0, 1).float()

    lr = (lr.cpu().numpy() * 2000).astype(np.uint16)
    hr = ((hr[0:3] / 2000).clip(0, 1) * 2000).astype(np.uint16)
    sr = (sr.cpu().numpy() * 2000).astype(np.uint16)

    results = {"lr": lr, "hr": hr, "sr": sr}

    return results