Dynamatrix/DiffBIR-OpenXLab

1from typing import overload2import torch3from torch.nn import functional as F4 5 6class Guidance:7    8    def __init__(self, scale, type, t_start, t_stop, space, repeat, loss_type):9        self.scale = scale10        self.type = type11        self.t_start = t_start12        self.t_stop = t_stop13        self.target = None14        self.space = space15        self.repeat = repeat16        self.loss_type = loss_type17    18    def load_target(self, target):19        self.target = target20 21    def __call__(self, target_x0, pred_x0, t):22        if self.t_stop < t and t < self.t_start:23            # print("sampling with classifier guidance")24            # avoid propagating gradient out of this scope25            pred_x0 = pred_x0.detach().clone()26            target_x0 = target_x0.detach().clone()27            return self.scale * self._forward(target_x0, pred_x0)28        else:29            return None30    31    @overload32    def _forward(self, target_x0, pred_x0): ...33 34 35class MSEGuidance(Guidance):36    37    def __init__(self, scale, type, t_start, t_stop, space, repeat, loss_type) -> None:38        super().__init__(39            scale, type, t_start, t_stop, space, repeat, loss_type40        )41    42    @torch.enable_grad()43    def _forward(self, target_x0: torch.Tensor, pred_x0: torch.Tensor):44        # inputs: [-1, 1], nchw, rgb45        pred_x0.requires_grad_(True)46        47        if self.loss_type == "mse":48            loss = (pred_x0 - target_x0).pow(2).mean((1, 2, 3)).sum()49        elif self.loss_type == "downsample_mse":50            # FIXME: scale_factor should be 1/4, not 451            lr_pred_x0 = F.interpolate(pred_x0, scale_factor=4, mode="bicubic")52            lr_target_x0 = F.interpolate(target_x0, scale_factor=4, mode="bicubic")53            loss = (lr_pred_x0 - lr_target_x0).pow(2).mean((1, 2, 3)).sum()54        else:55            raise ValueError(self.loss_type)56        57        print(f"loss = {loss.item()}")58        return -torch.autograd.grad(loss, pred_x0)[0]59