Delete ffdnet_model.py

ffdnet_model.py

@@ -1,50 +0,0 @@
-"""FFDNet denoiser (Zhang et al., 2018 — https://arxiv.org/abs/1710.04026).
-
-A arquitetura espelha exatamente a implementação oficial do KAIR
-(https://github.com/cszn/KAIR, arquivos `models/network_ffdnet.py` +
-`models/basicblock.py`) para que os pesos pré-treinados do release v1.0
-carreguem com `strict=True`, sem rename de chaves.
-
-Convs são empilhadas num único `nn.Sequential` (sem blocos aninhados),
-reproduzindo o comportamento do `basicblock.sequential(...)` do KAIR que
-achata filhos de `nn.Sequential`.
-
-Configurações oficiais:
-  - color:  in_nc=3, out_nc=3, nc=96, nb=12  (ffdnet_color.pth)
-  - gray:   in_nc=1, out_nc=1, nc=64, nb=15  (ffdnet_gray.pth)
-"""
-from __future__ import annotations
-
-import math
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-
-class FFDNet(nn.Module):
-    def __init__(self, in_nc: int = 3, out_nc: int = 3, nc: int = 96, nb: int = 12):
-        super().__init__()
-        self.sf = 2
-
-        layers: list[nn.Module] = []
-        layers.append(nn.Conv2d(in_nc * self.sf * self.sf + 1, nc, 3, 1, 1, bias=True))
-        layers.append(nn.ReLU(inplace=True))
-        for _ in range(nb - 2):
-            layers.append(nn.Conv2d(nc, nc, 3, 1, 1, bias=True))
-            layers.append(nn.ReLU(inplace=True))
-        layers.append(nn.Conv2d(nc, out_nc * self.sf * self.sf, 3, 1, 1, bias=True))
-        self.model = nn.Sequential(*layers)
-
-    def forward(self, x: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
-        h, w = x.shape[-2], x.shape[-1]
-        pad_b = int(math.ceil(h / 2) * 2 - h)
-        pad_r = int(math.ceil(w / 2) * 2 - w)
-        x = F.pad(x, (0, pad_r, 0, pad_b), mode="replicate")
-
-        x = F.pixel_unshuffle(x, self.sf)
-        noise_map = sigma.view(-1, 1, 1, 1).expand(x.shape[0], 1, x.shape[-2], x.shape[-1])
-        x = torch.cat([x, noise_map], dim=1)
-        x = self.model(x)
-        x = F.pixel_shuffle(x, self.sf)
-        return x[..., :h, :w]