from __future__ import annotations import os from typing import List, Sequence, Tuple from PIL import Image import torch from torch.utils.data import Dataset from torchvision import transforms as T _IMG_EXTS = (".png", ".jpg", ".jpeg", ".webp", ".bmp") def _list_images(root: str) -> List[str]: paths: List[str] = [] for dirpath, _, filenames in os.walk(root): for fn in filenames: if fn.lower().endswith(_IMG_EXTS): paths.append(os.path.join(dirpath, fn)) paths.sort() return paths class TextureFolderDataset(Dataset): """ 纹理图片数据集: - 读取目录下所有图片(含子目录) - 统一转换为 RGB - 缩放到 16×16 - 输出范围为 [-1, 1] 的张量(C,H,W) """ def __init__( self, root_dir: str, image_size: int = 16, channels: int = 3, random_flip: bool = True, ): self.root_dir = root_dir self.image_size = int(image_size) self.channels = int(channels) if self.channels not in (3, 4): raise ValueError("channels 仅支持 3(RGB) 或 4(RGBA)") self.paths = _list_images(root_dir) if len(self.paths) == 0: raise ValueError(f"数据集目录为空或未找到图片:{root_dir}") tfms = [ T.Resize((self.image_size, self.image_size), interpolation=T.InterpolationMode.NEAREST), ] if random_flip: tfms.append(T.RandomHorizontalFlip(p=0.5)) tfms.extend( [ T.ToTensor(), # [0,1] # 用 Normalize 替代 lambda,保证 Windows 多进程 DataLoader 可 pickle T.Normalize(mean=[0.5] * self.channels, std=[0.5] * self.channels), # [-1,1] ] ) self.transform = T.Compose(tfms) def __len__(self) -> int: return len(self.paths) def __getitem__(self, idx: int) -> torch.Tensor: path = self.paths[idx] # 使用 nearest resize,尽量保留像素风格边缘 with Image.open(path) as im: im = im.convert("RGBA" if self.channels == 4 else "RGB") x = self.transform(im) return x