from skimage.exposure import match_histograms from skimage import io import os from PIL import Image import torch import torchvision import torchvision.transforms as transforms def normalize(): MEAN = [0.485, 0.456, 0.406] STD = [0.229, 0.224, 0.225] return transforms.Normalize(mean = MEAN, std = STD) def denormalize(): # out = (x - mean) / std MEAN = [0.485, 0.456, 0.406] STD = [0.229, 0.224, 0.225] MEAN = [-mean/std for mean, std in zip(MEAN, STD)] STD = [1/std for std in STD] return transforms.Normalize(mean=MEAN, std=STD) def transformer(imsize = None, cropsize = None): transformer = [] if imsize: transformer.append(transforms.Resize(imsize)) if cropsize: transformer.append(transforms.RandomCrop(cropsize)) transformer.append(transforms.ToTensor()) transformer.append(normalize()) return transforms.Compose(transformer) def load_img(path, imsize = None, cropsize = None): transform = transformer(imsize = imsize, cropsize = cropsize) # torchvision.transforms supports PIL Images return transform(Image.open(path).convert("RGB")).unsqueeze(0) def tensor_to_img(tensor): denormalizer = denormalize() if tensor.device == "cuda": tensor = tensor.cpu() # tensor = torchvision.utils.make_grid(denormalizer(tensor.squeeze())) image = transforms.functional.to_pil_image(tensor.clamp_(0., 1.)) return image def save_img(tensor, path): pass def histogram_matching(image, reference): """ img: style image reference: original img output: style image that resembles original img's color histogram """ device = image.device reference = reference.cpu().permute(1, 2, 0).numpy() image = image.cpu().permute(1, 2, 0).numpy() output = match_histograms(image, reference, multichannel = True) return torch.Tensor(output).permute(2, 0, 1).to(device) def batch_histogram_matching(images, reference): """ images of shape BxCxHxW reference of shape 1xCxHxW """ reference = reference.squeeze() output = torch.zeros_like(images, dtype = images.dtype) B = images.shape[0] for i in range(B): output[i] = histogram_matching(images[i], reference) return output def statistics(f, inverse = False, eps = 1e-10): c, h, w = f.shape f_mean = torch.mean(f.view(c, h*w), dim=1, keepdim=True) f_zeromean = f.view(c, h*w) - f_mean f_cov = torch.mm(f_zeromean, f_zeromean.t()) u, s, v = torch.svd(f_cov) k = c for i in range(c): if s[i] < eps: k = i break if inverse: p = -0.5 else: p = 0.5 f_covsqrt = torch.mm(torch.mm(u[:, 0:k], torch.diag(s[0:k].pow(p))), v[:, 0:k].t()) return f_mean, f_covsqrt def whitening(f): c, h, w = f.shape f_mean, f_inv_covsqrt = statistics(f, inverse = True) whitened_f = torch.mm(f_inv_covsqrt, f.view(c, h*w) - f_mean) return whitened_f.view(c, h, w) def batch_whitening(f): b, c, h, w = f.shape whitened_f = torch.zeros(size = (b, c, h, w), dtype = f.dtype, device = f.device) for i in range(b): whitened_f[i] = whitening(f[i]) return whitened_f def coloring(style, content): s_c, s_h, s_w = style.shape c_mean, c_covsqrt = statistics(content, inverse = False) colored_s = torch.mm(c_covsqrt, whitening(style).view(s_c, s_h * s_w)) + c_mean return colored_s.view(s_c, s_h, s_w) def batch_coloring(styles, content): colored_styles = torch.zeros_like(styles, dtype = styles.dtype, device = styles.device) for i, style in enumerate(styles): colored_styles[i] = coloring(style, content[i]) return colored_styles def batch_wct(styles, content): whitened_styles = batch_whitening(styles) return batch_coloring(whitened_styles, content) class Image_Set(torch.utils.data.Dataset): def __init__(self, root_path, imsize, cropsize): super(Image_Set, self).__init__() self.root_path = root_path self.files = sorted(os.listdir(self.root_path)) self.transformer = transformer(imsize, cropsize) def __len__(self): return len(self.file_names) def __getitem__(self, index): image = Image.open(os.path.join(self.root_path + self.file_names[index])).convert("RGB") return self.transformer(image)