Spaces:
Sleeping
Sleeping
| import os | |
| import torch | |
| from PIL import Image | |
| import numpy as np | |
| from huggingface_hub import hf_hub_download | |
| from .rrdbnet_arch import RRDBNet | |
| from .utils import pad_reflect, split_image_into_overlapping_patches, stich_together, unpad_image | |
| HF_MODELS = { | |
| 2: dict(repo_id='ai-forever/Real-ESRGAN', filename='RealESRGAN_x2.pth'), | |
| 4: dict(repo_id='ai-forever/Real-ESRGAN', filename='RealESRGAN_x4.pth'), | |
| 8: dict(repo_id='ai-forever/Real-ESRGAN', filename='RealESRGAN_x8.pth'), | |
| } | |
| class RealESRGAN: | |
| def __init__(self, device, scale=4): | |
| self.device = device | |
| self.scale = scale | |
| self.model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=scale) | |
| def load_weights(self, model_path, download=True): | |
| if not os.path.exists(model_path) and download: | |
| config = HF_MODELS[self.scale] | |
| cache_dir = os.path.dirname(model_path) or 'weights' | |
| os.makedirs(cache_dir, exist_ok=True) | |
| hf_hub_download(repo_id=config['repo_id'], filename=config['filename'], local_dir=cache_dir) | |
| print('Weights downloaded to:', model_path) | |
| loadnet = torch.load(model_path, map_location=self.device) | |
| if 'params' in loadnet: | |
| self.model.load_state_dict(loadnet['params'], strict=True) | |
| elif 'params_ema' in loadnet: | |
| self.model.load_state_dict(loadnet['params_ema'], strict=True) | |
| else: | |
| self.model.load_state_dict(loadnet, strict=True) | |
| self.model.eval().to(self.device) | |
| def predict(self, lr_image, batch_size=4, patches_size=192, padding=24, pad_size=15): | |
| lr_image = np.array(lr_image) | |
| lr_image = pad_reflect(lr_image, pad_size) | |
| # PERBAIKAN 3: Terima n_patches_per_row dari fungsi split | |
| patches, p_shape, n_patches_per_row = split_image_into_overlapping_patches(lr_image, patch_size=patches_size, padding_size=padding) | |
| img = torch.FloatTensor(patches/255).permute((0,3,1,2)).to(self.device).detach() | |
| with torch.no_grad(): | |
| res = self.model(img[0:batch_size]) | |
| for i in range(batch_size, img.shape[0], batch_size): | |
| res = torch.cat((res, self.model(img[i:i+batch_size])), 0) | |
| sr_image = res.permute((0,2,3,1)).cpu().clamp_(0, 1).numpy() | |
| padded_size_scaled = tuple(np.multiply(p_shape[0:2], self.scale)) + (3,) | |
| scaled_image_shape = tuple(np.multiply(lr_image.shape[0:2], self.scale)) + (3,) | |
| # PERBAIKAN 4: Kirim n_patches_per_row ke fungsi stich_together | |
| np_sr_image = stich_together( | |
| sr_image, | |
| padded_image_shape=padded_size_scaled, | |
| target_shape=scaled_image_shape, | |
| padding_size=padding * self.scale, | |
| n_patches_per_row=n_patches_per_row | |
| ) | |
| sr_img = (np_sr_image * 255).astype(np.uint8) | |
| sr_img = unpad_image(sr_img, pad_size * self.scale) | |
| return Image.fromarray(sr_img) |