import numpy as np import torch import torch.nn as nn from einops.layers.torch import Rearrange class PatchFeatureExtractor(nn.Module): x_mean = torch.FloatTensor(np.array([0.485, 0.456, 0.406])[None, :, None, None]) x_std = torch.FloatTensor(np.array([0.229, 0.224, 0.225])[None, :, None, None]) def __init__(self, patch_num=256, input_shape=None): super(PatchFeatureExtractor, self).__init__() if input_shape is None: input_shape = [3, 512, 1024] self.patch_dim = 1024 self.patch_num = patch_num img_channel = input_shape[0] img_h = input_shape[1] img_w = input_shape[2] p_h, p_w = img_h, img_w // self.patch_num p_dim = p_h * p_w * img_channel self.patch_embedding = nn.Sequential( Rearrange('b c h (p_n p_w) -> b p_n (h p_w c)', p_w=p_w), nn.Linear(p_dim, self.patch_dim) ) self.x_mean.requires_grad = False self.x_std.requires_grad = False def _prepare_x(self, x): x = x.clone() if self.x_mean.device != x.device: self.x_mean = self.x_mean.to(x.device) self.x_std = self.x_std.to(x.device) x[:, :3] = (x[:, :3] - self.x_mean) / self.x_std return x def forward(self, x): # x [b 3 512 1024] x = self._prepare_x(x) # [b 3 512 1024] x = self.patch_embedding(x) # [b 256(patch_num) 1024(d)] x = x.permute(0, 2, 1) # [b 1024(d) 256(patch_num)] return x if __name__ == '__main__': from PIL import Image extractor = PatchFeatureExtractor() img = np.array(Image.open("../../src/demo.png")).transpose((2, 0, 1)) input = torch.Tensor([img]) # 1 3 512 1024 feature = extractor(input) print(feature.shape) # 1, 1024, 256