import torch from torch import nn class PatchEmbed2D(nn.Module): """ 2D Image to Patch Embedding. Args: img_size (tuple[int]): Image size. patch_size (tuple[int]): Patch token size. in_chans (int): Number of input image channels. embed_dim(int): Number of projection output channels. norm_layer (nn.Module, optional): Normalization layer. Default: None """ def __init__(self, img_size, patch_size, in_chans, embed_dim, norm_layer=None): super().__init__() self.img_size = img_size height, width = img_size h_patch_size, w_path_size = patch_size padding_left = padding_right = padding_top = padding_bottom = 0 h_remainder = height % h_patch_size w_remainder = width % w_path_size if h_remainder: h_pad = h_patch_size - h_remainder padding_top = h_pad // 2 padding_bottom = int(h_pad - padding_top) if w_remainder: w_pad = w_path_size - w_remainder padding_left = w_pad // 2 padding_right = int(w_pad - padding_left) self.pad = nn.ZeroPad2d((padding_left, padding_right, padding_top, padding_bottom)) self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) if norm_layer is not None: self.norm = norm_layer(embed_dim) else: self.norm = None def forward(self, x: torch.Tensor): B, C, H, W = x.shape assert H == self.img_size[0] and W == self.img_size[1], \ f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." x = self.pad(x) x = self.proj(x) if self.norm is not None: x = self.norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) return x class PatchEmbed3D(nn.Module): """ 3D Image to Patch Embedding. Args: img_size (tuple[int]): Image size. patch_size (tuple[int]): Patch token size. in_chans (int): Number of input image channels. embed_dim(int): Number of projection output channels. norm_layer (nn.Module, optional): Normalization layer. Default: None """ def __init__(self, img_size, patch_size, in_chans, embed_dim, norm_layer=None): super().__init__() self.img_size = img_size level, height, width = img_size l_patch_size, h_patch_size, w_patch_size = patch_size padding_left = padding_right = padding_top = padding_bottom = padding_front = padding_back = 0 l_remainder = level % l_patch_size h_remainder = height % l_patch_size w_remainder = width % w_patch_size if l_remainder: l_pad = l_patch_size - l_remainder padding_front = l_pad // 2 padding_back = l_pad - padding_front if h_remainder: h_pad = h_patch_size - h_remainder padding_top = h_pad // 2 padding_bottom = h_pad - padding_top if w_remainder: w_pad = w_patch_size - w_remainder padding_left = w_pad // 2 padding_right = w_pad - padding_left self.pad = nn.ZeroPad3d( (padding_left, padding_right, padding_top, padding_bottom, padding_front, padding_back) ) self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) if norm_layer is not None: self.norm = norm_layer(embed_dim) else: self.norm = None def forward(self, x: torch.Tensor): B, C, L, H, W = x.shape assert L == self.img_size[0] and H == self.img_size[1] and W == self.img_size[2], \ f"Input image size ({L}*{H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}*{self.img_size[2]})." x = self.pad(x) x = self.proj(x) if self.norm: x = self.norm(x.permute(0, 2, 3, 4, 1)).permute(0, 4, 1, 2, 3) return x