Image to Patch Embedding Args: patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. norm_layer (nn.Module, optional): Normalization
| 285 | |
| 286 | |
| 287 | class PatchEmbed(nn.Module): |
| 288 | """ Image to Patch Embedding |
| 289 | |
| 290 | Args: |
| 291 | patch_size (int): Patch token size. Default: 4. |
| 292 | in_chans (int): Number of input image channels. Default: 3. |
| 293 | embed_dim (int): Number of linear projection output channels. Default: 96. |
| 294 | norm_layer (nn.Module, optional): Normalization layer. Default: None |
| 295 | use_conv_embed (bool): Whether use overlapped convolution for patch embedding. Default: False |
| 296 | is_stem (bool): Is the stem block or not. |
| 297 | """ |
| 298 | |
| 299 | def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None, use_conv_embed=False, is_stem=False): |
| 300 | super().__init__() |
| 301 | patch_size = to_2tuple(patch_size) |
| 302 | self.patch_size = patch_size |
| 303 | |
| 304 | self.in_chans = in_chans |
| 305 | self.embed_dim = embed_dim |
| 306 | |
| 307 | if use_conv_embed: |
| 308 | # if we choose to use conv embedding, then we treat the stem and non-stem differently |
| 309 | if is_stem: |
| 310 | kernel_size = 7; padding = 2; stride = 4 |
| 311 | else: |
| 312 | kernel_size = 3; padding = 1; stride = 2 |
| 313 | self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding) |
| 314 | else: |
| 315 | self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) |
| 316 | |
| 317 | if norm_layer is not None: |
| 318 | self.norm = norm_layer(embed_dim) |
| 319 | else: |
| 320 | self.norm = None |
| 321 | |
| 322 | def forward(self, x): |
| 323 | """Forward function.""" |
| 324 | _, _, H, W = x.size() |
| 325 | if W % self.patch_size[1] != 0: |
| 326 | x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) |
| 327 | if H % self.patch_size[0] != 0: |
| 328 | x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) |
| 329 | |
| 330 | x = self.proj(x) # B C Wh Ww |
| 331 | if self.norm is not None: |
| 332 | Wh, Ww = x.size(2), x.size(3) |
| 333 | x = x.flatten(2).transpose(1, 2) |
| 334 | x = self.norm(x) |
| 335 | x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) |
| 336 | |
| 337 | return x |
| 338 | |
| 339 | |
| 340 | class FocalNet(nn.Module): |