Image to Patch Embedding
| 262 | |
| 263 | |
| 264 | class PatchEmbed(nn.Module): |
| 265 | """ Image to Patch Embedding |
| 266 | """ |
| 267 | |
| 268 | def __init__(self, |
| 269 | img_size=[32, 100], |
| 270 | in_channels=3, |
| 271 | embed_dim=768, |
| 272 | sub_num=2, |
| 273 | patch_size=[4, 4], |
| 274 | mode='pope'): |
| 275 | super().__init__() |
| 276 | num_patches = (img_size[1] // (2 ** sub_num)) * \ |
| 277 | (img_size[0] // (2 ** sub_num)) |
| 278 | self.img_size = img_size |
| 279 | self.num_patches = num_patches |
| 280 | self.embed_dim = embed_dim |
| 281 | self.norm = None |
| 282 | if mode == 'pope': |
| 283 | if sub_num == 2: |
| 284 | self.proj = nn.Sequential( |
| 285 | ConvBNLayer( |
| 286 | in_channels=in_channels, |
| 287 | out_channels=embed_dim // 2, |
| 288 | kernel_size=3, |
| 289 | stride=2, |
| 290 | padding=1, |
| 291 | act=nn.GELU, |
| 292 | bias_attr=False), |
| 293 | ConvBNLayer( |
| 294 | in_channels=embed_dim // 2, |
| 295 | out_channels=embed_dim, |
| 296 | kernel_size=3, |
| 297 | stride=2, |
| 298 | padding=1, |
| 299 | act=nn.GELU, |
| 300 | bias_attr=False)) |
| 301 | if sub_num == 3: |
| 302 | self.proj = nn.Sequential( |
| 303 | ConvBNLayer( |
| 304 | in_channels=in_channels, |
| 305 | out_channels=embed_dim // 4, |
| 306 | kernel_size=3, |
| 307 | stride=2, |
| 308 | padding=1, |
| 309 | act=nn.GELU, |
| 310 | bias_attr=False), |
| 311 | ConvBNLayer( |
| 312 | in_channels=embed_dim // 4, |
| 313 | out_channels=embed_dim // 2, |
| 314 | kernel_size=3, |
| 315 | stride=2, |
| 316 | padding=1, |
| 317 | act=nn.GELU, |
| 318 | bias_attr=False), |
| 319 | ConvBNLayer( |
| 320 | in_channels=embed_dim // 2, |
| 321 | out_channels=embed_dim, |