Image to Patch Embedding
| 115 | |
| 116 | |
| 117 | class PatchEmbed(nn.Module): |
| 118 | """ Image to Patch Embedding |
| 119 | """ |
| 120 | def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): |
| 121 | super().__init__() |
| 122 | num_patches = (img_size // patch_size) * (img_size // patch_size) |
| 123 | self.img_size = img_size |
| 124 | self.patch_size = patch_size |
| 125 | self.num_patches = num_patches |
| 126 | |
| 127 | self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) |
| 128 | |
| 129 | def forward(self, x): |
| 130 | B, C, H, W = x.shape |
| 131 | x = self.proj(x).flatten(2).transpose(1, 2) |
| 132 | return x |
| 133 | |
| 134 | |
| 135 | class VisionTransformer(nn.Module): |