| 221 | |
| 222 | |
| 223 | class Block(nn.Module): |
| 224 | def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., |
| 225 | drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm): |
| 226 | super().__init__() |
| 227 | self.norm1 = norm_layer(dim) |
| 228 | self.attn = Attention( |
| 229 | dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) |
| 230 | self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() |
| 231 | self.norm2 = norm_layer(dim) |
| 232 | mlp_hidden_dim = int(dim * mlp_ratio) |
| 233 | self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) |
| 234 | |
| 235 | def forward(self, x, return_attention=False): |
| 236 | y, attn = self.attn(self.norm1(x)) |
| 237 | if return_attention: |
| 238 | return attn |
| 239 | x = x + self.drop_path(y) |
| 240 | x = x + self.drop_path(self.mlp(self.norm2(x))) |
| 241 | return x |
| 242 | |
| 243 | |
| 244 | class PatchEmbed(nn.Module): |