An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted to the N-d case. https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
| 279 | |
| 280 | |
| 281 | class AttentionBlock(nn.Module): |
| 282 | """ |
| 283 | An attention block that allows spatial positions to attend to each other. |
| 284 | Originally ported from here, but adapted to the N-d case. |
| 285 | https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. |
| 286 | """ |
| 287 | |
| 288 | def __init__( |
| 289 | self, |
| 290 | channels, |
| 291 | num_heads=1, |
| 292 | num_head_channels=-1, |
| 293 | use_checkpoint=False, |
| 294 | use_new_attention_order=False, |
| 295 | ): |
| 296 | super().__init__() |
| 297 | self.channels = channels |
| 298 | if num_head_channels == -1: |
| 299 | self.num_heads = num_heads |
| 300 | else: |
| 301 | assert ( |
| 302 | channels % num_head_channels == 0 |
| 303 | ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" |
| 304 | self.num_heads = channels // num_head_channels |
| 305 | self.use_checkpoint = use_checkpoint |
| 306 | self.norm = normalization(channels) |
| 307 | self.qkv = conv_nd(1, channels, channels * 3, 1) |
| 308 | if use_new_attention_order: |
| 309 | # split qkv before split heads |
| 310 | self.attention = QKVAttention(self.num_heads) |
| 311 | else: |
| 312 | # split heads before split qkv |
| 313 | self.attention = QKVAttentionLegacy(self.num_heads) |
| 314 | |
| 315 | self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) |
| 316 | |
| 317 | def forward(self, x): |
| 318 | return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! |
| 319 | #return pt_checkpoint(self._forward, x) # pytorch |
| 320 | |
| 321 | def _forward(self, x): |
| 322 | b, c, *spatial = x.shape |
| 323 | x = x.reshape(b, c, -1) |
| 324 | qkv = self.qkv(self.norm(x)) |
| 325 | h = self.attention(qkv) |
| 326 | h = self.proj_out(h) |
| 327 | return (x + h).reshape(b, c, *spatial) |
| 328 | |
| 329 | |
| 330 | def count_flops_attn(model, _x, y): |