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.
| 347 | |
| 348 | |
| 349 | class AttentionBlock(nn.Module): |
| 350 | """ |
| 351 | An attention block that allows spatial positions to attend to each other. |
| 352 | Originally ported from here, but adapted to the N-d case. |
| 353 | https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. |
| 354 | """ |
| 355 | |
| 356 | def __init__( |
| 357 | self, |
| 358 | channels, |
| 359 | num_heads=1, |
| 360 | num_head_channels=-1, |
| 361 | use_checkpoint=False, |
| 362 | use_new_attention_order=False, |
| 363 | ): |
| 364 | super().__init__() |
| 365 | self.channels = channels |
| 366 | if num_head_channels == -1: |
| 367 | self.num_heads = num_heads |
| 368 | else: |
| 369 | assert ( |
| 370 | channels % num_head_channels == 0 |
| 371 | ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" |
| 372 | self.num_heads = channels // num_head_channels |
| 373 | self.use_checkpoint = use_checkpoint |
| 374 | self.norm = normalization(channels) |
| 375 | self.qkv = conv_nd(1, channels, channels * 3, 1) |
| 376 | if use_new_attention_order: |
| 377 | # split qkv before split heads |
| 378 | self.attention = QKVAttention(self.num_heads) |
| 379 | else: |
| 380 | # split heads before split qkv |
| 381 | self.attention = QKVAttentionLegacy(self.num_heads) |
| 382 | |
| 383 | self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) |
| 384 | |
| 385 | def forward(self, x, **kwargs): |
| 386 | # TODO add crossframe attention and use mixed checkpoint |
| 387 | return checkpoint( |
| 388 | self._forward, (x,), self.parameters(), True |
| 389 | ) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! |
| 390 | # return pt_checkpoint(self._forward, x) # pytorch |
| 391 | |
| 392 | def _forward(self, x): |
| 393 | b, c, *spatial = x.shape |
| 394 | x = x.reshape(b, c, -1) |
| 395 | qkv = self.qkv(self.norm(x)) |
| 396 | h = self.attention(qkv) |
| 397 | h = self.proj_out(h) |
| 398 | return (x + h).reshape(b, c, *spatial) |
| 399 | |
| 400 | |
| 401 | def count_flops_attn(model, _x, y): |