Fused `Conv2d()` followed by `downsample_2d()`. Padding is performed only once at the beginning, not between the operations. The fused op is considerably more efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of arbitrary o
(self, hidden_states, weight=None, kernel=None, factor=2, gain=1)
| 358 | self.out_channels = out_channels |
| 359 | |
| 360 | def _downsample_2d(self, hidden_states, weight=None, kernel=None, factor=2, gain=1): |
| 361 | """Fused `Conv2d()` followed by `downsample_2d()`. |
| 362 | Padding is performed only once at the beginning, not between the operations. The fused op is considerably more |
| 363 | efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of |
| 364 | arbitrary order. |
| 365 | |
| 366 | Args: |
| 367 | hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`. |
| 368 | weight: |
| 369 | Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be |
| 370 | performed by `inChannels = x.shape[0] // numGroups`. |
| 371 | kernel: FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * |
| 372 | factor`, which corresponds to average pooling. |
| 373 | factor: Integer downsampling factor (default: 2). |
| 374 | gain: Scaling factor for signal magnitude (default: 1.0). |
| 375 | |
| 376 | Returns: |
| 377 | output: Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and |
| 378 | same datatype as `x`. |
| 379 | """ |
| 380 | |
| 381 | assert isinstance(factor, int) and factor >= 1 |
| 382 | if kernel is None: |
| 383 | kernel = [1] * factor |
| 384 | |
| 385 | # setup kernel |
| 386 | kernel = torch.tensor(kernel, dtype=torch.float32) |
| 387 | if kernel.ndim == 1: |
| 388 | kernel = torch.outer(kernel, kernel) |
| 389 | kernel /= torch.sum(kernel) |
| 390 | |
| 391 | kernel = kernel * gain |
| 392 | |
| 393 | if self.use_conv: |
| 394 | _, _, convH, convW = weight.shape |
| 395 | pad_value = (kernel.shape[0] - factor) + (convW - 1) |
| 396 | stride_value = [factor, factor] |
| 397 | upfirdn_input = upfirdn2d_native( |
| 398 | hidden_states, |
| 399 | torch.tensor(kernel, device=hidden_states.device), |
| 400 | pad=((pad_value + 1) // 2, pad_value // 2), |
| 401 | ) |
| 402 | output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0) |
| 403 | else: |
| 404 | pad_value = kernel.shape[0] - factor |
| 405 | output = upfirdn2d_native( |
| 406 | hidden_states, |
| 407 | torch.tensor(kernel, device=hidden_states.device), |
| 408 | down=factor, |
| 409 | pad=((pad_value + 1) // 2, pad_value // 2), |
| 410 | ) |
| 411 | |
| 412 | return output |
| 413 | |
| 414 | def forward(self, hidden_states): |
| 415 | if self.use_conv: |
no test coverage detected