Feature fusion block.
| 380 | |
| 381 | |
| 382 | class FeatureFusionBlock_custom(nn.Module): |
| 383 | """Feature fusion block. |
| 384 | """ |
| 385 | |
| 386 | def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, size=None): |
| 387 | """Init. |
| 388 | |
| 389 | Args: |
| 390 | features (int): number of features |
| 391 | """ |
| 392 | super(FeatureFusionBlock_custom, self).__init__() |
| 393 | |
| 394 | self.deconv = deconv |
| 395 | self.align_corners = align_corners |
| 396 | |
| 397 | self.groups=1 |
| 398 | |
| 399 | self.expand = expand |
| 400 | out_features = features |
| 401 | if self.expand==True: |
| 402 | out_features = features//2 |
| 403 | |
| 404 | self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1) |
| 405 | |
| 406 | self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) |
| 407 | self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) |
| 408 | |
| 409 | self.skip_add = nn.quantized.FloatFunctional() |
| 410 | |
| 411 | self.size=size |
| 412 | |
| 413 | def forward(self, *xs, size=None): |
| 414 | """Forward pass. |
| 415 | |
| 416 | Returns: |
| 417 | tensor: output |
| 418 | """ |
| 419 | output = xs[0] |
| 420 | |
| 421 | if len(xs) == 2: |
| 422 | res = self.resConfUnit1(xs[1]) |
| 423 | output = self.skip_add.add(output, res) |
| 424 | # output += res |
| 425 | |
| 426 | output = self.resConfUnit2(output) |
| 427 | |
| 428 | if (size is None) and (self.size is None): |
| 429 | modifier = {"scale_factor": 2} |
| 430 | elif size is None: |
| 431 | modifier = {"size": self.size} |
| 432 | else: |
| 433 | modifier = {"size": size} |
| 434 | |
| 435 | output = nn.functional.interpolate( |
| 436 | output, **modifier, mode="bilinear", align_corners=self.align_corners |
| 437 | ) |
| 438 | |
| 439 | output = self.out_conv(output) |
no outgoing calls
no test coverage detected