Add TDAN backbone #316
Add TDAN backbone #316
Conversation
Codecov Report
@@ Coverage Diff @@
## master #316 +/- ##
==========================================
- Coverage 80.30% 80.05% -0.25%
==========================================
Files 177 178 +1
Lines 9280 9332 +52
Branches 1352 1357 +5
==========================================
+ Hits 7452 7471 +19
- Misses 1640 1673 +33
Partials 188 188
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
| generate the offsets. | ||
|
|
||
| Args: | ||
| in_channels (int): Same as nn.Conv2d. |
There was a problem hiding this comment.
Please refactor docstring of args or copy explain of these args from nn.Conv2d
| False. | ||
| """ | ||
|
|
||
| def __init__(self, *args, **kwargs): |
There was a problem hiding this comment.
How about the same as docstring rather than use *args and **kwargs?
There was a problem hiding this comment.
This class is similar to DeformConv2dPack. Therefore I think it is better to follow the definition in DeformConv2dPack.
Reference: https://mmcv.readthedocs.io/en/latest/_modules/mmcv/ops/deform_conv.html#DeformConv2dPack
|
|
||
| self.conv_offset = nn.Conv2d( | ||
| self.in_channels, | ||
| self.deform_groups * 2 * self.kernel_size[0] * self.kernel_size[1], |
There was a problem hiding this comment.
deform_groups is not in docstring
| super().__init__() | ||
|
|
||
| self.feat_extract = nn.Sequential( | ||
| ConvModule(3, 64, 3, padding=1), |
There was a problem hiding this comment.
How about declaring args such as in_channels, out_channels, mid_channels=64.
|
|
||
| self.reconstruct = nn.Sequential( | ||
| ConvModule(3 * 5, 64, 3, padding=1), | ||
| make_layer(ResidualBlockNoBN, 10, mid_channels=64), |
| lrs (Tensor): Input LR sequence with shape (n, t, c, h, w). | ||
|
|
||
| Returns: | ||
| Tensor: Output HR image with shape (n, c, 4h, 4w). |
There was a problem hiding this comment.
The returns in docstring are different from those in code.
|
Please do not merge for the moment. |
Okay now. |
* Add TDAN architecture * Modify docstring * Fix bug in unittest * Update backbone * Minor update * Change TDANNet arguments
TDAN: Temporally-Deformable Alignment Network for Video Super-Resolution, CVPR, 2020