add upsample neck #512
add upsample neck #512
Conversation
Codecov Report
@@ Coverage Diff @@
## master #512 +/- ##
==========================================
- Coverage 86.53% 86.13% -0.41%
==========================================
Files 98 99 +1
Lines 5133 5165 +32
Branches 829 835 +6
==========================================
+ Hits 4442 4449 +7
- Misses 534 557 +23
- Partials 157 159 +2
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
mmseg/models/necks/upsample_neck.py
Outdated
| class UpsampleNeck(nn.Module): | ||
| """Upsample Network.""" |
mmseg/models/necks/upsample_neck.py
Outdated
| self.in_channels = in_channels | ||
| self.out_channels = out_channels | ||
| self.scales = scales | ||
| self.num_outs = num_outs |
There was a problem hiding this comment.
num_outs == len(scales. We may remove num_outs.
mmseg/models/necks/upsample_neck.py
Outdated
| if len(inputs) == 1: | ||
| inputs = inputs * self.num_outs |
There was a problem hiding this comment.
Repeating the same scale may not be useful. We may remove this.
mmseg/models/necks/upsample_neck.py
Outdated
| in_channels, | ||
| out_channels, | ||
| scales=[0.5, 1, 2, 4], | ||
| num_outs=4): |
There was a problem hiding this comment.
We may support norm_cfg and act_cfg.
mmseg/models/necks/upsample_neck.py
Outdated
| size=list((np.array(inputs[i].shape[2:]) * | ||
| self.scales[i]).astype(int)), |
There was a problem hiding this comment.
We may use scale argument.
mmseg/models/necks/upsample_neck.py
Outdated
|
|
||
|
|
||
| @NECKS.register_module() | ||
| class UpsampleNeck(nn.Module): |
There was a problem hiding this comment.
We may rename it to something like MultiLevelNeck since we would also perform downsampling.
| Args: | ||
| in_channels (List[int]): Number of input channels per scale. | ||
| out_channels (int): Number of output channels (used at each scale). | ||
| scales (List[int]): Scale factors for each input feature map. |
There was a problem hiding this comment.
We fix this in the new PR.
* init * upsample v1.0 * fix errors * change to in_channels list * add unittest, docstring, norm/act config and rename Co-authored-by: xiexinch <test767803@foxmail.com>
* [Download] Smart downloading * add test * finish test * update * make style
…s so that they will be ignored in modelzoo statistics (open-mmlab#512)
A neck structure connect vit backbone and decoder_heads.
Default output of vit backbone is a downsample 16x feature map, we design a structure that convert the feature map to [32x, 16x, 8x, 4x] feature map list.