Add BasicVSR #245
Add BasicVSR #245
Conversation
Codecov Report
@@ Coverage Diff @@
## master #245 +/- ##
==========================================
+ Coverage 81.31% 81.35% +0.04%
==========================================
Files 156 158 +2
Lines 7503 7719 +216
Branches 1102 1141 +39
==========================================
+ Hits 6101 6280 +179
- Misses 1271 1296 +25
- Partials 131 143 +12
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
Sure, I will change it to |
|
|
||
| def __init__(self, mid_channels=64, num_blocks=30, spynet_pretrained=None): | ||
|
|
||
| super(BasicVSRNet, self).__init__() |
| """ | ||
|
|
||
| def __init__(self, in_channels, out_channels=64, num_blocks=30): | ||
| super(ResidualBlocksWithInputConv, self).__init__() |
| """ | ||
|
|
||
| def __init__(self, pretrained): | ||
| super(SPyNet, self).__init__() |
mmedit/models/restorers/basicvsr.py
Outdated
| train_cfg=None, | ||
| test_cfg=None, | ||
| pretrained=None): | ||
| super(BasicVSR, self).__init__(generator, pixel_loss, train_cfg, |
mmedit/models/restorers/basicvsr.py
Outdated
| self.generator.find_unused_parameters = False | ||
|
|
||
| # count training steps | ||
| self.step_counter = train_cfg.get('step_counter', |
There was a problem hiding this comment.
Proposal:
Add a TAIL component to the module, just like backbones and heads, and save step counters there, so that they could be automatically saved and restored.
mmedit/models/restorers/basicvsr.py
Outdated
|
|
||
| # Note: For Vimeo-90K, we use mirror extension. Hence, the output | ||
| # sequence has 14 frames. Only center frame is kept for Vimeo-90K. | ||
| if output.size(1) == 14: |
There was a problem hiding this comment.
I don't think the hard-coded 14 is good
|
Looks good generally. Some minor comments. |
* Add BasicVSR * fix isort * fix isort * rename to basicvsr_net.py * change step_counter to a key in train_cfg * rename and fix minor bug * rename to BasicVSRNet * Fix init and model
BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond