[WIP] Xception Backend #3383
[WIP] Xception Backend #3383
Conversation
|
|
| """Xception optimized for the ImageNet dataset, as specified in | ||
| https://arxiv.org/pdf/1610.02357.pdf.""" | ||
|
|
||
| def __init__(self, |
There was a problem hiding this comment.
- Please add complete docstrings for all newly added modules and functions as those in MMDetection V2.x.
mmdet/models/backbones/xception.py
Outdated
| self.conv2.out_channels) | ||
| # do relu here | ||
|
|
||
| self.block1 = Block( |
There was a problem hiding this comment.
- Use
self.layer1,self.layer2, which are more consistent with ResNets. - Can initialize the blocks/stage layers in a more general form as that in ResNet?
|
|
||
| def _init_weights(self, ): | ||
| for module in self.modules(): | ||
| if isinstance(module, nn.Conv2d): |
There was a problem hiding this comment.
Use the implemented initialization methods (constant_init, kaiming_init) implemented in MMCV, unless the initialization is special and not implemented in MMCV. If so, please also add comments in the code as a reminder.
|
Something to do before merge:
|
|
@ZwwWayne Thank you for the review. I have addressed partly your comments for the moment. I will do some more once I get more time. Thanks once again! |
|
Hi @roya0045！We are grateful for your efforts in helping improve this open-source project during your personal time. |
Draft PR for a Xception backend proposal