Iteration Plan v1.3.7 - June 2021

[ZwwWayne]

MMCV v1.3.6 has been released, see the plan and the release notes
We list our plan for v1.3.7, which will be released on June 26th.
The PRs from the communities will also be listed here.
Thus, developers could comprehensively know the plan and the progress.

Features:

• Porting mmcv for hip [Feature] Porting mmcv for hip #1022
• Support LR_ReduceOnPlateau [WIP] Support LR_ReduceOnPlateau #1033
• Support variables in base files for configs Support variables in base files for configs #1083

Enhancements:

• Add Windows CI [Feature] Add windows CI #1023
• Use more cpus in building Use more cpus in building #1046
• Decreasing the cycle amplitude after each period in cyclic lr decreasing the cycle amplitude after each period in cyclic lr #783
• Support image reading while ignoring EXIF orientation info Support image reading while ignoring EXIF orientation info #1091
• Optimize the PyTorch CUDA implementation for Criss Cross Attention Optimize the PyTorch CUDA implementation for Criss Cross Attention #1088

If you have created a PR and wish it to be merged in this version, feel free to leave a message in this issue. We will update our plan accordingly here.

[yeliudev]

Hi @ZwwWayne! May I ask that will the ConvModule inherit from BaseModule in the newer version of mmcv? Currently, the ConvModule may not be properly initialized using init_cfg.

class Net(BaseModule):

    def __init__(self, init_cfg=dict(type='Constant', layer='Conv2d', val=0)):
        super(Net, self).__init__(init_cfg)
        self.conv_module = ConvModule(128, 128, 3)

In this case, self.conv_module will be finally initialized by its own method (kaiming normal) but not constant since ConvModule can not determine whether itself has been initialized using self._is_init.

[zhouzaida]

Hi @ZwwWayne! May I ask that will the ConvModule inherit from BaseModule in the newer version of mmcv? Currently, the ConvModule may not be properly initialized using init_cfg.

class Net(BaseModule):

    def __init__(self, init_cfg=dict(type='Constant', layer='Conv2d', val=0)):
        super(Net, self).__init__(init_cfg)
        self.conv_module = ConvModule(128, 128, 3)

In this case, self.conv_module will be finally initialized by its own method (kaiming normal) but not constant since ConvModule can not determine whether itself has been initialized using self._is_init.

@MeowZheng

[MeowZheng]

Many thanks for you reporting this issue. It is a little difficult, as the inner parts have higher priority, but we will solve it asap.