[Refactor]: Unified parameter initialization #622
Conversation
|
Need to resolve conflicts. FCOS3D seems to have low accuracy. |
| 'please use "init_cfg" instead') | ||
| self.init_cfg = dict(type='Pretrained', checkpoint=pretrained) | ||
| else: | ||
| self.init_cfg = dict(type='Kaiming', layer='Conv2d') |
There was a problem hiding this comment.
According to L66-67, why use Kaiming init here?
There was a problem hiding this comment.
if we dont use the pretrained model, we want to use kaiming_init to init all Conv2d layer.
| self.init_cfg = dict( | ||
| type='Kaiming', | ||
| layer='Conv2d', | ||
| override=dict(type='Kaiming', name='heatmap')) |
There was a problem hiding this comment.
I think override can be deleted, as you have init heatmap in init_weights.
There was a problem hiding this comment.
Yes, I forget to delete it. thank you!
| layer='Conv2d', | ||
| override=dict( | ||
| type='Kaiming', | ||
| name=self.cls_head[-1], |
There was a problem hiding this comment.
the value of name is attribute name. I suggest to rewrite init_weights for initializing cls_head.
mmdet3d/__init__.py
Outdated
| @@ -17,7 +17,7 @@ def digit_version(version_str): | |||
| return digit_version | |||
|
|
|||
|
|
|||
| mmcv_minimum_version = '1.3.1' | |||
| mmcv_minimum_version = '1.3.5' | |||
mmdet3d/__init__.py
Outdated
| @@ -27,17 +27,17 @@ def digit_version(version_str): | |||
| f'MMCV=={mmcv.__version__} is used but incompatible. ' \ | |||
| f'Please install mmcv>={mmcv_minimum_version}, <={mmcv_maximum_version}.' | |||
|
|
|||
| mmdet_minimum_version = '2.10.0' | |||
| mmdet_maximum_version = '2.11.0' | |||
| mmdet_minimum_version = '2.12.0' | |||
mmdet3d/__init__.py
Outdated
| mmseg_minimum_version = '0.14.0' | ||
| mmseg_maximum_version = '0.14.0' | ||
| mmseg_minimum_version = '0.14.1' | ||
| mmseg_maximum_version = '0.15.0' |
There was a problem hiding this comment.
may set 1.0.0 for now. Usually there is not many BC-breakings.
Codecov Report
@@ Coverage Diff @@
## master #622 +/- ##
==========================================
- Coverage 51.01% 48.80% -2.22%
==========================================
Files 204 208 +4
Lines 15383 15858 +475
Branches 2488 2542 +54
==========================================
- Hits 7848 7739 -109
- Misses 7017 7625 +608
+ Partials 518 494 -24
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
docs/compatibility.md
Outdated
| @@ -4,6 +4,15 @@ This document provides detailed descriptions of the BC-breaking changes in MMDet | |||
|
|
|||
| ## MMDetection3D 0.15.0 | |||
|
|
|||
| ### Unified parameter initialization | |||
|
|
|||
| To unify the parameter initialization in OpenMMLab projects, MMCV supports `BaseModule` that accepts `init_cfg` to allow the modules' parameters initialized in a flexible and unified manner. Now the users need to explicitly call `model.init_weights()` in the training script to initialize the model (as in [here](https://github.com/open-mmlab/mmdetection3d/blob/master/tools/train.py#L183), previously this was handled by the detector. Please refer to PR #378 for details. | |||
tests/test_data/test_pipelines/test_augmentations/test_transforms_3d.py
Outdated
Show resolved
Hide resolved
Motivation
Refactoring model initialization methods based MMCV. For more information, please refer to open-mmlab/mmcv#780
Dependency
Training results