Kei-Chi Tse

[20220814_041356.log](https://github.com/open-mmlab/mmdetection/files/9374707/20220814_041356.log) [20220816_015156.log](https://github.com/open-mmlab/mmdetection/files/9374711/20220816_015156.log) ### [Update]: Experiment logs of _"res50"_ and _"res50+3 patterns"_ are uploaded. Compared to paper, as for _res50_, current mAP is **41.8** (**-0.4**) with **80** classes. As for _res50+3...

### **[Update]:** Previous experiments set dropout to 0.1, _**misaligned**_ with **official repo** as they use _**dropout 0.0**_ , which is proved to be beneficial in enhancing mAP in Sparse R-CNN...

### [Update]: Pretrained checkpoint of this version will be available **_after_** the refactor work is done in upcoming MMDetection 3.0. For details, please refer to [#8811](https://github.com/open-mmlab/mmdetection/pull/8811) and [#8763](https://github.com/open-mmlab/mmdetection/pull/8763).

## Short-term Plan 1. Modify code based on review. 2. _**ResNet**_ 50 mAP alignment with official repo.

### Update: **Forward results** and **losses** for a **batch** are **perfectly aligned** (_4 decimal places_) with **2.0** version and **official** repo!