[Feature] Support ConvNeXt-V2.

[RangiLyu]

Support ConvNeXt-V2

This PR gives an example of using ConvNeXt-V2 in mmdetection.

It requires the latest version of mmclassification-1.x.

Notice

The official finetuning setting of object detection is not released yet.

There are at least three settings unknown:

• What is stochastic depth rate {0.1, 0.2, 0.3, 0.4} mean?
• Does RPN use soft_nms?
• What is sweep layer-wise lr decay?

Training result:

01/14 20:43:26 - mmengine - �[4m�[37mINFO�[0m - Evaluating bbox...
Loading and preparing results...
DONE (t=0.94s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=39.44s).
Accumulating evaluation results...
DONE (t=6.56s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.529
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.727
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.587
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.359
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.573
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.670
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.694
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.694
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.694
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.523
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.742
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.836
01/14 20:44:14 - mmengine - �[4m�[37mINFO�[0m - bbox_mAP_copypaste: 0.529 0.727 0.587 0.359 0.573 0.670
01/14 20:44:14 - mmengine - �[4m�[37mINFO�[0m - Evaluating segm...
Loading and preparing results...
DONE (t=2.69s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *segm*
DONE (t=43.78s).
Accumulating evaluation results...
DONE (t=6.45s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.464
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.699
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.508
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.273
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.498
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.649
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.601
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.601
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.601
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.436
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.639
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.750
01/14 20:45:11 - mmengine - �[4m�[37mINFO�[0m - segm_mAP_copypaste: 0.464 0.699 0.508 0.273 0.498 0.649
01/14 20:45:12 - mmengine - �[4m�[37mINFO�[0m - Epoch(val) [36][625/625] coco/bbox_mAP: 0.5290 coco/bbox_mAP_50: 0.7270 coco/bbox_mAP_75: 0.5870 coco/bbox_mAP_s: 0.3590 coco/bbox_mAP_m: 0.5730 coco/bbox_mAP_l: 0.6700 coco/segm_mAP: 0.4640 coco/segm_mAP_50: 0.6990 coco/segm_mAP_75: 0.5080 coco/segm_mAP_s: 0.2730 coco/segm_mAP_m: 0.4980 coco/segm_mAP_l: 0.6490

The box AP is aligned with the paper, but the mask AP has a gap of 0.2%.

[yuangpeng]

Could you please provide your reproduction log or result?

[RangiLyu]

Could you please provide your reproduction log or result?

I trained it last weekend. The box AP is aligned with the paper, but the mask AP has a gap of 0.2%.
here is the log:
20230113_110947.log

[yuangpeng]

Could you please provide your reproduction log or result?

I trained it last weekend. The box AP is aligned with the paper, but the mask AP has a gap of 0.2%. here is the log: 20230113_110947.log

Thanks! I just noticed my problem, because I only reproduced the cascade mask rcnn convnext v1 base, the result of box AP 52.35 is my baseline. I keep the same config and change the backbone and pretrained weight to v2, then I only get the box AP 52.48. Now I am curious whether v2 has little improvement on cascade mask rcnn.