[WIP][Feature] Support RTMDet-Ins fast training #649
Conversation
RangiLyu
changed the title
| persistent_workers = True | ||
|
|
||
| # -----train val related----- | ||
| # Base learning rate for optim_wrapper. Corresponding to 8xb16=64 bs |
There was a problem hiding this comment.
| # Base learning rate for optim_wrapper. Corresponding to 8xb16=64 bs | |
| # Base learning rate for optim_wrapper. Corresponding to 8xb32=256 bs |
| multi_label=True, | ||
| # The number of boxes before NMS | ||
| nms_pre=1000, | ||
| score_thr=0.05, # Threshold to filter out boxes. |
There was a problem hiding this comment.
这里可以加个注释说明：实例分割任务相比目标检测后处理速度更慢，因此需要加大 score_thr 和减少 nms_pre 和 max_per_img 等参数
| dict(type='mmdet.PackDetInputs') | ||
| ] | ||
|
|
||
| train_pipeline_stage2 = [ |
There was a problem hiding this comment.
可以用简化写法： _base_.train_pipeline_stage2[2].ratio_range=random_resize_ratio_range
| for i, (weight, bias) in enumerate(zip(weights, biases)): | ||
| x = F.conv2d( | ||
| x, weight, bias=bias, stride=1, padding=0, groups=num_inst) | ||
| x = torch.einsum('nij,njhw->nihw', weight, x) |
There was a problem hiding this comment.
我们可以保留 F.conv2d( 的实现，并注释掉。说明新的写法更快？
There was a problem hiding this comment.
conv2d 的 kernel shape 和 einsum 有点区别
| for i, (weight, bias) in enumerate(zip(weights, biases)): | ||
| x = F.conv2d( | ||
| x, weight, bias=bias, stride=1, padding=0, groups=num_inst) | ||
| x = torch.einsum('nij,njhw->nihw', weight, x) |
| @@ -0,0 +1,340 @@ | |||
| _base_ = ['../_base_/default_runtime.py'] | |||
There was a problem hiding this comment.
配置应该是放到 config/rtmdet/ins_seg 下比较好？更好管理？你觉得呢
|
Is this still a planned feature or WIP? Would love to be able to train the RTMDet-Ins model but I'm not 100% sure what changes I need to make. It seems like part of this PR has been merged for YOLOv5-Ins but part has not, so I'm not sure what the plan is here. Any update here would be much appreciated! Also related to #853. |
|
@RangiLyu hi, may I ask what is the current status of this feature? Look forward to seeing the fast training for Rtmdet-instance-segmentation |
Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.
Motivation
Training RTMDet-Ins-s costs 19h. It is 2.8 times faster than MMDetection.