Learning rate shows an abnormal changing trend and AP =0 while training queryinst model with my own dataset about queryinst HOT 3 CLOSED

Hi! Thanks for your interest on our work.
The first issue is normal for QueryInst. QueryInst is trained with two learning rate adjustment strategies: warmup and step down. The warmup strategy provides an additive learning rate for the first few iterations (~1K iters) and is set up here. The step down strategy decreases learning at certain epochs and it was set up here. So in the total training process, learning rate will firstly increase to the default setting within ~1k iterations and then decrease by 10 at 27th and 33rd epochs.
For the second issue, the training log shows that there seems no problem during your training process. So I suggest you check out your evaluation and inference code to find the way out.

from queryinst.

@vealocia Thanks for your comment :) ! I do have a deeper understanding for the learning rate adjustment strategies of QueryInst. That helps a lot.

However, for the second issue, I still can't figure it out. My evaluation and inference code remain as the original version of GitHub repository, I didn't change anywhere of them after I downloaded the code from Github. Based on a lot of experiments, I noticed that : Bbox AP and Bbox AR usually perform well with high MAP values, but Segms AP and AR still keep as zeros all the time.

For example: 2021-08-25 12:30:34,516 - mmdet - INFO - Epoch(val) [30][385] bbox_mAP: 0.8950, bbox_mAP_50: 1.0000, bbox_mAP_75: 0.9650, bbox_mAP_s: 0.7940, bbox_mAP_m: 0.9260, bbox_mAP_l: 0.9100, bbox_mAP_copypaste: 0.895 1.000 0.965 0.794 0.926 0.910, segm_mAP: 0.0000, segm_mAP_50: 0.0000, segm_mAP_75: 0.0000, segm_mAP_s: 0.0000, segm_mAP_m: 0.0000, segm_mAP_l: 0.0000, segm_mAP_copypaste: 0.000 0.000 0.000 0.000 0.000 0.000

I attached the whole log below just in case. Thanks!
20210825_090242.log

from queryinst.

Here, we only deal with issues that cannot reproduce the results reported in our paper. It is unlikely for us to solve other problems.

I'm closing this issue but let us know if you have any further questions related to reproducing the COCO results

from queryinst.