facebookresearch / Detic

Code release for "Detecting Twenty-thousand Classes using Image-level Supervision".
Apache License 2.0
1.88k stars 210 forks source link

Can not reproduce the result on LVIS box supervised training #85

Closed YifanXu74 closed 2 years ago

YifanXu74 commented 2 years ago

Hi! I run the following command: python train_net.py --config-file configs/BoxSup-C2_Lbase_CLIP_R5021k_640b64_4x.yaml --num-gpus 8 and get

^[[32m[10/23 17:57:34 detectron2]: ^[[0mEvaluation results for lvis_v1_val in csv format:
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: Task: bbox
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: 11.6125,18.8525,11.9556,8.6461,15.4798,18.7985,4.9094,9.6765,16.7192
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: Task: segm
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
^[[32m[10/23 17:57:34 d2.evaluation.testing]: ^[[0mcopypaste: 10.4488,17.3474,10.6957,7.4254,14.9531,17.8749,4.7808,8.9843,14.5744

It is far from the officially trained one:

[10/23 17:22:49 detectron2]: Evaluation results for lvis_v1_val in csv format:
[10/23 17:22:49 d2.evaluation.testing]: copypaste: Task: bbox
[10/23 17:22:49 d2.evaluation.testing]: copypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
[10/23 17:22:49 d2.evaluation.testing]: copypaste: 33.6902,46.4293,35.4045,24.4995,43.3657,49.3483,17.3725,34.3475,40.1283
[10/23 17:22:49 d2.evaluation.testing]: copypaste: Task: segm
[10/23 17:22:49 d2.evaluation.testing]: copypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
[10/23 17:22:49 d2.evaluation.testing]: copypaste: 30.0988,44.0328,31.9440,20.6293,41.1902,46.9922,16.2368,30.9377,35.2550

Here is my training log: link

I am not sure what goes wrong here. Thanks for the help! Best regards!

YifanXu74 commented 2 years ago

Hi! I found that I have loaded a wrong ResNet 50 pre-trained checkpoint when training LVIS. When loading resnet50_miil_21k.pkl, the problem is solved. Thanks~

YiZhiTangZong commented 3 months ago

Hi! I found that I have loaded a wrong ResNet 50 pre-trained checkpoint when training LVIS. When loading resnet50_miil_21k.pkl, the problem is solved. Thanks~

It might be a stupid question... but where can I get that resnet50_miil_21k.pkl? I could only find the pth version and failed to convert it.