Can not replicate the final result.

[JiuqingDong]

Hi, I run this ' python train_net.py --num-gpus 3 --config-file configs/VLDet_OVCOCO_CLIP_R50_1x_caption.yaml' Here is the result. The result is lower than yours.

[03/08 12:51:07] d2.evaluation.coco_evaluation INFO: Seen bbox AP50: 45.93429908569812 [03/08 12:51:07] d2.evaluation.coco_evaluation INFO: Unseen bbox AP50: 28.641015836555734 [03/08 12:51:07] detectron2 INFO: Evaluation results for coco_generalized_del_val in csv format: [03/08 12:51:07] d2.evaluation.testing INFO: copypaste: Task: bbox [03/08 12:51:07] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [03/08 12:51:07] d2.evaluation.testing INFO: copypaste: 24.7339,41.4114,25.6790,11.3174,27.5064,34.1043

Batch : 24 gpu: 3

Could you give me some suggestions?

[clin1223]

Hi,

Thanks for your interest. The main results are trained on 8 GPUs. 3 GPUs and small batchsize may drop the performance.

[JiuqingDong]

Hi, thank you for your code. I have one more question that is your text encoder frozen or learnable?

[clin1223]

It's frozen.

[JiuqingDong]

Hi, how to do this ablation study? I can not replicate your result. Maybe I do a wrong experiment like 'one to many'?

[dreamwish1998]

@clin1223 @JiuqingDong When I was doing base training on the LVIS dataset, I encountered the following error. I used the configuration file "Base-C2_L_R5021k_640b64_4x.yaml". Have you ever encountered this problem? Or, were you able to run the training script successfully?

-- Process 1 terminated with the following error: Traceback (most recent call last): File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/multiprocessing/spawn.py", line 59, in _wrap fn(i, args) File "/home/weixj/VLDet/detectron2/detectron2/engine/launch.py", line 126, in _distributed_worker main_func(args) File "/home/weixj/VLDet/train_net.py", line 246, in main do_train(cfg, model, resume=args.resume) File "/home/weixj/VLDet/train_net.py", line 154, in do_train loss_dict = model(data) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, kwargs) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/parallel/distributed.py", line 799, in forward output = self.module(*inputs[0], *kwargs[0]) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(input, kwargs) File "/home/weixj/VLDet/vldet/modeling/meta_arch/custom_rcnn.py", line 169, in forward ann_type=ann_type, classifier_info=classifier_info) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, kwargs) File "/home/weixj/VLDet/vldet/modeling/roi_heads/vldet_roi_heads.py", line 217, in forward ann_type=ann_type, classifier_info=classifier_info) File "/home/weixj/VLDet/vldet/modeling/roi_heads/vldet_roi_heads.py", line 121, in _forward_box ann_type=ann_type, classifier_info=classifier_info) File "/home/weixj/VLDet/vldet/modeling/roi_heads/vldet_roi_heads.py", line 331, in _run_stage classifier_info=classifier_info) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, *kwargs) File "/home/weixj/VLDet/vldet/modeling/roi_heads/vldet_fast_rcnn.py", line 573, in forward cls_scores, proj_region, zs_weight = self.cls_score(x, ann_type) File "/home/weixj/anaconda3/envs/VLDet/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(input, kwargs) TypeError: forward() takes 2 positional arguments but 3 were given

[JiuqingDong]

Sorry, I didn't train on the Lvis dataset. I just try to use COCO and I got a result. It is lower than the author's