The problem of training result！

[qqqqhhqq]

When I train the precls of SHA_GCL on single card RTX3090, my batch_size is set to 8 and Solver.max_iter is set to 70000, the final result on the test cannot reach the experimental result in the paper, can you help me to answer the question?

[qqqqhhqq]

My training script nohup python tools/relation_train_net.py --config-file "configs/SHA_GCL_e2e_relation_X_101_32_8_FPN_1x.yaml" \ GLOBAL_SETTING.DATASET_CHOICE 'VG' \ GLOBAL_SETTING.RELATION_PREDICTOR 'TransLike_GCL' \ GLOBAL_SETTING.BASIC_ENCODER 'Hybrid-Attention' \ GLOBAL_SETTING.GCL_SETTING.GROUP_SPLIT_MODE 'divide4' \ GLOBAL_SETTING.GCL_SETTING.KNOWLEDGE_TRANSFER_MODE 'KL_logit_TopDown' \ MODEL.ROI_RELATION_HEAD.USE_GT_BOX True \ MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL True \ SOLVER.IMS_PER_BATCH 8 \ TEST.IMS_PER_BATCH 2 \ DTYPE "float16" \ SOLVER.MAX_ITER 70000 \ SOLVER.VAL_PERIOD 5000 \ SOLVER.CHECKPOINT_PERIOD 5000 \ GLOVE_DIR /home/new702/disk2/QH/scene_graph/datasets/vg/glove \ OUTPUT_DIR ./output/SHA_GCL_VG_PredCls_test

[dongxingning]

Thanks for your question. I guess this performance gap may derive from the difference of GPU and hardware setup. Besides, we find a normal phenomenon that the performance of mR@k may fluctuates around about 1%. Anyway, it is encouraged to use your reproduced result in your comparison if needed.

[qqqqhhqq]

Thank you for your reply, and thank you for your contribution to the community. I will take your comments and then do some comparative experiments with my reproduction as a benchmark.

[qqqqhhqq]

I have modified the structure of the model, but during training, the value 'nan' will appear

[liuzhenqi1999]

KeyError: 'Non-existent config key: MODEL.ROI_BOX_HEAD.VG_NUM_CLASSES'，What is the reason for this，Have you encountered it