STAMP静态推理复现

[zzm000]

作者你好，请问在stamp的静态推理复现中，我复现出来的效果并不理想，下面是我实验运行的步骤，请问是有哪里出现了问题吗
mkdir opencood/logs/Baselines/stage1/OPV2V_m0_v2xvit
mkdir opencood/logs/Baselines/stage1/OPV2V_m1_v2xvit
mkdir opencood/logs/Baselines/stage1/OPV2V_m3_v2xvit

cp opencood/hypes_yaml/opv2v/GenComm_yamls/baselines/stage1/m0_v2xvit.yaml opencood/logs/Baselines/stage1/OPV2V_m0_v2xvit/config.yaml
cp opencood/hypes_yaml/opv2v/GenComm_yamls/baselines/stage1/m1_v2xvit.yaml opencood/logs/Baselines/stage1/OPV2V_m1_v2xvit/config.yaml
cp opencood/hypes_yaml/opv2v/GenComm_yamls/baselines/stage1/m3_v2xvit.yaml opencood/logs/Baselines/stage1/OPV2V_m3_v2xvit/config.yaml

python opencood/tools/train.py -y None --model_dir opencood/logs/Baselines/stage1/OPV2V_m0_v2xvit/
python opencood/tools/train.py -y None --model_dir opencood/logs/Baselines/stage1/OPV2V_m1_v2xvit/
python opencood/tools/train.py -y None --model_dir opencood/logs/Baselines/stage1/OPV2V_m3_v2xvit/
== ==
python opencood/tools/heal_tools.py merge_and_save
opencood/logs/Baselines/stage1/OPV2V_m3_v2xvit
opencood/logs/Baselines/stage1/OPV2V_m0_v2xvit
opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m3_v2xvit
python opencood/tools/heal_tools.py merge_and_save
opencood/logs/Baselines/stage1/OPV2V_m1_v2xvit
opencood/logs/Baselines/stage1/OPV2V_m0_v2xvit
opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit

mkdir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit
mkdir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m3_v2xvit

cp opencood/hypes_yaml/opv2v/GenComm_yamls/baselines/stage2/STAMP/m0m1_v2xvit.yaml opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit/config.yaml
cp opencood/hypes_yaml/opv2v/GenComm_yamls/baselines/stage2/STAMP/m0m3_v2xvit.yaml opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit/config.yaml

python opencood/tools/train_stamp.py -y None --model_dir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit
python opencood/tools/train_stamp.py -y None --model_dir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m3_v2xvit

python opencood/tools/heal_tools.py merge_and_save opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m3_v2xvit opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit opencood/logs/Baselines/stage2/STAMP/OPV2V_m1m3_v2xvit_infer

python opencood/tools/inference.py --model_dir opencood/logs/Baselines/stage2/STAMP/OPV2V_m1m3_v2xvit_infer

[jeffreychou777]

你好 如果stage1训练出来的模型是正确的，在stage2训练stamp时应该收敛比较快。可以观察训练stamp时的loss，如果正常收敛，考虑可能是inference前checkpoint合并出了问题，如果stamp训练不收敛，考虑stage1训练问题或者stage2前checkpoints合并问题。

[zzm000]

您好，我推理了stage1的模型结果，应该是正常的，在global_sort的AP50下，m0为0.86，m1为0.92，m3为0.88，我在stage2训练STAMP时似乎确实没有正常收敛，m0m1的Loss在3.多，m0m3的 Loss在30多，应该是checkpoint合并出了问题？

在stamp的融合中，是否应该是使用merge_and_save_fianl，并入下述方法合并ckp？（m1为ego）
python opencood/tools/heal_tools.py merge_and_save
m0_v2xvit
m1_v2xvit
STAMP/m0m1_v2xvit
python opencood/tools/train_stamp.py -y None --model_dir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m1_v2xvit

python opencood/tools/heal_tools.py merge_and_save
m0_v2xvit
m3_v2xvit
STAMP/m0m3_v2xvit
python opencood/tools/train_stamp.py -y None --model_dir opencood/logs/Baselines/stage2/STAMP/OPV2V_m0m3_v2xvit

python opencood/tools/heal_tools.py merge_and_save
m0m3_v2xvit
m0m1_v2xvit
STAMP/m1m3_v2xvit_infer