Is there something wrong with the original cfg file, please?

[ZhiXianZ]

Command: sh scripts/meta_training_pascalvoc_split1_resnet101.sh

ValueError: Milestone must be smaller than total number of updates: num_updates=10000, milestone=10000

version: 0.5

cfg File:
SOLVER:
IMS_PER_BATCH: 4
BASE_LR: 0.002
STEPS: (15000, 20000)
MAX_ITER: 20000
CHECKPOINT_PERIOD: 10000

Has the author encountered this problem?

Environment:
Ubuntu18+torch1.8+cuda11.0 detetron2-v0.5

[GuangxingHan]

Hi,
The default cfg files work well on my local machine.
It seems that the error might be related to the software version.
Actually I used an old version of pytorch (1.6.0) and detectron (0.2.1) and have not tested our codes under other versions.
Hope this will be useful to you.
Guangxing

[ZhiXianZ]

Hi, The default cfg files work well on my local machine. It seems that the error might be related to the software version. Actually I used an old version of pytorch (1.6.0) and detectron (0.2.1) and have not tested our codes under other versions. Hope this will be useful to you. Guangxing

OK. I also asked the same question on Detectron2, maybe I need to find other way to solve this problem. By the way, Is the other paper (Meta Faster R-CNN) also the same configuration and environment?

[GuangxingHan]

OK. I also asked the same question on Detectron2, maybe I need to find other way to solve this problem. By the way, Is the other paper (Meta Faster R-CNN) also the same configuration and environment?

Yes, we used the same environment in the two repos.

[ZhiXianZ]

OK. I also asked the same question on Detectron2, maybe I need to find other way to solve this problem. By the way, Is the other paper (Meta Faster R-CNN) also the same configuration and environment?

Yes, we used the same environment in the two repos.

Thank you for your reply and good luck with your work.

[ZhiXianZ]

Excuse me, I have another problem:
What's the difference between fsod_train_net_fewx.py and fsod_train_net.py?
They look so similar, can I keep only one line of commands in meta_training_pascalvoc_split1_resnet101.sh?

[GuangxingHan]

Excuse me, I have another problem: What's the difference between fsod_train_net_fewx.py and fsod_train_net.py? They look so similar, can I keep only one line of commands in meta_training_pascalvoc_split1_resnet101.sh?

Both of the two scripts are crucial.

We first use fsod_train_net_fewx.py to train the baseline model following this repo FewX, which is reorganized in our fewx module.

Then we add the proposed heterogeneous GCNs and use fsod_train_net.py to train the whole model, which is defined in our QA_FewDet module.

The two modules fewx and QA_FewDet are different, and the two-step meta-training is crucial for our final performance. If we only use the QA_FewDet module, the training is unstable.

[ZhiXianZ]

Excuse me, I have another problem: What's the difference between fsod_train_net_fewx.py and fsod_train_net.py? They look so similar, can I keep only one line of commands in meta_training_pascalvoc_split1_resnet101.sh?

Both of the two scripts are crucial.

We first use fsod_train_net_fewx.py to train the baseline model following this repo FewX, which is reorganized in our fewx module.

Then we add the proposed heterogeneous GCNs and use fsod_train_net.py to train the whole model, which is defined in our QA_FewDet module.

The two modules fewx and QA_FewDet are different, and the two-step meta-training is crucial for our final performance. If we only use the QA_FewDet module, the training is unstable.

Thank you for your excellent work and your reply.