low performance on a single GPU

[squirrel16]

I run python train.py --gpu 0 on a 6G TITAN Z, and in epoch240, only achieve mAP 73.7.(vgg16_reduced,300x300,voc07+12trainval, voc07 test)

Did this result from my single GPU? Should I use 4 GPU instead? But it is said that

perfect training parameter set is yet to be discovered for multi-gpu

I wonder under what condition,you achieve MAP77.9？Thanks a lot!

[edmBernard]

hyperparameter change in function of multi-gpu or not, batch size etc ... If you want to achieve the best MAP on different configuration, you need to adjust them.

[ghost]

@squirrel16 The same problem. I reproduced ssd 300 with vgg16_reduced and achieved mAP 73.1, using 4-gpus. Then I downloaded the released ssd 300 model, evaluated on voc07 using legacy symbol, and achieved only 75.7 (compared with published 77.8). As the author said, the vgg16 is broken. You may try reproducing inception-ssd and share your result.

Thanks.

[bowenc0221]

I had the same problem, mAP is about 73.7 trained on a single P100 with CUDA8 and using the default setting. Did anyone figure out the correct hyperparameter?