search

YoungXIAO13 / FewShotDetection

(ECCV 2020) PyTorch implementation of paper "Few-Shot Object Detection and Viewpoint Estimation for Objects in the Wild"
http://imagine.enpc.fr/~xiaoy/FSDetView/
MIT License
212 stars 33 forks source link

coco training time #6

Closed hzhupku closed 4 years ago

hzhupku commented 4 years ago

hi, i use your script run/train_coco.sh to perform base training on coco dataset. But the epoch is set to be 21, batch_size as 4, which makes training time very large, since coco is a large dataset. I am wondering about this setting.

YoungXIAO13 commented 4 years ago

Yes, it indeed makes the training much longer. I didn't try sth specific on tuning the learning scheduel on different datasets. But I don't think keeping the longer training time should be a big issue as there are much richer training samples in coco, thus require more optimization iterations to achieve a good perf.

hzhupku commented 4 years ago

May I ask about your training time on coco? Since it cost me almost two weeks to finish the training(1 TITANXP).

YoungXIAO13 commented 4 years ago

It alos took me ~10 days to finish the whole training on COCO with one TitanX-Pascal, which is indeed very long. As I don't have many GPUs available in hands, I did not try parallelize it with multiple GPUs, but it should be ~K times faster if you have K GPUs. For other recently published code on few-shot object detection, you can look at TFA and MPSR to get a more efficient implementations.