search

ucbdrive / few-shot-object-detection

Implementations of few-shot object detection benchmarks
Apache License 2.0
1.1k stars 223 forks source link

The issue about base train and novel train #72

Closed myh12138 closed 3 years ago

myh12138 commented 3 years ago

During I train my own dataset,I met some problems, my owm dataset only has two classes. I have only one GPU 2080Ti , therefore I adjust the bachsize. However the result is so bad, the mAP is almost 0.1. Also I find this program not has novel1_1shot.yaml in voc , threfore i copy one form coco and make some change , is there any problem with this? I hope anyone can answer my questions.

zyxfater commented 3 years ago

Also I find this program not has novel1_1shot.yaml in voc

jinweiLiu commented 3 years ago

Did you find a solution? Where is the 'novel1_1shot.yaml' in voc? Look forward to your reply.

g-lyc commented 3 years ago

I have same problem，no solution. T.T

a07913838438 commented 3 years ago

no solution

thomasehuang commented 3 years ago

Apologies for the late response.

Regarding training on one GPU, the mAP being 0.1 is a little weird. You can reopen the issue and let me know if you still have this issue.

Regarding the novel file for VOC, see my answer in this issue. The reason why it's missing is because we don't do that step for VOC. I will update the README to make it more clear, sorry for the confusion.