search

zjhuang22 / maskscoring_rcnn

Codes for paper "Mask Scoring R-CNN".
MIT License
1.9k stars 378 forks source link

pretrained ms_rcnn model #21

Open fanyix opened 5 years ago

fanyix commented 5 years ago

Thanks for the great work! It seems the link provided in README is for ImageNet pretrained models? Could you please provide the R-50/R-101 MS-RCNN model that is used to produce your results in the paper?

Csysmile commented 5 years ago

Sorry, I have a same question, how to find the R-50/R-101 MS-RCNN model?

huanhuancao commented 5 years ago

@fanyix @Csysmile Hello, are you meaning that the pkl files the author provides in readme can not directly used to test? Have you run test_net.py and got similar results with the paper?

Csysmile commented 5 years ago

@maomaochongchh Yes , the the pkl can not be download, I found the R-50 in other way. However I have another problem now. After a few minutes of training, the error will be reported as "the shape pf the mask[261888] at index 0 does not match the indexed tensor [0] at index 0". l put my data in maskrcnn-benchmark as the author mentions in the end and it works normally. Have you meet a similar problem?

huanhuancao commented 5 years ago

sorry, I have not. When I noticed that it needs 15 days to train it with 4 gpus, I stopped it. I just want to run the test_net.py now. If you run it successfully, please tell me, thank you, and if I meet the similar problem later, I will tell you. But I suggest that you can try debug and print the value. @Csysmile So where did you get the R-50 pkl?

Csysmile commented 5 years ago

@maomaochongchh I downloaded it at https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md

YubinXie commented 5 years ago

+1 Could you please provide the R-50/R-101 MS-RCNN model that is used to produce your results in the paper?

zjhuang22 commented 5 years ago

I've put the trained model in the corresponding link.

XiaoLaoDi commented 5 years ago

@fanyix @Csysmile https://github.com/XiaoLaoDi/maskscoring_rcnn here I provide the demo.py and the pretrained models on coco2017, but i am still not sure whether it matches the results of the paper, hoping can help you.