search

facebookarchive / C3D

C3D is a modified version of BVLC caffe to support 3D ConvNets.
Other
1.16k stars 507 forks source link

testing_ucf101.sh accuracy problem #376

Closed ofeyliu closed 6 years ago

ofeyliu commented 6 years ago

Hello! Thanks for your C3D work!

I met a problem when I tried to run C3D-v1.1/examples/c3d_ucf101_finetuning/testing_ucf101.sh. I had downloaded c3d_resnet18_ucf101_r2_iter_90000.caffemodel weights but my accuracy was low as belows:

Could you tell me what reason may cause this problem? Thank you!

xianyuanliu commented 6 years ago

I met the same problems. The loss is between 3~4, and the top-1 is around 0.15, while top-5 is 0.35. Do you solve it?

ofeyliu commented 6 years ago

@liulinyi007 I might figure out this problem. I found issue 354 in the closed issue. But the prototxt was almost same in c3d_ucf101_finetuning folder except the last fully connected layer which name was "fc8" where .prototxt file in folder was "fc8-new". So I wonder maybe I used prototxt file with wrong caffemodel weights which had wrong name. My solution is firstly finetuned train_resnet18_r2.prototxt in UCF101 dataset. When I have trained and got c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel, I tested it and I got 80.71% top-1 accuracy.

xianyuanliu commented 6 years ago

Thanks very much. I'll try this immediately Another question is that I finetuned train_resnet18_r2.prototxt on c3d_resnet18_sports1m_r2_iter_2800000.caffemodel and met the same problem. Did you do it before?

ofeyliu commented 6 years ago

Sorry late to reply! I haven't met this problem. But I wonder there may exist some configure problem if you don't revise the code. The file I downloaded from dropbox, I think, is different with prototxt file in the repository folder. By the way, recently I choose Pytorch version Res3D, may this help you.

xianyuanliu commented 6 years ago

Thxxx! I found the solution. I compiled the C3D without CUDNN, and then the question disappeared. However, I do not know why yet.