Test accuracy is low on cub

[MyChocer]

I download the provided trained model from the website
(https://drive.google.com/open?id=0B0ldy05JZFSaS1pIUjBaSUxMTUU)
and test the cub dataset. But I get only 0.003 test accuracy and high loss up to 5.29.
And, I used the script in the folder(examples/compact_bilinear) and I didn't modify anything.
Could you help me to solve this problem?

[gy20073]

If you could point me to a repo, I could probably help to figure out the issue. Otherwise, I don't have any clues....

[sandareka]

@MyChocer Were you able to resolve this? I'm also having the same issue? Accuracy with pretrained models is really low.
@gy20073 I would be really grateful if you could help us to resolve this.

Thank you in advance!

[gy20073]

@sandareka I'm not really able to help if you don't provide exact steps to reproduce your problem.

[sandareka]

Hi @gy20073 ,

Thank you very much for your prompt reply.

This is my approach.

I'm trying to classify CUB bird dataset. For that I integrated your new caffe layers to my version of caffe as you have explained in the readme. I added all the files and rebuilt. Build was sucessful. Then I tried to use your model(provided prototxt file and .caffemodel from google drive) to classify images. However, classification is almost always incorrect. What I want to confirm is if the .caffemodel files given in the google drive are weights of the compact bilear model trained for classification of CUB birds dataset?

I hope I'm clear now. Your help with this regard is highly appreciated.

Thank you!

[gy20073]

ok, I understand your steps now. Have you tried using the old version of Caffe that comes with my repo? If you could try that and get normal accuracy, then it indicates that the weights are not compatible with newer version of Caffe. Otherwise there might be some details we are missing.

…

On Oct 9, 2017, at 7:55 PM, Sandareka Wickramanayake ***@***.***> wrote: fy CUB bird dataset. For that I integrated your new caffe layers to my version of caffe as you have explained in the readme. I added all the files a

[MyChocer]

@gy20073 @sandareka .I found that it worked well in cuda7-5 ,cudnnv4,but failed in cuda 8 with cudnn v4 or v5.Beside, I try to upgrade the caffe version and it works in cuda8 ,cudnnv5.

[sandareka]

Hi @gy20073 ,

No I didn't try with your version of caffe. I will give it a try. But as per my understanding .caffemodel contains weights (numbers) and they shouldn't depend on caffe version. Please correct me if I'm wrong. Further, I have used pretrained models trained with earlier caffe versions with my current version of caffe and they worked perfectly. Anyway I will try with your older version.

Thank you very much for your cooperation.

[sandareka]

@MyChocer Did you mean when you upgrade caffe version it worked well? May I please know your caffe version? I have cuda 8 and cuddn v5.