Cifar-10 example gets no prediction right.

[Robmosh]

Hi there,

I am relatively new to caffe and thus am trying to understand the examples.

Yesterday I have trained the cifar 10 example with the full_solver and no changes in the .prototxt.

After training I now want to make predictions in Python. Therefor I use

MODEL_FILE_FULL= '/usr/local/lib/python2.7/dist-packages/caffe/examples/caffe/cifar-10/cifar10_full.prototxt'

PRETRAINING_FILE_FULL= '/usr/local/lib/python2.7/dist-packages/caffe/examples/caffe/cifar-10/cifar10_full_iter_70000.caffemodel.h5'

IMAGE_FILE_LIST = list of images

mean = np.load('/usr/local/lib/python2.7/dist-packages/caffe/examples/caffe/cifar-10/mean.npy).reshape((3,32,32))

net = caffe.Classifier(MODEL_FILE_FULL, PRETRAINING_FILE_FULL, mean)

input_img = caffe.io.load_image(IMAGE_FILE_LIST.pop(0))

prediction = net.predict([input_img])

print 'prediction class:', prediction.argmax()

My problem now is, that i get class 3 as prediction for every image. I tested images from every class and always get the same result, where the probability of class the is around 8.8-01, thus very high for all images.

Has anybody had similar problems and a solution for this?

I already tested it with the short training and downloaded the cifar-10 data new.

Thanks!!!

[Robmosh]

Can'take anybody help me with this problem?

[Hossein Hasanpour]

Why dont you use the ipython notebook sample(located in examples directory) for this? its way easier and straight forward. use that and if the problem still persists post your question

[Robmosh]

Which example do you mean? There is one for classification, in which a model is loaded via 

net = caffe.Net(model,_def, model_weights, caffe.TEST)

and the an image will be predicted with this net via

net.forward()

net.blo 

Then the probobilitys get printed.

I think that this is exactly the same like i do, with the different that i use the caffe.Classifier instead of caffe.Net.

[zxdz...@gmail.com]

hi, I had got the same problem as yours. I use caffe.Net to predict, and always get the class 4. Have you solve this problem? I will be pgrateful if you or someone can help me. Thanks a lot!

在 2016年4月29日星期五 UTC+8下午8:07:21，Robmosh写道：