Proper Classifer setup for GoogleNet - Cannot reproduce test accuracy

[echeng]

What is the proper way to modify the Classifying Imagenet (http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/classification.ipynb) example to use the BVLC GoogleNet model?

Following the Flickr style example, I fine-tuned the BVLC GoogleNet model for my application and am seeing test accuracy of about 95%.  However, when I modify the Imagenet example above to use my fine-tuned GoogleNet model and actually classify images, it seems to be performing much worse.

Per this topic: https://groups.google.com/forum/#!topic/caffe-users/Xze-SMvYCvs it looks like it might be due to a mismatch in preprocessing, but as far as I can tell, the means, channel swapping, and image sizes in the example are correct for GoogleNet.

Any suggestions?  What are the common ways I might be misusing the model?

Thanks.

[echeng]

Turns out I had a bug.  To fine-tune, I renamed the classifier layers in train_val.prototxt, but forgot to modify the name of the corresponding deploy.prototxt layer.  Once I fixed this, I was able to reproduce the test accuracy reported by Caffe during training using the default preprocessing parameters specified in the Classifying Imagenet ipython notebook example. 

[S.Majid Azimi]

how to did you do that?

I get the error that i can't be broadcast from 3,224,224 to 3,227,227 which I can't understand because the size of cat image is 360 480 3 which should be converted by transformer.preprocess to 1,3,227,227. I don't know from 224 is coming from?

[auro tripathy]

I'm getting the same error. I switched the models and the prototxt file for GoogLeNet (from the reference model). 

net = caffe.Net(caffe_root + 'models/bvlc_googlenet/deploy.prototxt',

                caffe_root + 'models/bvlc_googlenet/bvlc_googlenet.caffemodel',

                caffe.TEST)

Error is the same as yours.

Help in understanding this is deeply appreciated. 

Thank you.

--------------------------------------------------------------------------

ValueError                                Traceback (most recent call last)

<ipython-input-4-0b9723a6825b> in <module>()

      1 print(caffe_root)

----> 2 net.blobs['data'].data[...] = transformer.preprocess('data', caffe.io.load_image(caffe_root + 'examples/images/cat.jpg'))

      3 #net.blobs['data'].data[...] = caffe.io.load_image(caffe_root + 'examples/images/taj-mahal.jpg')

      4 out = net.forward()

      5 print("Predicted class is #{}.".format(out['prob'].argmax()))

ValueError: could not broadcast input array from shape (3,224,224) into shape (50,3,227,227)

----------------------------------------------------------------------------

[auro tripathy]

matching the dimensions seems to fix it. 

old

# set net to batch size of 50

net.blobs['data'].reshape(50,3,227,227)

new

# set net to batch size of 50

net.blobs['data'].reshape(50,3,224,224)

[johnn...@gmail.com]

I get the same error that I can't classify that cat.jpg, because the size is not right. The cat.jpg is 480*360*3, but in bvlc_googlenet/deploy.prototxt, the input_shape is 10*3*224*224. I do not know the details. Do you find any solutions? 

Thanks!

[Giulia]

Thanks ekc, I had EXACTLY the same problem, and browsing I found your post which saved me from days of debugging!!!!

Giulia