bvlc_googlenet preprocessing parameters

[Yoann]

Hi all,

The parameters of the images used as input of the bvlc GoogleNet are not written explicitly.

From the train_val.prototxt, since it is using the same dataset as the bvlc reference model, I assume that the preprocessing for the input pictures for the training phase is the same. Am I right?

Properties:

channel_swap 2,1,0

image_scale 0-255

transpose 2,0,1

I'm trying to fine-tune googlenet and my pictures need to be processed in the same way.

Thanks,

Yoann

[npit]

Where are you getting these properties from?

[Yoann]

From the defaults parameters of the python script "classify.py" for the bvlc_reference_caffenet.

[echeng]

I too am trying to use a fine-tuned GoogleNet model, but when using the default preprocessing parameters you mention above, the results do not match what the test accuracy reports during training.  

I've posted a similar topic here: https://groups.google.com/forum/#!topic/caffe-users/3Xgr6mF34Wo

[echeng]

Turns out I had a bug -- modified the name of the classifier layers in train_val.prototxt, but did not modify deploy.prototxt to match.  Once I did, everything worked as expected using the parameters specified in the ImageNet Classification ipython notebook example (http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/classification.ipynb) which match the preprocessing parameters you cite above.