Python wrapper: classifier.py does not initialize mean if no mean file is provided

[wendlerc]

Hello,

I just tried to switch off the mean normalization like it is proposed in http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/imagenet_classification.ipynb . When doing so I ran into an AttributeError since when not giving a meanfile to the classifier constructor, it won't initialize self.mean which is needed in the later _Net_preprocess function in pycaffe. For now I fixed it by initializing it in the Classifier constructor.

class Classifier(caffe.Net):
    """
    Classifier extends Net for image class prediction
    by scaling, center cropping, or oversampling.
    """
    def __init__(self, model_file, pretrained_file, image_dims=None,
                 gpu=False, mean_file=None, input_scale=None, channel_swap=None):

I skip some stuff there.

       if mean_file:
           self.set_mean(self.inputs[0], mean_file)
       #Chris Wendler
       if mean_file == None:
           self.mean = {}

Correct me if I understood something wrong.

Best regards,

Chris

[kubiakb]

Can anyone explain why mean file is necessary for classification?
Model was trained with images centered with means, but while classification process we are showing new image. In this case we need to calculated and use mean for that new image image.
What is the reason of reading training set means file?

Regards,
Blazej

[wendlerc]

@kubiakb http://metaoptimize.com/qa/questions/12584/preprocessing-images-centering-each-pixel
Have a look at this this maybe helps.

[shelhamer]

Fixed by #733.