Caffe always gives the same prediction in Python, but training accuracy is good

[Alina Dima]

I have a trained model with caffe (through the command line) where I get an accuracy of 63%(according to the log files). However, when I try to run a script in Python to test the accuracy (acc2.py), I get all the predictions in the same class, with very similar prediction values, but not quite identical. My goal is to compute the accuracy per class, but in the example I just display the predictions for a given image.

Here are some examples of predictions:

[ 0.20748076  0.20283087  0.04773897  0.28503627  0.04591063  0.21100247] (label 0)

[ 0.21177764  0.20092578  0.04866471  0.28302929  0.04671735  0.20888527] (label 4)

[ 0.19711637  0.20476575  0.04688895  0.28988105  0.0465695   0.21477833] (label 3)

[ 0.21062914  0.20984225  0.04802448  0.26924771  0.05020727  0.21204917] (label 1)

It seems to be a problem with preprocessing, but I can't figure out what. The initial data is grayscale, but we convert it to rgb by duplicating the channels. What could be causing the problem?

[Alina Dima]

Edit: This is the actual prototxt used

[Alina Dima]

Update: Found the bug.

The finetuning was done on a net which takes 227x227 images, but I specified 256x256 in the prototxt file