I expect mean subtraction to be same between .prototxt and python's transfomer.set_mean. but No.

[Peerajak Witoonchart]

Hi all

I am expecting the transform_param{} to calculate the same thing as transformer.set_mean. When these values do not agree over the very same image, I think something must be wrong.

Suppose I initailize a caffe solver with the following test section. I found that the solver.test_nets[0] will preprocess the image different from python caffe transformer.set_mean.
If I remove both the transform_param{} in prototxt and Transfomer.set_mean in python code out of the code, my net.data layer will be the same value. However, when there is transform_param{}, and Transformer, the net.data differs.

layer {
  name: "cifar"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    mean_file: "examples/Softmax/mean.binaryproto"
  }
  data_param {
    source: "examples/Softmax/cifar10_testdata_pycreated"
    batch_size: 100
    backend: LMDB
  }
}

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import Image
import scipy.io
# load input and configure preprocessing
test_batch_mat_path='examples/Softmax/cifar-10-batches-mat/test_batch_mat'
i_test_batch_Obj = scipy.io.loadmat(test_batch_mat_path + '/test_batch.mat' )
X = i_test_batch_Obj.get('data').reshape((10000,3,32,32)).astype('float')
im = X[0,:,:,:].squeeze().transpose((1,2,0))
print im.shape
im = im[:, :, (2, 1, 0)]


print np.load('examples/Softmax/mean.npy').mean(1).mean(1)

transformer = caffe.io.Transformer({'data': net_softmax.blobs['data'].data.shape})
transformer.set_mean('data', np.load('examples/Softmax/mean.npy').mean(1).mean(1))
#transformer.set_mean('data', np.load('python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
#transformer.set_mean('data', mean_arr[0]) # ImageNet mean
transformer.set_transpose('data', (2,0,1))
transformer.set_channel_swap('data',(2,1,0))#(2,1,0))
transformer.set_raw_scale('data',1.0)
# make classification map by forward and print prediction indices at each location
transformed_image = transformer.preprocess('data', im)
net_softmax.blobs['data'].data[...] = transformed_image
net_softmax.forward()