Python wrapper and extract_feature.cpp gives different fc7 layer values for the same image ?

[erogol]

I created a single image path file_list2.txt and I give it to caffe_reference model as

layers {
name: "data"
type: IMAGE_DATA
top: "data"
top: "label"
image_data_param {
source: "file_list2.txt"
batch_size: 1
new_height: 256
new_width: 256
}
transform_param {
crop_size: 224
mean_file: "mean_file.binaryproto"
mirror: false
}
}

layers {
name: "conv1"
type: CONVOLUTION
bottom: "data"
top: "conv1"
blobs_lr: 1
....

the I call extract_features.cpp as

./build/tools/extract_features.bin models/CaffeRef_iter_100000.caffemodel models/caffe_reference/train_val_feat_ext.prototxt fc7 /path/to/features 1

so it takes 1 image for 1 instance batch

Then for the same image I run following python code

net = caffe.Classifier(MODEL_FILE, "models/CaffeRef_iter_100000.caffemodel ",
mean=np.load(mean_file.npy'),
channel_swap=(2,1,0),
image_dims = [256,256],
raw_scale=255)

img_path = 'same/file/path'
img = caffe.io.load_image(img_path)
plt.imshow(img)
x = net.predict([img], oversample=False)

MODEL_FILE is same with last one except it has different data input;

input: "data"
input_dim: 10
input_dim: 3
input_dim: 224
input_dim: 224
layers {
name: "conv1"
type: CONVOLUTION
bottom: "data"
top: "conv1"
convolution_param {
num_output: 96
.....

So normally I ecpect to have the same feature vector for this image from these two different approach. However they are different. I compare both feature vector on Python like below to see that whether only the non-zero values are different or activated units are different too.

X[X>0] = 1
x[x>0] = 1
print len(np.nonzero(X-x)[0])
output: 845

I guess this needs a bit of attention from the community. Please direct me to enhance the observation.

[shelhamer]

Keep to the thread on caffe-users: https://groups.google.com/forum/?utm_medium=email&utm_source=footer#!msg/caffe-users/r2w27xzSLu8/MUff_buD6yEJ

[erogol]

I believe this is not users problem since Python wrapper gives different output compared to caffe backhand. I guess something in the pycaffe preprocessing breaks the given image.