C++ vs Python feature extraction

[Brant Chee]

Hi Everyone,

    I'm new to Caffe and trying to determine if I can get the same output for python and c++ feature extraction.  I ran the example C++ feature extraction over the fish-bike.jpg

I get an array like:

[ 0.          0.          1.52356386 ...,  0.          0.          0.        ]

However, when I do feature extraction in python my feature array looks like:

net.predict([caffe.io.load_image('examples/images/fish-bike.jpg')], oversample=False)
features = copy.deepcopy(net.blobs['fc7'].data[0])
print (features)

[ 0.          0.          2.89886808 ...,  0.          0.          0.        ]

I think it has something to do with differences between the imagenet_val.prototext transformation params and the configuration I'm using for the python Classifier:

  transform_param {

    mirror: false

    crop_size: 227

    mean_file: "data/ilsvrc12/imagenet_mean.binaryproto"

  }

  image_data_param {

    source: "examples/_temp/file_list.txt"

    batch_size: 50

    new_height: 256

    new_width: 256

  }

vs 

net = caffe.Classifier(MODEL_FILE,
   PRETRAINED,
   mean=np.load(MEAN).mean(1).mean(1),
   input_scale=1.0, raw_scale=255,
   channel_swap=(2,1,0),
   image_dims=(227, 227))

Does anyone know the mapping between new_height, new_width and crop size and the corresponding Classifier parameters?

Thanks,
Brant

[jZero]

I also experiencing the same issue.

Have you figure it out?

[Gavin Hackeling]

I am also trying to extract the same features using the Python wrapper and the extract_features tool. Has anyone figured this out?