This is a Python reimplementation of this repo. It includes:
-
Replacing MATLAB code with PyCaffe during evaluation phase
-
Gives exactly the same mAP value for the default caffemodel (
KevinNet_CIFAR10_48.caffemodel
) on default dataset (cifar-10), i.e. 0.897373
- Caffe with its Python wrapper
- OpenCV's Python wrapper
- Common Python packages(numpy, matplotlib, etc)
-
Prepare dataset and pretrained model as claimed in caffe-cvprw15. You can also download the caffemodel file via BaiduYun with password(3doq)
-
run this bash command:
python py/run_cifar10.py
which will compute mAP and precision at different level, also gives the following output:
-
py/ilsvrc2012_mat_mean.pkl
is converted frommatlab/caffe/ilsvrc_2012_mean.mat
of caffe-cvprw15. -
In MatCaffe, blobs are with order (W,H,C,N), and in PyCaffe, it use (N,C,H,W)