Ilya Nelkenbaum

@cxy200927099 exactly. Original code had ReLU layers after fully connected instead of PReLU (as in original paper) and that is the reason for missing alpha in current model.

@cxy200927099 I am still working on improving landmark detection, and it will take some time. In order to solve it you have to train model from scratch before you are...

@ddxu it should be solved by: https://github.com/AITTSMD/MTCNN-Tensorflow/commit/519b453975a8dbdc7b0eecdb4f8735b7b6c80d46

@flankechen Hi, you can see the original code that was written by the author's: https://github.com/kpzhang93/MTCNN_face_detection_alignment/blob/master/code/codes/MTCNNv1/model/det2.prototxt There is PReLU. The arg scope is valid for conv2d ops only and not fully_connected

The affine is: 

I don't have access to the PC where the data is stored, so I just manually retyped (hopefully wo typos) from JPEG above. Just remove .zip extension after downloading: [aff.npz.zip](https://github.com/nipy/nibabel/files/7001460/aff.npz.zip)

I will try to acquire the original data and upload here when I will get it.

Affine taken from NIFTI: [aff.npz.zip](https://github.com/nipy/nibabel/files/7002190/aff.npz.zip)

Yep, looks like ordinary axes. It will take me a while since I don't have direct access to this PC, I will update soon.

@peteryuX ?