Open xingwangsfu opened 7 years ago
The flip layer and the alpha parameter are implemented in my caffe https://github.com/happynear/caffe-windows/tree/ms . Anyway, you may directly delete the alpha parameters and several layers such as flip_data
, concat_data
, slice_fc5
and eltmax_fc5
, then extract features of both front face and mirror face and do element-wise max operation on them. It should be same with my model definition. I do this just because I believe in the so-called "end-to-end" learning. In fact, there is no gain in accuracy after I integrate the flip and eltmax into cnn training.
The solver and train_val is actually the same with center-face, with only the step values changed. I don't know the best step values. I just observe and decrease the learning rate manually.
Hi, When I tried to use your centerloss model trained on MS-Celeb-1M, I met some problems: 1, clearly, "Flip" layer is not supported in official caffe version. Would you mind providing the source of your caffe? 2, In some convolutional layers, there is an 'alpha' parameter in weight_filler, which is also not supported in offiical caffe version.
BTW, would you mind sharing your solver.prototxt and train_val.prototxt, since it seems like yours is a little different from the original centerloss repo?
Thanks.