We use a modified version of keras which add a generic Maxout layer.

a config file must provide with two functions:

• get() which returns a dict of 'model', 'data_train' and 'data_val'
• get_data(list_path) The content of file located list_path is two image paths a line, name X and y respectively. This function should return a list of numpy array whose axes is (c, h, w), representing a multi-channel image. We run the model given by get() on (1, c, h, w) image and it should produce a image whose size is identical to y