mk-a / kerasextras

KerasExtras (Keras==2.2.4) with Tensorflow 2 support

Keras functions that you've always wanted!

• BatchNormalization that supports float16!
• Random cropping for augmentation
• Foreground sparse accuracy metric (if backgroud is set to '0')
• Background spare accuracy (if backgroud is set to '0')
• Sparse accuracy ignoring void (last) label
• Mean IOU (support for multi class labels)
• Segmentation generator: yields images, masks and sample weights
• Cyclic learning rate

Note: it works only when the output of the net is flattened. e.g., if the output is a mask of (M, N, #Classes) then you must add a reshape layer to make it (M*N, #Classes) and your labels are one hot encoded (and not sparse)