fixpoint is a important function for the embedding device, such as arm.

it will make network inference much quicker in the mobile device.

nvidia's calibration method is one for this purpose.

but end2end fixpoint finetuning is the more suitable for speed up and accuracy. especially for the any bit wide optimization.

one example is ristretto. based on it , i have done something and trying to use it in the product.

but it is one caffe framework, i want to move it into the mxnet framework.

see my github: https://github.com/starimpact/ristretto_caffe

  • No labels

1 Comment

  1. Ming Zhang

    i have implemented good one in pytorch.

    i think mxnet should support hooks for each layer, both for forward and backward.