This repository aims at prototyping a pure numpy
interface above mxnet backend. The key features include:
- Autograd support.
- Nature MXNet symbol integration.
- Graceful fallback for missing operations.
- Transparent device and partition specification.
The project is still a work-in-progress. You could look at this tutorial to understand its concept. Documents and installation instructions are coming soon!
In one word, if you have a numpy
code, you could replace the import
by:
import minpy.numpy as np
# other numpy codes remain the same
and you could have:
- Auto differentiation support.
- Speed up with some operations executed on GPUs.
- Missing operations will not cause "NO IMPLEMENTATION" exception.
- Directly call Caffe's Layer abstraction without any code change.
- Switch between
numpy
's operators and Caffe's operator as you wish.