MXNet is 8 times slower running a simple example than my code with numpy.
I am providing a zip file with the necessary files to reproduce the problem.
Steps to reproduce
- Uncompress the zip file.
- Use Jupyter with python3 to open both examples files
- Compare the processing time between examples
What I have tried to solve it
- Running in my GPU
- Running in my CPU
- Running hybrid and non hybrid NN, both net.hybridize and loss_function.hybridize
System: Ubuntu 18.04.4 LTS 64-bit
MXNet version: mxnet-cu102mkl 1.6.0
Python version: Python 3.6.9
CPU: AMD Ryzen 9 3950x
GPU: GeForce RTX 2080 SUPER
RAM: 64 GiB
With Numpy: 1.9 seconds (obviously in cpu)
With MXNet: 15.7 seconds (in cpu without hybridize)
With MXNet: 14.3 seconds (in gpu without hybridize)
With MXNet: 6.6 seconds (in cpu fully hybridized)
With MXNet: 6.1 seconds (in gpu fully hybridized)
I hope this is the place to report this type of problems and get some help. Looking forward for your responses.
I have uploaded the necessary files into this github post: