Accelerate large-scale machine learning with NP on MXNet
summary
NP on MXNet claims a drop-in replacement for numpy that keeps numpy APIs while dispatching ops to CPU and GPU; microbenchmarks show about 1.2× on CPU and 30–40× on GPU for matrix dot, and it adds automatic differentiation and extra deep‑learning ops. The demo builds an MLP on Fashion‑MNIST (batch size 256), trains with SGD and weight decay, verifies gradients against numpy, and notes language bindings plus plans to use Apache TVM for wider hardware deployment.
AI-written summary · about
chapters (9)
- 0:00MXNet and NP package
- 1:29NumPy interoperability
- 3:17Parallel acceleration
- 5:48NPX deep learning operators
- 9:01Automatic differentiation
- 16:38Building model with NPX
- 26:57Deployment overview
- 29:25Future plans
- 30:12Q&A
related talks
29mAccelerate NumPy Data Science Workloads and Deep Learning Applications
30mApache MXNet 2.0: Towards Synergistic ML and DL with Standardization
30mAnyone can Build Great Deep Learning Applications - Deep Numpy
42mAnima Anadkumar, Principal Scientist, Amazon Web Services, Endowed Professor, CalTech
38mThe new kid on the block: deep learning with GluonNLP
01h 16mDeep Dive into Deep Learning with MXNet
33mHigh Performance Deep Learning on Edge Devices With
20mApache MXNet, a highly memory efficient deep learning framework