HighEffiNNI

This repository is a high-efficiency training framework with NNI for PyTorch.

Requirements

To install requirements:

PyTorch >= 1.6
NNI >= 2.9
easydict >= 1.9

News

• In NNI 2.9, we can directly launch the visual website of NNI instead of configuring an ssh tunnel.
• Now, we can integrate the NNI's hyperparameter tuning into your project in a simple and quick way. You can compare nni_minist.py and minist.py to find these.
• A solution about the error (Failed to establish a new connection) in 2.8 or 2.9 of NNI can be found in config.yml.

config.yml

A example of config.yml can be found in here. More experiment config references can see here.

Quick start

To start up the NNI, run this command:

nnictl create --config config.yml -p 8140

To watch the running of NNI, run this command:

nnictl top 

To stop the NNI, run this command:

nnictl stop [--all] ([id])

TODO:

• To integrate the distributed training framework horovod into NNI.
• To develop more efficient training skills.