nni is an automatic machine learning framework open sourced by Microsoft
Can make machine learning tuning more convenient
Installation is very simple, just use pip
pip install nni
This time I am using the latest version, 2.0
But it cannot be used at this time, because it will appear:
https://github.com/microsoft/nni/issues/3276
The solution is also very simple, directly change ruamel.yaml to ruamel_yaml
Edit and modify the content under the path below
vim /opt/AN/lib/python3.7/site-packages/nni/tools/nnictl/common_utils.py
vim /opt/AN/lib/python3.7/site-packages/nni/experiment/config/base.py
vim /opt/AN/lib/python3.7/site-packages/nni/tools/package_utils/__init__.py
Then it can be used normally
root@lizong:/opt# nnictl --help
usage: nnictl [-h] [--version]
{ss_gen,create,resume,view,update,stop,trial,experiment,platform,webui,config,log,algo,package,tensorboard,top}
...
use nnictl command to control nni experiments
positional arguments:
{ss_gen,create,resume,view,update,stop,trial,experiment,platform,webui,config,log,algo,package,tensorboard,top}
ss_gen automatically generate search space file from trial
code
create create a new experiment
resume resume a new experiment
view view a stopped experiment
update update the experiment
stop stop the experiment
trial get trial information
experiment get experiment information
platform get platform information
webui get web ui information
config get config information
log get log information
algo control nni builtin tuner, assessor and advisor
algorithms
package this argument is replaced by algo
tensorboard manage tensorboard
top monitor the experiment
optional arguments:
-h, --help show this help message and exit
--version, -v
Next, run the official example:
nnictl create --config nni/examples/trials/mnist-pytorch/config.yml
Note: git clone is required before running
git clone -b v2.0 https://github.com.cnpmjs.org/Microsoft/nni.git
After running successfully, you can observe the running result on the webpage:
Specific access address ------ running machine ip: 8080
We can see that the above picture has been running for 48m54s, and the best score is 99 points
Look at the hyperparameter curve:
The operation introduction will appear: