NNDy

INTRO

This is the user guide for NNDy The code is thoroughly commented so this guide will focus on the configuration, installation of packages and general usage of the optimization package as it was implemented by me

For some info about the algorithm and the architecture refer to the notebook 01/28 - NN-Dy explanatory intro of the lab's OneNote under People's NOtes/Fabio

The original and bare code can be found in this git Versions of the script already adapted and ready to use for different sequences of the experiment are instead to be found in the main lab git, in a dedicated folder within the labscript sequences folder

If you're using the package on the main control pc or in the main preplab pc, there's no need for setup

Skip to COST FUNCTION section

MLOOP installation

Install M-LOOP package from pip in the same python environment where you installed labscript

pip install M-LOOP

See here https://m-loop.readthedocs.io/en/stable/install.html#installing-with-pip

Check dependencies if needed

TENSORFLOW INSTALLATION FOR GPU (not for windows :( )

If you want the NN to be built and run with your GPU you need to make sure that the tensorflow version installed leverages cuda and your GPU (if you install the tensorflow package from pip or via M-LOOP dependencies this is not usually the case)

Follow here https://www.tensorflow.org/install/pip

DOWNLOAD CODE FROM THIS GIT

Into the sequences folder of your local labscript installation create a dedicated folder for NNDy where you copy:

• NNDy.py
• NNDy_Interface.py
• runmanager_remote.py
• experiment_remote.py - not necessary, used only for testing and debugging

Copy the script:

• DA/newRun_NNDy.py

into the lyse folder of your local labscript installation, here: ${labscrit_suite_folder}$\userlib\analysislib\{experiment_name}

RUNMANAGER_REMOTE.PY

Modify the path where it will look for the sequences, this will be the shared_drive path of the specified in the during configuration of the labscript software (see https://labscriptsuite.org/en/latest/setup/configuration/ for more details)

COST FUNCTION

In the first lines of NNDy.py you can indicate the file that will perform the analysis of the shot(s) at each iteration

How should this file be written? - see /DA/TestDA.py for more instructions The essential part of the file is the function

def cost(hdf_file):

This is the function that will be transferred to the internal structure of the interface (see NNDy_Interface.py) and be called after the experiment sequence is finished

The function reads the hdf_file of the shot and returns a dictionary with

• cost
• bad flag
• uncertainity

Within this cost function the analysis routine is initiated and handled to return the desired observable from which you want to estimate the cost The analysis can be customized in any way you desire, as long as it relies of the hdf file to read the results of the shot and the cost function returns the dictionary as explained above

NNDy SCRIPT - here comes the fun part

Read carefully through the NNDy.py before running the optimizer as all the degrees of freedom offered and the details that you need to take care of are explained in the comments

As a recap:

• Indicate the name of the routine that you want to run - aka the experimet
• Indicate the name of the cost/DA script that - aka the cost function
• Set at least one halting condition
• Set the names, the initial values and the boundaries of the input parameters

Optional functionalities

• Set values of fixed, global parameters
• Recall as training dataset the data of past runs
• Handle trust region
• Activate update of hyperparameters

BEFORE HITTING RUN

Open runmager, blacs and lyse Make sure that they're correctly configured and that your sequence works fine - try hitting engage from runmanager

Then in lyse check the box for newRun_NNDy.py

Now you can run NNDy.py