Designing an end-to-end Pipeline for Developing and Deploying IoT Solutions on Embedded Neuromorphic Platforms

M.Sc. Thesis Project
Author: Marco Bramini
Thesis Link: https://webthesis.biblio.polito.it/29384/

How To Use

Initialize Environment

Create Conda Environment

This project was specifically thought to work on Linux OS, having Conda installed. A fully comprehensive virtual environment can be installed on such OS using the following command:

conda env create -f=requirements.txt -n myenv

In case of different operating system, the user must manually build the virtual environment and install packages one by one:

conda create env -n myenv
conda activate myenv
conda install pip
pip install “rockpool[all]”
pip install nni
pip install tonic
pip install -U "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install ipykernel

Unpack dataset

Use 7-Zip to unpack the multipart zip file. The extraction must produce the following files:

wisdm_watch_full_40.npz
wisdm_watch_full_40_encoded.npy
wisdm_watch_full_40_classes.json

The provided scripts assume these files are contained in a root folder named data/.

Task Definition

The task-definition folder contains all the scripts needed for the Task Definition step.

Generate tasks employing the DTW-based distance metric

1. Configure the global settings in the head of the script task-definition/generate_tasks_dtw.py.
2. Run the script with the command:

   cd task-definition
   python generate_tasks_dtw.py
   

Generate tasks employing the KLD-based distance metric

1. Configure the global settings in the head of the script task-definition/generate_tasks_kld.py.
2. Run the script with the command:

   cd task-definition
   python generate_tasks_kld.py
   

Plot the KDE for a specific combination of classes

1. Configure the global settings in the head of the script task-definition/plot_kde.py.
2. Run the script with the command:

   cd task-definition
   python plot_kde.py
   

Model Generation

Run Architecture Search

1. Configure the training pipeline editing the file training-pipeline/run_xylo_rockpool_train.py

   In particular:

   • Select the model architecture editing the line 37, selecting one of the predefined architectures in the library:

     xylo_networks.get_ff_simple(...)
     xylo_networks.get_ff_deep(...)
     xylo_networks.get_ff_deep_res(...)
     xylo_networks.get_ff_deep_deep_res(...)
     xylo_networks.get_rec_simple(...)
     xylo_networks.get_rec_deep(...)
     

   • Edit the parameter input_params["enabled_classes"] at line 177, with the class labels associated to the task 7CB

2. Run NNI with the following commands:

   cd training-pipeline
   python run_nni_hpo.py --config_path nni_experiment_configs/xylo_as.json --port 32000
   

   It is possible to follow the process on the NNI GUI at localhost:32000

3. At the end of the process, the full list of experiments and model checkpoints will be stored in the folders training-pipeline/experiments/{EXPERIMENT_ID}/{TRIAL_ID}

   The best performing model can be also retrieved with the script training-pipeline/get_best_trial_id.py, after setting the experiment id at line 4

Run HPO

1. Configure the training pipeline editing the file training-pipeline/run_(dynapse2|xylo)_rockpool_train.py

   In particular:

   • Edit the parameter input_params["enabled_classes"] at line 177, with the class labels associated to the selected task

2. Run NNI with the following commands:

   cd training-pipeline
   python run_nni_hpo.py --config_path nni_experiment_configs/(dynapse2|xylo)_hpo.json --port 32000
   

3. At the end of the process, the full list of experiments and model checkpoints will be stored in the folders training-pipeline/experiments/{EXPERIMENT_ID}/{TRIAL_ID}

   The best performing model can be also retrieved with the script training-pipeline/get_best_trial_id.py, after setting the experiment id at line 4

Run Extended Training

1. Configure the training pipeline editing the file training-pipeline/run_(dynapse2|xylo)_rockpool_train.py

   In particular:

   • Edit the parameter input_params["enabled_classes"] at line 177, with the class labels associated to the selected task

2. Directly run the training pipeline with the following command:

   cd training-pipeline
   PYTHONPATH="lib" python run_(dynapse2|xylo)_rockpool_train.py
   

3. At the end of the process, the folder best_model/ will contain the checkpoint of the trained model

Hardware Deployment

Run Quantization Tuning

This process needs for a trained model. The tuning parameters will be automatically applied in the model checkpoint.

1. Configure the NNI Quantization Tuning experiment configuration by editing the file training-pipeline/nni_experiment_configs/(dynapse2|xylo)_tuning_(task).json

   In particular:

   • Change the model path at line 2.

2. Run NNI with the following commands:

   cd training-pipeline
   python run_nni_hpo.py --config_path nni_experiment_configs/(dynapse2|xylo)_tuning_(task).json --port 32000
   

Hardware Configuration Generation

Refer to the Python Notebooks in dynapse2-deploy/and xylo-deploy/ for the generation of the hardware configuration.