Overview

This tutorial introduces how to make your data exploration and model building process more interactive and exploratory by using the combination of JupyterLab, HoloViews, and PyTorch. HoloViews is a set of Python libraries that offers simple yet powerful visualization and GUI building tools which, together with other data analysis libraries (eg. pandas, geopandas, numpy) and machine learning framework (eg. PyTorch, Tensorflow) can make your modeling procedure more interactive and exploratory. I will start by introducing four core HoloViews libraries (Holoviews, GeoViews, Panel and Param) and demonstrate basic examples on how we can essentially replace any "Matplotlib.pyplot" calls with equivalents in HoloViews. You will see how this opens up the possibilities to directly interact with your visualization by eg. hovering over the graph to inspect values, querying RGB values of an image, or Lat/Lon values on your map.

Following the introduction of the HoloViews libraries, I will demonstrate how to embellish each step of your machine learning workflow with HoloViews. First, you will learn to easily turn your PyTorch codes into a simple GUI that encaptulates the state of your model (or alternatively, the state of your training session). This GUI explicitly exposes your model parameters and training hyperparameters (eg. learning rate, optimizer settings, batch size) as directly tunable parameters. Compared to conventional ways of specifying the hyperparameter settings with the help of 'argparse' library or config files, this GUI approach focuses on the experimental nature of modeling and integrates seamlessly with Jupyter notebooks. After training a neural network model using our own GUI in the notebook, I will demonstrate how to understand the model by visualizing the intermediate layers with HoloViews and test the model with test images directly sampled from HoloViews visualization.

To illustrate these steps, I will focus on the problem of classfying different types of roads on satellite images, defined as a multi-class semantic segmentation problem. Starting from the data exploration to the trained model understanding, you will learn different ways to explore the data and models by easily building simple GUIs in a Jupyter notebook.

In summary, by the end of the talk you will have learned: - how to make your data exploration more intuitive and experimental using HoloViews libraries - how to turn your model script into a simple GUI that allows interactive hyperparameter tuning and model exploration - how to monitor the training process in realtime - how to quickly build a GUI tool to inspect the trained models in the same Jupyter notebook

The provided example codes will be a great starting point to experiment these tools on your own datasets and tasks.

Outline

This tutorial will consists of five main sections. I will first introduce the basic concepts behind Holoviews/Geoviews and Panel which are the main libraries we are going to use to add interactive exploration tools for data exploration and model training/evaluation, all in a single Jupyter notebook. This will take ~15 minutes. The rest of the tutorial will flow in the order of the general neural network training workflow, while integrating these libraries at each step. I will leave the last <10 minutes for questions.

• Step 0: Introduction to Holoviews/Geoviews and Panel [15mins]
• Step 1: Explore your dataset with Holoviews/Geoviews [15mins]
• Step 2: Build an easily-configurable neural network model with param [15mins]
• Step 3: Monitor your training process through an interactive GUI [15mins]
• Step 4: Analyze your learned model on new images + Understand what your model has learned by looking at intermediate feature maps with Holoviews and Panel [15mins]
• Q/A [5~10 mins]