Browse the notes index
Every note on this site is filed under one of four subject areas: TensorFlow and neural networks, the machine learning workflow, D3 and data visualisation, and the JavaScript data stack. Pick a subject to see everything filed under it.
Topic categories
TensorFlow and neural networks
How a TensorFlow program is structured, what the training loop actually repeats, and how to tell a model that is learning from one that is memorising.
Machine learning workflow
The plain mechanics either side of the model: preparing columns so they compare fairly, reading the result honestly, and picking the lighter tool when it fits.
D3 and data visualisation
Scales, axes, force simulations and text layouts, described as the small functions they are rather than as chart recipes.
JavaScript data stack
What the Node and MongoDB layer around a model is responsible for, and what it should refuse to take on.
Current reading
Questions these notes answer
- How do a graph and a session relate in TensorFlow?
- What does a training loop actually repeat?
- Why are features scaled before fitting?
- How is a confusion matrix read?
- What does overfitting look like in a learning curve?
- How does a force layout decide where nodes go?
- How does a D3 scale map data to pixels?
- How is a word cloud built from raw text?
- Where does a Node and MongoDB layer fit under a model?
- Which Python workflow should be reached for first?
When something is not working
- The loss will not go down at all
- What a training loop repeats
- Accuracy is high on training data only
- Overfitting and regularisation
- The chart renders but every point sits at the same coordinate
- D3 scales and axes
- The force graph never settles
- Force directed graphs in D3
- The model fits in a notebook but not in the service
- Feature scaling before fitting
Stack and work archive
Working stack
I am a Full Stack Web, Deep Learning & D3 Visualizations Engineer working for Adappt ltd.(UK), building amazing products and solutions for enterprises around the world. I also write articles on programming, do talks in local meetups and host webinars for positive karma.
Realtime analytics with Python & D3 - ChennaiPy (May'16)
Projects, talks and video
- Ranked Keywords Word Cloud from Hillary Clinton's Emails
- Papyrrus - Minimalistic Medium Style Web-based Text Editor
- Neural Network module for Pattern Recognition and Function Approximation
- Map Visualisation - Who is Smoking Weed in the World?
- Labelled & Curved Force Directed Graph with Variable Image Radius
- Tensorflow Mechanics - ChennaiPy (May'16)
- Introduction to Deep Learning - Geek Night (May'16)
- Introduction to Machine Learning - Chennaipy (April'16)
- Tensorflow Mechanics - ChennaiPy (May'16)