peijli/97-tensorflowjs-quick-start

Maybe this is NOT a good starting place for your next big TensorBoard project...

Forked from Angular Firebase for TFJS starter app (BlueML)

"TRUST THE PLAN"

Do I sound like I believe in QAnon bullshit? Hello, FBI. I never did.

Short-term Goals

• Go through the TensorFlow.JS training modules
• Go through the TensorBoard training modules
• Build a Fashion MNIST image classifying module OR
• Build off the MobileNet image classifying module OR
• Adopt some other model from tensorflow/tfjsmodels
• Construct a TensorBoard logging applet out of it
• Publish the applet through [TensorBoard.dev]

For the Next Few Months...

• Go through source code of tensorflow/tfjsmodels and Google Teachable Machine
• Find out ways to embed TensorBoard.dev visualizations into webpages
• Any chance of manually writing the TensorFlow log files from preexisting data?
• Any chance of presenting the graph only and NOT the potentially sensitive data?
• Build a neat little in-house Python package for generating a visualization module from an existing model and a JavaScript applet to produce a webpage.
• Try out more complex and realistic data generated from other SOCR projects

Learning Materials

Useful Links

• Google Machine Learning Crash Course
• TensorBoard Training Modules
• tensorflow/tensorboard repository

Tensorflow.js

• https://js.tensorflow.org/
• Fireship Tutorial
• freeCodeCamp.org video tutorial
• Converting an existing Keras model into a TensorFlow.js model

Convolutional Neural Networks

• Stanford CS 231N
• 3Blue1Brown video

Vegi's Proposed Approach

1. I believe Dimensionality reduction and Tensorboard might not be your first priority. Prof. Dinov's students have done a fair bit of work on Dimensionality Reduction and you have my initial TensorBoard work on SOCR.

2. The primary goal for your team in my opinion should be TensorFlow.js, i.e., you should leverage TensorFlow.js framework to do data processing, training, and prediction right in the browser without any standalone server.

3. This is how I would lay out the work.

   • Level 1: Learn Image Classification Algorithms. Implement a web app where a user uploads an image on your web app and your model predicts. For this, you can start your work on 2D ABIDE images.
   • Level 2: Now train the model on 3D images. Enable your web app to facilitate uploading and data processing of 3D images.
   • Level 3: Enrich each image with complimentary features such as Age, sex, demographics etc. Now the user can choose to get the predicts solely based on images or can get more accurate predictions with complimentary features he inputs on the web app.

Tensorflow.js MNIST Angular Demo

This demo imports an MNIST ConvNet trained in Keras Python, then makes predictions with TensorFlow.js

• clone it, cd into it, npm install && ng serve

Use a Different Keras Model

tensorflowjs_converter --input_format keras keras/yourWeights.h5 src/assets