Transfer Learning in TensorFlow.js

Examples of transfer learning using TensorFlow.js. Here, tfjs-node will be used.

Introduction

Modern, state-of-the-art models typically have millions of parameters and can take inordinate amounts of time to fully train. Transfer learning shortcuts a lot of this training work by taking a model trained on one task and repurposing it for a second related task. We do this by removing the final layer(s) of the pre-trained model and then train a new, much smaller model on top of the output of the truncated model. A major advantage of this technique is that much less training data is needed to train an effective model for new classes.

However, take note that in order for this to be effective, model features learned from the first task should be general. That is, the features should be suitable for both the first and second tasks.

To illustrate this in TensorFlow.js, we will use the Fashion-MNIST dataset.

Instructions

Download and extract the Fashion-MNIST dataset into the root of the project. Your structure should look like this:

tfjs-transfer-learning
│
└───fashion-mnist
    │   fashion-mnist_train.csv
    │   fashion-mnist_test.csv

Train a model from scratch using the first five classes of the dataset. After training, the model will be saved in ./fashion-mnist-tfjs.
```
node train-model.js
```
Train a classifier for the other five classes using transfer learning. However, this time, using only a fraction of the training dataset for each class. Training should be much faster now, and test accuracy should be similar to the first model.
```
node transfer-learn.js
```

pvaneck/tfjs-transfer-learning

Transfer Learning in TensorFlow.js

Introduction

Instructions