date-conversion-attention

TensorFlow.js Example: Date Conversion Through an LSTM-Attention Model

See this example live!

Overview

This example shows how to use TensorFlow.js to train a model based on long short-term memory (LSTM) and the attention mechanism to achieve a task of converting various commonly seen date formats (e.g., 01/18/2019, 18JAN2019, 18-01-2019) to the ISO date format (i.e., 2019-01-18).

We demonstrate the full machine-learning workflow, consisting of data engineering, server-side model training, client-side inference, model visualization, and unit testing in this example.

The training data is synthesized programmatically.

Model training in Node.js

For efficiency, the training of the model happens outside the browser in Node.js, using tfjs-node or tfjs-node-gpu.

To run the training job, do

yarn
yarn train

By default, the training uses tfjs-node, which runs on the CPU. If you have a CUDA-enabled GPU and have the CUDA and CuDNN libraries set up properly on your system, you can run the training on the GPU by:

yarn
yarn train --gpu

Monitoring model training with TensorBoard

The Node.js-based training script allows you to log the loss values to TensorBoard. Compared to printing loss values to the console, which the training script performs by default, logging to tensorboard has the following advantanges:

1. Persistence of the loss values, so you can have a copy of the training history available even if the system crashes in the middle of the training for some reason, while logs in consoles are more ephemeral.
2. Visualizing the loss values as curves makes the trends easier to see (e.g., see the screenshot below).

To do this in this example, add the flag --logDir to the yarn train command, followed by the directory to which you want the logs to be written, e.g.,

yarn train --logDir /tmp/date-conversion-attention-logs

Then install tensorboard and start it by pointing it to the log directory:

# Skip this step if you have already installed tensorboard.
pip install tensorboard

tensorboard --logdir /tmp/date-conversion-attention-logs

tensorboard will print an HTTP URL in the terminal. Open your browser and navigate to the URL to view the loss curves in the Scalar dashboard of TensorBoard.

Using the model in the browser

To see the trained model in action in the browser, do:

yarn
yarn watch

Visualization of the attention mechanism

In the page opened by the yarn watch command, you can generate random input date strings by clicking the "Random" button. A converted date string will appear in the output text box each time a new input date string is entered. You may also manually enter a date in the input-date box. But make sure that the date falls into a range between the years 1950 and 2050, as this is the range of dates that the model is trained on. See date_format.js for more details.

In addition to converting the date and showing the output, the page visualizes the attention matrix used by the trained model to convert the input date string to the output one in (e.g., see the image below).

Each column of the attention matrix corresponds to a character in the input string and each column corresponds to a character in the output string. A darker color in a cell of the attention matrix indicates a greater attention paid by the model to the corresponding input character when generating the corresponding output character.

Running unit tests

The data and model code in this example are covered by unit tests. To run the unit tests:

cd ../
yarn
cd date-conversion-attention
yarn
yarn test