DEMO.md

Load TensorFlow models with deeplearn-tensorflow

This tutorial demonstrates how to load a TensorFlow model with deeplearn-tensorflow directly into deeplearn.js.

We will modify the existing MNIST demo to simplify the needed steps.

Prepare and build the MNIST demo

Clone the deeplearn.js project

To build the MNIST demo, we need to clone the source code with:

git clone https://github.com/PAIR-code/deeplearnjs.git
cd deeplearnjs

Install yarn

Yarn is a requirement of deeplearn.js and could be installed with the following command:

npm install --global yarn

Train the MNIST model

Please make sure you have Python and TensorFlow installed to train the MNIST model by running:

python demos/mnist/fully_connected_feed.py

Copy the trained MNIST model

The following command will copy the trained model to the mnist demo dirctory:

cp -R /tmp/tensorflow/mnist/logs/fully_connected_feed demos/mnist/

(Note: The trained model could be at a different location depending on the used operation system.)

Process dependencies

To load and resolve all dependencies run the following npm command:

npm run prep

Add deeplearn-tensorflow dependencies over yarn

To be able to use the deeplearn-tensorflow integration add deeplearn-tensorflow as dependencies for the demos with:

cd demos
yarn add deeplearn-tensorflow
cd ..

Prepare and run the MNIST demo

First run the unchanged MNIST demo to make sure everything is setup correctly:

./scripts/watch-demo demos/mnist

To view the MNIST demo visit http://127.0.0.1:8080/demos/mnist/ in your browser.

Modify demos/mnist/mnist.ts

To load the trained TensorFlow model with deeplearn-tensorflow, the following modifications needs to be done for demos/mnist/mnist.ts.

Remove 'CheckpointLoader'

// tslint:disable-next-line:max-line-length
- import {Array1D, Array2D, CheckpointLoader, ENV, NDArray, Scalar} from 'deeplearn';
+ import {Array1D, Array2D, ENV, NDArray, Scalar} from 'deeplearn';

Import 'TensorflowLoader'

import {Array1D, Array2D, ENV, NDArray, Scalar} from 'deeplearn';
+ import {TensorflowLoader} from 'deeplearn-tensorflow';

Replace existing CheckpointLoader with TensorflowLoader

- // manifest.json lives in the same directory as the mnist demo.
- const reader = new CheckpointLoader('.');
- reader.getAllVariables().then(vars => {

+ // Tensorflow reader demo.
+ const tensorflowReader = new TensorflowLoader(NDArray);
+ tensorflowReader.loadRemoteFiles('fully_connected_feed/model.ckpt-999').then((vars) => {

Reload MNIST demo

Visit http://127.0.0.1:8080/demos/mnist/ to see the new loader in action.