This repo provides the legacy loader for porting a TensorFlow model to deeplearn.js. Note that this loader is deprecated.

import {CheckpointLoader} from 'deeplearn-legacy-loader';
const reader = new CheckpointLoader(pathToCheckpointDir);
reader.getAllVariables().then(vars => {
  // vars maps a variable name to a Tensor.
});

All the necessary resources used in this tutorial are stored in the demo/ directory.

To demonstrate the porting steps, we will use a fully connected neural network that predicts hand-written digits from the MNIST dataset. The code is forked from the official TensorFlow MNIST tutorial.

Before we start, make sure you have TensorFlow installed.

First, we clone this repository. We cd into the base dir and train the model in TensorFlow by running:

python demo/fully_connected_feed.py

The training should take ~1 minute and will store a model checkpoint in /tmp/tensorflow/mnist/tensorflow/mnist/logs/fully_connected_feed/.

Next, we need to port the weights from the TensorFlow checkpoint to a format the loader undestands. We provide a script that does this. We run it from the base directory:

python python/dump_checkpoint_vars.py \
  --model_type=tensorflow \
  --output_dir=demo/ \
  --checkpoint_file=/tmp/tensorflow/mnist/logs/fully_connected_feed/model.ckpt-1999

The script will save a set of files (one file per variable, and a manifest.json) in the demo/ directory. The manifest.json is a simple dictionary that maps variable names to files and their shapes:

{
  ...,
  "hidden1/weights": {
    "filename": "hidden1_weights",
    "shape": [784, 128]
  },
  ...
}

To read the weights, we need to create a CheckpointLoader and point it to the manifest file. We then call loader.getAllVariables() which returns a dictionary that maps variable names to Tensors. At that point, we are ready to write our model. Here is a snippet demonstrating the use of CheckpointLoader:

import * as dl from 'deelearn';
import {CheckpointLoader} from 'deeplearn-legacy-loader';

// manifest.json is in the same dir as index.html.
const varLoader = new CheckpointLoader('.');

varLoader.getAllVariables().then(vars => {
  // Get Tensor of variables casted with expected dimension.
  const hidden1W = vars['hidden1/weights'];
  const hidden1B = vars['hidden1/biases'];

  // Write your model here...
});

For details regarding the full model code, see demo/mnist.ts.

To run the mnist demo, run yarn run-demo from the base dir. This compiles the typescript code and runs an http-server on port 8080 that serves the static html/js files.

yarn
yarn run-demo

>> Starting up http-server, serving demo/
>> Available on:
>>   http://127.0.0.1:8080
>>   http://192.168.1.136:8080
>> Hit CTRL-C to stop the server

You should see a simple page showing test accuracy of ~90% measured using a test set of 50 mnist images stored in demo/sample_data.json.