mnist demo not working

[dano1234]

No description provided.

[shiffman]

I pushed a "quick fix" by just adding back in deeplearn.js locally to the project. @cvalenzuela, what do you think is the best pathway to end-users to accessing native elements of deeplearn from within coding importing p5ml.js? It could be a static reference like p5ML.deeplearn.Thing but maybe best to keep deeplearn in the global scope?

[cvalenzuela]

Since we can load the complete deeplearn library with p5ML I think making it global makes more sense. I'll push an update to reflect that.

[cvalenzuela]

Also, should we leave deeplearn graph api as it is? or build something around it? The mnist demo is works by building the whole thing in deeplearn:

function buildModelGraph(checkpoints) {
  var g = new deeplearn.Graph();
  var input = g.placeholder('input', [784]);
  var hidden1W = g.constant(checkpoints['hidden1/weights']);
  var hidden1B = g.constant(checkpoints['hidden1/biases']);
  var hidden1 = g.relu(g.add(g.matmul(input, hidden1W), hidden1B));
  var hidden2W = g.constant(checkpoints['hidden2/weights']);
  var hidden2B = g.constant(checkpoints['hidden2/biases']);
  var hidden2 = g.relu(g.add(g.matmul(hidden1, hidden2W), hidden2B));
  var softmaxW = g.constant(checkpoints['softmax_linear/weights']);
  var softmaxB = g.constant(checkpoints['softmax_linear/biases']);
  var logits = g.add(g.matmul(hidden2, softmaxW), softmaxB);
  return [input, g.argmax(logits)];
}

Do you think we should modify/simplify this?

[shiffman]

I'm not sure, but probably? I think we hold off for now as working through other examples and scenarios will inform our thinking. I could imagine a keras-like layer perhaps above deeplearn for architecting a model? We could also do a simple MNIST demo with the NeuralNetwork class perhaps?

I also like the idea of avoiding MNIST and using something more artist-friendly for a classification demo as discussed in #8.

[cvalenzuela]

This relates to the old examples and API.
Closing this since we don't have an MNIST example and https://github.com/CodingTrain/Toy-Neural-Network-JS address this demo better!