Documentation

[jayenashar]

Is the README the only documentation? In particular, I'm looking to create a tiny_dnn-compatible 2d convolutional layer using CompiledNN, in order to benchmark it against a new one I created "tiny_jnn", but I don't want to replace the rest of our pipeline - just the one layer.

[fthielke]

Yes, currently our only documentation are the Readme and comments in the code.

Afaik, tinydnn's convolutional layer supports the same parameters as our Conv2D, so I suppose you would want to instantiate a CompiledNN model with just that one layer.
The problem is though that we currently only support loading Keras models without the possibility of constructing a Model yourself, which admittedly is a needless constraint.

[jayenashar]

I think the only extra parameter in tiny_dnn would be dilation, which I saw somewhere is not supported in CompiledNN. But that's ok as most of our layers have no dilation.

I thought I could use the Conv2DCompiler class directly, but I couldn't figure out how. I looked at https://github.com/bhuman/CompiledNN/blob/master/Src/CompiledNN/CompiledNN.cpp#L160 as well.

[ahasselbring]

You can also compile a single node, see for example https://github.com/bhuman/CompiledNN/blob/master/Tests/Layers/UpSampling2D.cpp. I have an unfinished Conv2D test, this an excerpt from it:

  static const Node& buildNode(Conv2DLayer* l, const std::array<unsigned int, 2>& strides, const std::array<unsigned int, 2>& kernelSize,
                               bool hasBiases, ActivationFunctionId activation, PaddingType padding, unsigned int filters,
                               unsigned int height, unsigned int width, unsigned int channels)
  {                               
    l->nodes.clear();
    l->strides = strides;                                           
    l->weights.reshape(kernelSize[0], kernelSize[1], channels, filters);
    l->biases.resize(hasBiases ? filters : 0);
    l->hasBiases = hasBiases;                                         
    l->activationId = activation;
    l->padding = padding;
     
    l->nodes.emplace_back(l);
    Node& n = l->nodes.back();
    n.inputs.emplace_back(nullptr, 0, 0);
    n.inputDimensions.push_back({height, width, channels});
    l->calcOutputDimensions(n);                 
    for(std::size_t i = 0; i < n.outputDimensions.size(); ++i)
      n.outputs.emplace_back(l, 0, i);
    return n;
  }

called by

CompiledNN c;
Conv2DLayer l;
const Node& n = buildNode(&l, {stride, stride}, {kernelSize, kernelSize}, true, ...);
// ... copy weights to l.weights, copy biases to l.biases
c.compile(n, CompilationSettings());
// ... fill c.input(0)
c.apply();
// ... obtain output from c.output(0)

[jayenashar]

oh fantastic. i will figure out how to make this work as a tiny_dnn layer and report back.