12
$\begingroup$

There is an great example of exporting a NetGraph (that was trained in Mathematica) into the MXNet's format.

However, I think there's a bug or something is wrong with the answer, because once exported the file refuses to load into mxnet.

I've cleaned the old answer (which is broken) into something that works:

<< MXNetLink`;
<< NeuralNetworks`;
<< GeneralUtilities`;
net = NetChain[{ConvolutionLayer[20, {5, 5}], ElementwiseLayer[Ramp], 
   PoolingLayer[{2, 2}, {2, 2}], ConvolutionLayer[50, {5, 5}], 
   ElementwiseLayer[Ramp], PoolingLayer[{2, 2}, {2, 2}], 
   FlattenLayer[], DotPlusLayer[500], ElementwiseLayer[Ramp], 
   DotPlusLayer[10], SoftmaxLayer[]}, 
  "Output" -> NetDecoder[{"Class", Range[0, 9]}], 
  "Input" -> NetEncoder[{"Image", {28, 28}, "Grayscale"}]]
resource = ResourceObject["MNIST"];
trainingData = ResourceData[resource, "TrainingData"];
testData = ResourceData[resource, "TestData"];
trained = 
 NetTrain[net, trainingData, ValidationSet -> testData, 
  MaxTrainingRounds -> 3]
jsonPath = "~/MNIST-symbol.json";
paraPath = "~/MNIST-0000.params";
Export[jsonPath, ToMXJSON[trained][[1]], "String"];
f[str_] := 
  If[StringFreeQ[str, "Arrays"], str, 
   StringReplace[
    StringSplit[str, ".Arrays."] /. {a_, b_} :> 
      StringJoin[{"arg:", a, "_", b}], {"Weights" -> "weight", 
     "Biases" -> "bias"}]];
plan = ToMXPlan[trained];
NDArrayExport[paraPath, 
 NDArrayCreate /@ KeyMap[f, plan["ArgumentArrays"]]]

Now, yes we can load it back into MMA alright, but when it comes time to try it with MXNet for reals, we start hitting various snags when trying to load the symbol/params into a module so we can actually use the network:

import mxnet as mx (* python code that loads an runs the model *)
sym, arg_params, aux_params = mx.model.load_checkpoint('MNIST', 0)
from sklearn.datasets import fetch_mldata
mnist = fetch_mldata('MNIST original')
batch_size = 2
iter = mx.io.NDArrayIter(mnist['data'][:6], mnist['target'][:6], batch_size)
mod = mx.mod.Module(symbol=sym, context=mx.cpu(), data_names=['Input'], label_names=None)
mod.bind(for_training=False, data_shapes = [('Input', (batch_size, 1, 28, 28))])
mod.set_params(arg_params, aux_params)
print(mod.predict(iter))

enter image description here enter image description here

I've heard this functionality is coming in future versions, but I'd really like to start using the nets I've trained in a real setting. Any ideas?

$\endgroup$
  • 1
    $\begingroup$ It seems that the internal structure of the neural network framework is under constant development. Many things have changed from version 11.0 to 11.1.1. It would be best if this answer can be addressed by wolfram employees who knows more about the internal structure. $\endgroup$ – xslittlegrass Jun 24 '17 at 5:44
  • 2
    $\begingroup$ same problem.Even the net is simple.net = NetInitialize@NetChain[{5, 3}, "Input" -> 2]; $\endgroup$ – partida Sep 5 '17 at 5:35
4
$\begingroup$

In my question Example of NetChain running in MXNet cause Error and re-post in Wolfram Community

Thanks for Mike Sollami gave nice code to get the perfect result.

enter image description here

enter image description here

enter image description here

enter image description here

$\endgroup$
1
$\begingroup$

The best way to do that, it is change the way of exporting by Mathematica and it is fully explained in https://mathematica.stackexchange.com/a/162479/54287

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.