3
$\begingroup$

I watched the this video on YouTube and copied the code. Why it does not work in M12? I mean net[inData] and rnnLayerForward[param, inData, state] don't give the same result? What change since then and how can I fix it? Slides can be downloaded here.

    rnnTimeStep[params_Association, input_, state_] := 
      Tanh[params["StateWeights"].state + params["InputWeights"].input + 
        params["Biases"]];
    rnnLayerForward[params_Association, input_, state_] := 
     Rest@FoldList[rnnTimeStep[params, #2, #1] &, state, input]

sequenceLength = 3;
featureSize = 2;
rnnStateSize = 1;
inData = RandomReal[1, {sequenceLength, featureSize}];
state = ConstantArray[0, rnnStateSize];

net = NetInitialize@
  BasicRecurrentLayer[rnnStateSize, "Input" -> Dimensions@inData]

param = Normal@NetExtract[net, "Arrays"]

net[inData]

rnnLayerForward[param, inData, state]
$\endgroup$
  • $\begingroup$ param = Normal[#] & /@ NetExtract[net, "Arrays"] $\endgroup$ – Alexey Golyshev Jul 26 at 7:00
  • $\begingroup$ Thanks. Can you please post as an answer, so I can accept it. param = Normal/@ NetExtract[net, "Arrays"] is also fine. $\endgroup$ – OkkesDulgerci Jul 26 at 13:48
3
$\begingroup$
rnnTimeStep[params_Association, input_, state_] := 
  Tanh[params["StateWeights"].state + params["InputWeights"].input + params["Biases"]];

rnnLayerForward[params_Association, input_, state_] := 
 Rest@FoldList[rnnTimeStep[params, #2, #1] &, state, input]

sequenceLength = 3;
featureSize = 2;
rnnStateSize = 1;
inData = RandomReal[1, {sequenceLength, featureSize}];
state = ConstantArray[0, rnnStateSize];

net = NetInitialize@BasicRecurrentLayer[rnnStateSize, "Input" -> Dimensions@inData];

param = Normal[#] & /@ NetExtract[net, "Arrays"];
net[inData]

{{-0.85818}, {-0.987016}, {-0.941039}}

rnnLayerForward[param, inData, state]

{{-0.85818}, {-0.987016}, {-0.941039}}

$\endgroup$

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.