Matlab R2016a - MacOsX - Neural tool - Index exceeds matrix dimensions.

I am using Matlab R2016a for MacOsX and I'm having an Index exceeds matrix dimensions.

Can you fix this problem?

Thanks,

>> load house_dataset;
>> whos
  Name               Size             Bytes  Class     Attributes

    houseInputs       13x506            52624  double              
    houseTargets       1x506             4048  double              

>> net=newfit(houseInputs, houseTargets, 20);
Index exceeds matrix dimensions.

Error in initnw>initialize_layer (line 168)
       range(inputStart(j):inputStop(j),:) =
       temp2((inputStart(j):inputStop(j))-inputStart(j)+1,:);

Error in initnw (line 93)
    out1 = initialize_layer(in1,in2);

Error in initlay>initialize_network (line 147)
      net = feval(initFcn,net,i);

Error in initlay (line 89)
    out1 = initialize_network(in1);

Error in network/init (line 32)
  net = feval(initFcn,net);

Error in newff>new_5p1 (line 235)
net = init(net);

Error in newff>create_network (line 129)
  net = new_5p1(varargin{:});

Error in newff (line 101)
  out1 = create_network(varargin{:});

Error in newfit>create_network (line 98)
net = newff(varargin{:});

Error in newfit (line 79)
  out1 = create_network(varargin{:});

>>

[Walter Roberson]

Try transposing your input to be 506 x 13

[Orlando Stein Junior]

Making transposing did not work. Same thing.

[Greg Heath]

I did both without error.

Greg

[Orlando Stein Junior]

For me, didn't work. Do you trying in a Mac or PC version?

[Orlando Stein Junior]

In my initnw.m, is this the i have the initialize_layer that occurs the error. Is the same that you have in your version Greg?

I attached the initnw.m module of my installation. Maybe help.

%%SUPPORT FUNCTIONS

function net = initialize_layer(net,i)

% Calculate source indices
inputInds = find(net.inputConnect(i,:));
numInputs = length(inputInds);
layerInds = find(net.layerConnect(i,:));
numLayers = length(layerInds);

% Get source sizes and delays
inputSizes = zeros(numInputs,1);
inputDelays = zeros(numInputs,1);
for j=1:numInputs
  inputDelays(j) = length(net.inputWeights{i,inputInds(j)}.delays);
  inputSizes(j) = net.inputWeights{i,inputInds(j)}.size(2);
end
totalInputSize = sum(inputSizes);

layerSizes = zeros(numLayers,1);
layerDelays = zeros(numInputs,1);
for j=1:numLayers
  layerDelays(j) = length(net.layerWeights{i,layerInds(j)}.delays);
  layerSizes(j) = net.layerWeights{i,layerInds(j)}.size(2);
end
totalLayerSize = sum(layerSizes);

totalSourceSize = totalInputSize + totalLayerSize;

% Calculate range indices
inputStart = cumsum([1; inputSizes]);
inputStop = cumsum(inputSizes);
layerStart = cumsum([1; layerSizes])+totalInputSize;
layerStop = cumsum(layerSizes)+totalInputSize;

% Get source ranges
range = zeros(totalSourceSize,2);
for j=1:numInputs
  irange = net.inputs{inputInds(j)}.processedRange;
  if ~isempty(irange)
    % ODJ 4/1/02 Avoid problem with delays and one column weights
    temp1=size(irange,1)*inputDelays(j);
    if temp1~= inputStop(j)-inputStart(j)-1
       temp2 = repmat(irange,inputDelays(j),1);
       range(inputStart(j):inputStop(j),:) = temp2((inputStart(j):inputStop(j))-inputStart(j)+1,:);
    else  
       range(inputStart(j):inputStop(j),:) = repmat(irange,inputDelays(j),1);
    end
  end
end
for j=1:numLayers
  lrange = feval(net.layers{layerInds(j)}.transferFcn,'output');
  if any(~isfinite(lrange))
    lrange = [max(lrange(1),-1) min(lrange(2),1)];
  end
  range(layerStart(j):layerStop(j),:) = lrange(ones(layerSizes(j),1),:);
end

% Get transferFcn info
transferFcn = net.layers{i}.transferFcn;
active = feval(transferFcn,'active');

% Check layer and sources for compatibility with Nguyen-Widrow method
ok = 1;
if ~net.biasConnect(i), ok = 0; end
if ~all(isfinite(active)), ok = 0; end
if ~strcmp(net.layers{i}.netInputFcn,'netsum'), ok = 0; end
for j=1:numInputs
  if ~strcmp(net.inputWeights{i,inputInds(j)}.weightFcn,'dotprod')
    ok = 0;
  end
end
for j=1:numLayers
  if ~strcmp(net.layerWeights{i,layerInds(j)}.weightFcn,'dotprod')
    ok = 0;
  end
end

% Use Nguyen-Widrow method if network checks out ok
if ok
  [w,b] = calcnw(range,net.layers{i}.size,active);

% Otherwise use RANDS
else
  sizeRows = 0;
  for j=1:numInputs
    if(net.inputWeights{i,inputInds(j)}.size(1)>sizeRows)
      sizeRows = net.inputWeights{i,inputInds(j)}.size(1);
    end
  end
  for j=1:numLayers
    if(net.layerWeights{i,layerInds(j)}.size(1)>sizeRows)
      sizeRows = net.layerWeights{i,layerInds(j)}.size(1);
    end
  end
  w = rands(sizeRows,totalSourceSize);
  if net.biasConnect(i)
    b = rands(net.layers{i}.size,1);
  end
end

for j=1:numInputs
  net.IW{i,inputInds(j)} = w(1:net.inputWeights{i,inputInds(j)}.size(1),inputStart(j):inputStop(j));
end
for j=1:numLayers
  net.LW{i,layerInds(j)} = w(1:net.layerWeights{i,layerInds(j)}.size(1),layerStart(j):layerStop(j));
end
if net.biasConnect(i)
  net.b{i} = b;
end
end

[Orlando Stein Junior]

[Greg Heath]

newfit is obsolete. Do you get the same result using

 fitnet

or

 feedforwardnet?

[Orlando Stein Junior]

Greg, I have the same issue using grafic tool. ? I included a picture with the variables of your module in matlab , the line that is the problem and you can check that the variable tmp2 line 168 will give the exception because its size is not correct as shown in the workspace variables. I'm doing a very simple example and does not work . For me this is a bug.