Error for dlarray format, but why?

6 views (last 30 days)
Song Decn
Song Decn on 1 Jun 2021
Answered: Ben on 20 Jun 2023
K>> lstm(dlX, hiddenState, initialCellState, inputWeights, ...
recurrentWeights, bias)
Error using deep.internal.dlarray.validateWeights (line 9)
'U' dimension (if not a formatted dlarray, second dimension) of weights must have size
NumFeatures, where NumFeatures is the size of the 'C' dimension of the input data.
Can any expert help me to solve this issue? Also I am still quite confused about the concept with the format labels C S T U B
Is there any simple explanation for tutorial for their usage?
Many thks
  1 Comment
Matt J
Matt J on 18 Jun 2023
Edited: Matt J on 18 Jun 2023
We need to examine dims(dlX) and the sizes of all your input variables.

Sign in to comment.

Answers (1)

Ben
Ben on 20 Jun 2023
This error appears to be thrown if the inputWeights have the wrong size, e.g. you can take this example code from help lstm
numFeatures = 10;
numObservations = 32;
sequenceLength = 64;
X = dlarray(randn(numFeatures,numObservations,sequenceLength), 'CBT');
% Create formatted dlarrays for the lstm parameters with three
% hidden units.
numHiddenUnits = 3;
H0 = dlarray(randn(numHiddenUnits,numObservations),'CB');
C0 = dlarray(randn(numHiddenUnits,numObservations),'CB');
weights = dlarray(randn(4*numHiddenUnits,numFeatures),'CU');
recurrent = dlarray(randn(4*numHiddenUnits,numHiddenUnits),'CU');
bias = dlarray(randn(4*numHiddenUnits,1),'C');
% Apply an lstm calculation
[Y,hiddenState,cellState] = lstm(X,H0,C0,weights,recurrent,bias);
If you now make weights the wrong size in the 2nd dimension you get the error:
errorWeights = dlarray(randn(4*numHiddenUnits,numFeatures+1),'CU');
lstm(X,H0,C0,errorWeights,recurrent,bias); % throws error
This suggests your inputWeights have the wrong size to use lstm. The inputWeights require a size of 4*NumHiddenUnits x NumFeatures, and they can either be a dlarray with format labels or without:
% both of these are valid - the format label U is just to specify that this
% dimension doesn't correspond to any of the standard named labels S -
% spatial, C - channel, T - time, B - batch.
weights = dlarray(randn(4*numHiddenUnits,numFeatures),'CU');
weights = dlarray(randn(4*numHiddenUnits,numFeatures));
If you list the sizes as @Matt J says then we can debug the issue further.

Tags

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!