Hi, everyone! dlgradient throws an error message as Undefined function 'dlgradient' for input arguments of type 'cell'.

I am new to NN tools and functions in MatLab, so probably there is a simple mistake I am doing. I have the the code that follows. I had this code running, but then I spotted few bugs in other parts of the code, that I fixed, probably something is linked to dlgradient. I created the weights and biases arrays by myself, not using any off-the-shelf tools of Matlab. The code below are a part of a class where I try to manage all NN computations. Thank you all in advance.

Alim

function [totalLoss, result] = trainingByBatches(obj,data,coef)

totalLoss = 0;

NoOfRecords = length(data(1,:));

BatchSize = floor(NoOfRecords*obj.batchSize);

NoOfBatches = floor(NoOfRecords/BatchSize);

params = coef; % cells of 6 arrays. There are two hidden layers 1 input and 1 output so 3 arrays of weight and 3 vectors for biases

for i=0:NoOfBatches-1

batch=data(:,i*BatchSize+1:(i+1)*BatchSize); %slide through data to extract batches

times = obj.t(i*BatchSize+1:(i+1)*BatchSize); %time coordinates for each data array

batch = [times;batch]; %combine time coordinates and data itself for decreasing # of parameters to pass

ll = [0;obj.l]; %space coordinates for each data record in batch

batch = [ll, batch]; %combine space coordinates with data for reducing number of parameters to pass

[grad, loss] = dlfeval(@train,obj,batch,params); %wraping in dlfeval

for jj = 1:obj.NL-1 %gradient descent workflow obj.LR - learning rate

params{1,jj} = params{1,jj} - obj.LR * grad{1,jj};

params{2,jj} = params{2,jj} - obj.LR * grad{2,jj};

end

totalLoss = totalLoss + loss;

end

result = params;

end

function [grads, loss] = train(obj, data, params) %a fuction to find gradient from

l = data(2:end,1); %x coordinates - extracting space coordinates

times = data(1,2:end); %time coordinates - extracting time coordinates

batch = data(2:end,2:end); % temperature values - extracting the data itself

curLayer = [batch'; l']; %input information for NN

output = NN(params,curLayer); %this function just does forward pass of input values through NN

loss = piDiffEqn(batch,output,obj.l,times); % this is a loss function that returns a double

grads = dlgradient(loss,params); % HERE I HAVE ERROR MESSAGE

end

[Dyuman Joshi]

You should check what the syntax and the data type of inputs is expected to be for dlgradient. Go through the linked documentation page.

[Alim]

Thank you. I assume that the problem is something you mentioned, but the code worked with the parameters as it is now. I am looking through the documentation. thank you

Answer 1

[Alim]

Problem solved. batch variable in the code above needs to be wrapped into dlarray.