Input structure for BiLSTM

I keep getting this error when i run my BiLSTM model "Error using trainNetwork (line 165)

Invalid training data. Predictors must be a N-by-1 cell array of sequences, where N is the number of

sequences. All sequences must have the same feature dimension and at least one time step."

My X_train data is a 134949x1 cell with Nx1 nested cells that hold cells with 70x1 double

and my Y_train data is a 134949x1 cell with Nx1 nested cells that hold cells with 65x1 double.

I do not know what is wrong as the error says it should be N-by-1 cell array and I believe that is how my data is stored.

[Pravin Jagtap]

Hello Sanjana,

I would like to know the way you maintained the structure of data has any significance? I need some more information on the description of data than only sizes of cell arrays so that we can think of reshaping/unwrapping the data in the required form.

Kind Regards

~Pravin

[Sanjana Sankar]

Hello!

Yes, basically I am storing data as words, split down to letters for the input and correspondingly, into phoneme in the output. Hence I cannot just expand all the cells. I need the demarcation between words.

Answer 1

[Pravin Jagtap]

Hello Sanjana,

Refer to following example in order to understand the desired structures of cell arrays of training data for Bi-LSTM model by going through the example:

https://www.mathworks.com/help/deeplearning/examples/classify-sequence-data-using-lstm-networks.html;jsessionid=dce7b01b40907817f81d4af4f702

In your case, there are 3 nested cell arrays whereas example mentioned above uses 2 nested cell arrays. Important thing to note here is we can have 2-D cell structure inside the main cell array. In your case you can combine N X 1 and 70 X 1 into one cell array 2-D cell array) which will be consistent with the desired structure of cell array as per above example.

To achieve the unwrapping/reshaping you can use the horizontal or vertical concatenation using following code:

%% Generating the Data
>>data={{1;2;3};{4;5;6};{7;8;9}};

%% Vertical concatenation
>>cell2mat(vertcat(data{:}));

%% Horizontal concatenation
>>cell2mat(horzcat(data{:}));

In general, for Bi-LSTM, make sure ‘X_Train‘ should have cell array size of ‘n X 1’ and each cell represents ‘f X l’ cell array size where f is number of features and l is length of each cell array data

( Note: f should be fixed whereas l can vary and follow the example given above to understand the desired structure of cell array for Bi-LSTM model.)

Kind Regards

~Pravin