stftLayer error: "Size of input in T dimension must be greater than or equal to window length" despite correct input shape

Hi everyone,

I'm encountering a confusing issue when using the stftLayer in MATLAB's Deep Learning Toolbox. I'm trying to pass a dlarray signal to an stftLayer inside a dlnetwork. The input signal has size:

size(signal) - [1 24 2401]

dims(signal) - 'CBT'

The stfftLayer configuration is:

        winLength = 256;
        overlap = 128;
        fftLength = 256;
        stft_net = [sequenceInputLayer(size(signal,1))
            stftLayer("Window",winLength,"FFTLength",fftLength,"OverlapLength",overlap,"Name","stft")];
        stftNet = dlnetwork(stft_net);
        

However, when I am trying to use this model, I am receiving the following error:

        Error using stftLayer/validateProperties (line 238)
Window length must be greater than one sample.

Error in stftLayer (line 166)
            this = validateProperties(this);
                   ^^^^^^^^^^^^^^^^^^^^^^^^
 

The signal properties are shown below for further clarity

K>> size(signal)

ans =

           1          24        2401

K>> dims(signal)

ans =

    'CBT'
    

What could be causing the stftLayer to misinterpret the time dimension and throw the error, even when T is clearly larger than the window length? Any insights would be much appreciated.

Answer 1

[Matt J]

The window must be specified as a vector. Did you want a flat window? If so, then,

signal=rand(1,24,2401);
winLength = 256;
overlap = 128;
fftLength = 256;
stft_net = [sequenceInputLayer(height(signal)),
            stftLayer("Window",ones(1,winLength),"FFTLength",fftLength,...
            "OverlapLength",overlap,"Name","stft")]

stft_net = 
  2×1 Layer array with layers:

     1   ''       Sequence Input   Sequence input with 1 dimensions
     2   'stft'   STFT             stftLayer

[Shubham Baisthakur]

Hi Matt,

Thanks for the answer, however this did not fix the problem. I am still receiving the following error when trying to wrap stft_net in dlnetwork:

>>dlnetwork(stft_net)

Error using dlnetwork/initialize (line 600)
Invalid network.

Error in dlnetwork (line 182)
                net = initialize(net, dlX{:});
                      ^^^^^^^^^^^^^^^^^^^^^^^
Caused by:
    Layer 'stft': Error using the predict function in layer stftLayer. The function threw an error and could not be executed.
        Error using stftLayer/predict (line 183)
        Size of input in T dimension must be greater than or equal to window length.

[Matt J]

signal=rand(1,24,2401);
winLength = 256;
overlap = 128;
fftLength = 256;
stft_net = [sequenceInputLayer(height(signal)),
            stftLayer("Window",ones(1,winLength),"FFTLength",fftLength,...
            "OverlapLength",overlap,"Name","stft")]

stft_net = 
  2×1 Layer array with layers:

     1   ''       Sequence Input   Sequence input with 1 dimensions
     2   'stft'   STFT             stftLayer

stftNet = dlnetwork(stft_net, networkDataLayout(size(signal),'CBT'))

stftNet = 
  dlnetwork with properties:

         Layers: [2×1 nnet.cnn.layer.Layer]
    Connections: [1×2 table]
     Learnables: [1×3 table]
          State: [0×3 table]
     InputNames: {'sequenceinput'}
    OutputNames: {'stft'}
    Initialized: 1

  View summary with summary.

[Matt J]

Or,

signal=rand(1,24,2401);
winLength = 256;
overlap = 128;
fftLength = 256;
stft_net = [sequenceInputLayer(height(signal),MinLength=2041),
            stftLayer("Window",ones(1,winLength),"FFTLength",fftLength,...
            "OverlapLength",overlap,"Name","stft")]

stft_net = 
  2×1 Layer array with layers:

     1   ''       Sequence Input   Sequence input with 1 dimensions
     2   'stft'   STFT             stftLayer

stftNet = dlnetwork(stft_net)

stftNet = 
  dlnetwork with properties:

         Layers: [2×1 nnet.cnn.layer.Layer]
    Connections: [1×2 table]
     Learnables: [1×3 table]
          State: [0×3 table]
     InputNames: {'sequenceinput'}
    OutputNames: {'stft'}
    Initialized: 1

  View summary with summary.

[Shubham Baisthakur]

Thanks, Matt! Setting the MinLength resolved the error. Could you explain why that fixed it?

[Matt J]

It can't initialize the network unless it has some way of knowing how long your time sequences are.