MATLAB Answers

Sofía
1

Is it possible to implement a LSTM layer after a CNN layer?

Asked by Sofía
on 26 Apr 2018
Latest activity Answered by Shounak Mitra on 11 Jul 2019
I'm trying to implement a CNN layer + a LSTM layer, but I have an error: "Network: Incompatible layer types". Is it not possible to implement this combination in MATLAB or am I just writing it not properly?
My code:
layers = [ ...
sequenceInputLayer(inputSize)
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2)
lstmLayer(numHiddenUnits,'OutputMode','last')
fullyConnectedLayer(numClasses)
softmaxLayer
classificationLayer
];
Error:
Error using trainNetwork (line 154)
Invalid network.
Caused by:
Network: Incompatible layer types. The network contains layer types not supported with recurrent layers.
Detected recurrent layers:
layer 6 (LSTM)
Detected incompatible layers:
layer 2 (Convolution)
layer 3 (Batch Normalization)
layer 5 (Max Pooling)
Layer 2: Input size mismatch. Size of input to this layer is different from the expected input size.
Inputs to this layer:
from layer 1 (output size 500)

  1 Comment

I would like to do this as well.

Sign in to comment.

5 Answers

Answer by Mona
on 19 Sep 2018
 Accepted Answer

As far as I know, no, you can't combine the two. You can train a CNN independently on your training data, then use the learned features as an input to your LSTM. However, learning and updating CNN weights while training an LSTM is unfortunately not possible.

  0 Comments

Sign in to comment.


Answer by charu
on 9 Jul 2018

use bilstmLayer layer instead of lstm layer as in example
inputSize = 12;
numHiddenUnits = 100;
numClasses = 9;
layers = [ ...
sequenceInputLayer(inputSize)
bilstmLayer(numHiddenUnits,'OutputMode','last')
fullyConnectedLayer(numClasses)
softmaxLayer
classificationLayer]

  1 Comment

I have the same problem by using a bilstm Layer. The error message is :
if true
Error using trainNetwork (line 154)
Invalid network.
Error in test_spa_REG (line 168)
net = trainNetwork(XTR,TTR,Layers,options);
Caused by:
Network: Incompatible layer types. The network contains layer types not supported with recurrent layers.
Detected recurrent layers:
layer 9 (BiLSTM)
Detected incompatible layers:
layer 1 (Image Input)
layer 2 (Transposed Convolution)
layer 'temp1' (Convolution)
layer 5 (Average Pooling)
and 1 other layers.
Layer 10: Input size mismatch. Size of input to this layer is different from the expected input size.
Inputs to this layer:
from layer 9 (output size 20)
Is it possible to combine CNN with LSTM layer ?

Sign in to comment.


Answer by sotiraw sotiroglou on 24 Mar 2019

Matlab 2019a is out. And it claims it can do this cnn - rnn combination.
Could someone give us an example?

  0 Comments

Sign in to comment.


Answer by sotiraw sotiroglou on 24 Mar 2019

Matlab 2019a is out there , and it claims it can do this rnn cnn combination.
I dont know the details, but i write this answer to encourage everyone with the same issue to search and maybe help with an example

  0 Comments

Sign in to comment.


Answer by Shounak Mitra on 11 Jul 2019

Hello Everyone,
As of 19a, MATLAB supports workflows containing both CNN and LSTM layers.
Please check the link that contains an example showing the CNN+LSTM workflow --> https://www.mathworks.com/help/deeplearning/examples/classify-videos-using-deep-learning.html

  0 Comments

Sign in to comment.