Why do I need to add "+1" to the input of my Keras LSTM-model?

Question

MathWorks Support Team on 8 Oct 2021

0
Link

Direct link to this question

https://nl.mathworks.com/matlabcentral/answers/1469416-why-do-i-need-to-add-1-to-the-input-of-my-keras-lstm-model

Edited: MathWorks Support Team on 31 May 2022

The LSTM-model is trained in Keras. The input comes from 'datfile'. If the trained net is imported to Matlab and the inputs from 'datfile' are added, the outputs of the layers do not match. If '1' is added to the inputs, then they match. Why must '1' be added for the same results?

Sign in to answer this question.

Answer 1

MathWorks Support Team on 31 May 2022

0
Link

Direct link to this answer

https://nl.mathworks.com/matlabcentral/answers/1469416-why-do-i-need-to-add-1-to-the-input-of-my-keras-lstm-model#answer_804021

Edited: MathWorks Support Team on 31 May 2022

Essentially, a 'wordEmbeddingLayer' acts as a lookup table. We assign each word a single index. And that index is used by the lookup table in order to get the corresponding embedding vector. Now the reason why this +1 is required is due to the fact that MATLAB uses 1-based indexing rather than 0-based indexing like python. Because of this, each index is shifted by 1 when we load in the lookup table, so the input data should also be shifted by 1 to accommodate this.

This has also been documented in the "Tips" section for importKerasNetwork, importKerasLayers, importTensorFlowNetwork, and importTensorFlowLayers:

https://www.mathworks.com/help/deeplearning/ref/importkerasnetwork.html#d123e88361