[Q] If I trained LSTM model with fixed-len input but when predicting I used variable length input is that affect the performance?

nlp
1

I create the training set as a window of 11 size and stride 1 to the end of the corpus like this
x =
[1,2,3,4,5,6…]
[2,3,4,5,6,7…]
[3,4,5,6,7,8…]

y=
[0,0,0,0,0,1…]
[0,0,0,0,1,0…]
[0,0,0,1,0,0…]
For example my training set x shape is 512,11 (512 is bs and 11 is inputs) but when I predict the input shape is 1,20. Is it affect the performance of the model ?. I think that I trained the model with fix-length so in the predicting process I should do the same as training.

English is not my mother tongue if something is not clear pls tell me, Thank you

2

It will have some effect on the performance. But I can’t tell you whether the effect will be good or bad.

Test it.

1 Like