reason for subtracting 0.5

[anujgupta82]

https://github.com/tannerbohn/char2vec/blob/master/main.py#L101

Not at all obvious why should one subtract 0.5 from 1-hot vectors in X ?

[tannerbohn]

It's not necessary since the first layer is linear (and even with other activations since we have biases) I probably added that when playing around -- seeing what affected what :p

[anujgupta82]

Cool.
Also did you systematically evaluated the embeddings ?

[tannerbohn]

I have not! I suspect they may be useful for predicting properties/generating encodings for out-of-vocab words... But yet to be explored