Chessboard issue with output layer(s)

[aaparab]

I'm facing a problem with how to organize the last layer. For each of the 64 squares, I need to use a softmax to identify whether the piece in that square is among empty, white king, white queen, ..., black pawn (13 possibilities). The attempt for a vanilla neural network is outlined here. You are welcome to discuss.

[sgsawant]

The link (see outlined here in OP) doesn't work for me.

On a related note, I would say it's better to write a small lightweight function which can tell whether the square is empty or occupied. The primary impact this will have is that more training muscle will be devoted to distinguishing between the remaining classes; information stored within the network will be more relevant to the task at hand & more distributed to the classes.

See LeCun's commentary as well. Pawns occur more frequently than most of the other pieces. The frequency probably needs to be normalized in the training datasets.

Shashank G. Sawant, PhD (352) 226-1688 sgsawant@gmail.com sgsawant@gmail.com linkedin.com/in/sgsawant

On Tue, 24 Nov 2020 at 14:39, Abhishek Parab notifications@github.com wrote:

I'm facing a problem with how to organize the last layer. For each of the 64 squares, I need to use a softmax to identify whether the piece in that square is among empty, white king, white queen, ..., black pawn (13 possibilities). The attempt for a vanilla neural network is outlined here http://./code/golf.ipynb. You are welcome to discuss.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/aaparab/png2pgn/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AANJ5KOPJ3D4QFAIT6BW42LSRQY3BANCNFSM4UBP4NGQ .

[aaparab]

• The link is now updated.
• The idea of writing a small function to identify if a square is empty or not is very good, I can try it.
• Another follow-up is to have a sub-architecture identify the color of the piece once the previous function identifies a nonempty square.
• I would like to use as little computer vision techniques as possible. I would rather have a slightly complex CNN architecture do the above two functions.
• The other idea of weighing pieces by frequency of occurrence is not so good for this problem because these are not valid positions in a chess game. They are random auto-generated positions.
• The question I'm posing here is mainly of how to design (the last layer of) a neural network assuming a non-empty square is found.