search

lifrordi / DeepStack-Leduc

Example implementation of the DeepStack algorithm for no-limit Leduc poker
https://www.deepstack.ai/
894 stars 212 forks source link

How to reduce the GPU memory used when compute preflop strategy in last 20 iterations? #37

Closed Kailianghu closed 5 years ago

Kailianghu commented 5 years ago

@lifrordi hi, In the Deepstack paper said that compute preflop strategy in last 20 iterations was query flop network, which means that input tensor to network size is [22100* 10, 2001] (10 nodes), if the network was 7 layers as the paper said, the need memory is large then 8GB, and they speed is slower than paper said. Do you use suit isomorphism or some other technique to reduce the public cards from 22100 to same value?

Kailianghu commented 5 years ago

I think you use the public state abstraction