Files
new/card_model/model.py
e2hang ed2fadb625 What
2026-04-20 20:25:35 +08:00

108 lines
3.5 KiB
Python

"""CardModel network architecture for poker equity prediction.
Dual-head model:
- equity_head: Sigmoid output -> scalar equity prediction
- histogram_head: Softmax output -> 50-bin equity distribution
"""
import torch
import torch.nn as nn
from .config import EMBEDDING_DIM, MLP_HIDDEN, NUM_BINS, VOCAB_SIZE, PAD_TOKEN
class CardModel(nn.Module):
"""Neural network predicting equity and equity histogram from cards.
Architecture:
1. Embedding layer for 52 cards + 1 PAD token
2. Hole cards: embed -> sum -> hole_emb
3. Board cards: embed -> sum -> board_emb
4. Concatenate [hole_emb, board_emb] -> MLP backbone
5. Dual output heads:
- equity_head: MLP -> Sigmoid -> scalar
- histogram_head: MLP -> Softmax -> 50-dim distribution
"""
def __init__(
self,
vocab_size: int = VOCAB_SIZE,
embedding_dim: int = EMBEDDING_DIM,
mlp_hidden: list = None,
num_bins: int = NUM_BINS,
):
"""Initialize CardModel.
Args:
vocab_size: Number of token types (52 cards + 1 pad).
embedding_dim: Dimension of card embeddings.
mlp_hidden: List of hidden layer sizes for the backbone MLP.
num_bins: Number of bins for the equity histogram.
"""
super().__init__()
if mlp_hidden is None:
mlp_hidden = list(MLP_HIDDEN)
# Embedding: 53 tokens (0-51 cards + 52 pad) -> embedding_dim
self.embedding = nn.Embedding(vocab_size, embedding_dim, padding_idx=PAD_TOKEN)
# Input to backbone: concat of hole_emb and board_emb
backbone_input_dim = embedding_dim * 2
# Build backbone MLP
layers = []
in_dim = backbone_input_dim
for hidden_dim in mlp_hidden:
layers.append(nn.Linear(in_dim, hidden_dim))
layers.append(nn.ReLU())
layers.append(nn.LayerNorm(hidden_dim))
in_dim = hidden_dim
self.backbone = nn.Sequential(*layers)
# Equity head: scalar output with Sigmoid
self.equity_head = nn.Sequential(
nn.Linear(mlp_hidden[-1], 32),
nn.ReLU(),
nn.Linear(32, 1),
nn.Sigmoid(),
)
# Histogram head: num_bins output with Softmax
self.histogram_head = nn.Sequential(
nn.Linear(mlp_hidden[-1], 64),
nn.ReLU(),
nn.Linear(64, num_bins),
nn.Softmax(dim=-1),
)
def forward(self, x_hole: torch.Tensor, x_board: torch.Tensor):
"""Forward pass.
Args:
x_hole: [batch, 2] int64 tensor of hole card IDs.
x_board: [batch, 5] int64 tensor of board card IDs (padded with 52).
Returns:
pred_equity: [batch, 1] float32, predicted equity in [0, 1].
pred_histogram: [batch, 50] float32, predicted equity distribution (sums to 1).
"""
# Embed hole cards and sum -> [batch, embedding_dim]
hole_emb = self.embedding(x_hole).sum(dim=1)
# Embed board cards and sum -> [batch, embedding_dim]
# PAD_TOKEN (52) has zero embedding due to padding_idx
board_emb = self.embedding(x_board).sum(dim=1)
# Concatenate -> [batch, embedding_dim * 2]
combined = torch.cat([hole_emb, board_emb], dim=-1)
# Backbone
features = self.backbone(combined)
# Dual heads
pred_equity = self.equity_head(features)
pred_histogram = self.histogram_head(features)
return pred_equity, pred_histogram