| """Compact ResNet policy/value network for the AChess distillation dataset.""" |
|
|
| from __future__ import annotations |
|
|
| import torch |
| from torch import nn |
| from torch.nn import functional as F |
|
|
|
|
| class ResBlock(nn.Module): |
| def __init__(self, channels: int): |
| super().__init__() |
| self.bn1 = nn.BatchNorm2d(channels) |
| self.conv1 = nn.Conv2d(channels, channels, 3, padding=1, bias=False) |
| self.bn2 = nn.BatchNorm2d(channels) |
| self.conv2 = nn.Conv2d(channels, channels, 3, padding=1, bias=False) |
| nn.init.zeros_(self.bn2.weight) |
|
|
| def forward(self, x): |
| h = F.relu(self.bn1(x), inplace=True) |
| h = self.conv1(h) |
| h = self.bn2(h) |
| h = self.conv2(h) |
| return F.relu(x + h, inplace=True) |
|
|
|
|
| class ChessNet(nn.Module): |
| def __init__(self, channels: int = 128, blocks: int = 10, planes: int = 20): |
| super().__init__() |
| self.stem = nn.Sequential( |
| nn.Conv2d(planes, channels, 3, padding=1, bias=False), |
| nn.BatchNorm2d(channels), |
| nn.ReLU(inplace=True), |
| ) |
| self.body = nn.Sequential(*[ResBlock(channels) for _ in range(blocks)]) |
| |
| self.policy = nn.Conv2d(channels, 64, 1) |
| self.value_conv = nn.Conv2d(channels, 1, 1) |
| self.value_bn = nn.BatchNorm2d(1) |
| self.value_fc = nn.Linear(64, 256) |
| self.value_out = nn.Linear(256, 1) |
| self.wdl_out = nn.Linear(256, 3) |
| self._init_weights() |
|
|
| def _init_weights(self): |
| for module in self.modules(): |
| if isinstance(module, nn.Conv2d): |
| nn.init.kaiming_normal_(module.weight, nonlinearity="relu") |
| elif isinstance(module, nn.Linear): |
| nn.init.xavier_uniform_(module.weight) |
| nn.init.zeros_(module.bias) |
|
|
| def forward(self, x, legal_mask=None): |
| h = self.body(self.stem(x)) |
| logits = self.policy(h).flatten(1) |
| if legal_mask is not None: |
| logits = logits.masked_fill(~legal_mask, -1e4) |
| v = F.relu(self.value_bn(self.value_conv(h)), inplace=True).flatten(1) |
| v = F.relu(self.value_fc(v), inplace=True) |
| value = torch.tanh(self.value_out(v)).squeeze(-1) |
| wdl = self.wdl_out(v) |
| return logits, value, wdl |
|
|
