Add generic file for NNet

Author: mikhail

pytorch/__init__.py

@@ -0,0 +1,80 @@
+import numpy
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self, game, args, layers: int = 4):
+        # game params
+        self.board_x, self.board_y = game.getBoardSize()
+        self.action_size = game.getActionSize()
+        self.args = args
+
+        # nnet params
+        self.layers = layers
+        self.conv = []
+        self.batchnorm = []
+        self.fc = []
+        self.fcbn = []
+
+        super(Model, self).__init__()
+
+        self._setup()
+
+    def _setup(self):
+        # Create Conv layers
+        in_channels = 1
+        kernel_size = int(float(min(self.board_x, self.board_y)) / self.layers)
+        # if kernel_size < 3:
+        kernel_size = 3
+        paddings = [1, 1, 0, 0]
+        for i in range(self.layers):
+            conv = nn.Conv2d(in_channels, self.args.num_channels, kernel_size, stride=1, padding=paddings[i])
+            self.add_module(f'conv{i}', conv)
+            self.conv.append(conv)
+            in_channels = self.args.num_channels
+
+        # Prepare Batch Normalization
+        for i in range(self.layers):
+            bn = nn.BatchNorm2d(self.args.num_channels)
+            self.batchnorm.append(bn)
+            self.add_module(f'batchnorm{i}', bn)
+
+        # Prepare features
+        in_features = self.args.num_channels * (self.board_x - self.layers) * (self.board_y - self.layers)
+        # self.fc1 = nn.Linear(self.args.num_channels * (self.board_x-4)*(self.board_y-4), 1024)
+
+        out_features = 0
+        for i in range(self.layers - 2):
+            out_features = (self.args.feature_multiplier * (self.layers - 2 - i))
+            linear = nn.Linear(in_features, out_features)
+            self.fc.append(linear)
+            self.add_module(f'fc{i}', linear)
+            bn = nn.BatchNorm1d(out_features)
+            self.fcbn.append(bn)
+            self.add_module(f'batchnorm1d{i}', bn)
+            in_features = out_features
+
+        self.fc_pi = nn.Linear(out_features, self.action_size)
+        self.fc_v = nn.Linear(out_features, 1)
+
+    def forward(self, s: torch.Tensor):
+        s = s.view(-1, 1, self.board_x, self.board_y)
+
+        for i in range(self.layers):
+            s = F.relu(self.batchnorm[i](self.conv[i](s)))
+
+        size = int(numpy.prod(list(s.size())))
+        s = s.view(-1, size)
+
+        fs = self.fc[0](s)
+        bs = self.fcbn[0](fs)
+        tensor = F.relu(bs)
+        s = F.dropout(tensor, p=self.args.dropout, training=self.training)
+        s = F.dropout(F.relu(self.fcbn[1](self.fc[1](s))), p=self.args.dropout, training=self.training)
+
+        pi = self.fc_pi(s)
+        v = self.fc_v(s)
+
+        return F.log_softmax(pi, dim=1), torch.tanh(v)