Merge pull request #35 from hijkzzz/dev

add option for no gui in training mode

config.py

@@ -2,6 +2,7 @@
     # gomoku
     'n': 15,                                    # board size
     'n_in_row': 5,                              # n in row
+    'use_gui': True,                           # show gomoku gui in training mode
 
     # mcts
     'libtorch_use_gpu' : True,                  # libtorch use cuda

src/learner.py

@@ -31,6 +31,7 @@ def __init__(self, config):
         # gomoku
         self.n = config['n']
         self.n_in_row = config['n_in_row']
+        self.use_gui = config['use_gui']
         self.gomoku_gui = GomokuGUI(config['n'], config['human_color'])
         self.action_size = config['action_size']
 
@@ -60,13 +61,13 @@ def __init__(self, config):
         self.nnet = NeuralNetWorkWrapper(config['lr'], config['l2'], config['num_layers'],
                                          config['num_channels'], config['n'], self.action_size, config['train_use_gpu'], self.libtorch_use_gpu)
 
+    def learn(self):
         # start gui
-        t = threading.Thread(target=self.gomoku_gui.loop)
-        t.start()
+        if self.use_gui:
+            t = threading.Thread(target=self.gomoku_gui.loop)
+            t.start()
 
-    def learn(self):
         # train the model by self play
-
         if path.exists(path.join('models', 'checkpoint.example')):
             print("loading checkpoint...")
             self.nnet.load_model()
@@ -282,6 +283,10 @@ def get_symmetries(self, board, pi, last_action):
         return l
 
     def play_with_human(self, human_first=True, checkpoint_name="best_checkpoint"):
+        # gomoku gui
+        t = threading.Thread(target=self.gomoku_gui.loop)
+        t.start()
+
         # load best model
         libtorch_best = NeuralNetwork('./models/best_checkpoint.pt', self.libtorch_use_gpu, 12)
         mcts_best = MCTS(libtorch_best, self.num_mcts_threads * 3, \

src/neural_network.py

@@ -5,7 +5,7 @@
 
 import torch
 import torch.nn as nn
-from torch.optim import Adam
+from torch.optim import AdamW
 from torch.autograd import Variable
 import torch.nn.functional as F
 
@@ -146,7 +146,7 @@ def __init__(self, lr, l2, num_layers, num_channels, n, action_size, train_use_g
         if self.train_use_gpu:
             self.neural_network.cuda()
 
-        self.optim = Adam(self.neural_network.parameters(), lr=self.lr, weight_decay=self.l2)
+        self.optim = AdamW(self.neural_network.parameters(), lr=self.lr, weight_decay=self.l2)
         self.alpha_loss = AlphaLoss()
 
     def train(self, example_buffer, batch_size, epochs):