def get_agent(num_filters: int) -> AlphaZeroArgMaxAgent:
path_to_weights = os.path.join('pretrained', f'ttt_4blocks_{num_filters}filters.pth')
config['num_resnet_filters'] = num_filters
config['value_head_hidden_dim'] = num_filters
net = dual_resnet(game, config)
net.load_state_dict(torch.load(path_to_weights))
return AlphaZeroArgMaxAgent(game, state_encoder, net, config)
def get_agent(model_name: str) -> AlphaZeroArgMaxAgent:
path_to_weights = os.path.join('pretrained', f'ttt_{model_name}.pth')
if model_name.startswith('dualres'):
config['num_res_blocks'] = 4
net = dual_resnet(game, config)
else:
config['num_res_blocks'] = 8
net = resnet(game, config)
net.load_state_dict(torch.load(path_to_weights))
return AlphaZeroArgMaxAgent(game, state_encoder, net, config)
def get_agent(num_blocks: int) -> AlphaZeroArgMaxAgent:
path_to_weights = os.path.join('pretrained', f'ttt_{num_blocks}blocks_256filters.pth')
config['num_res_blocks'] = num_blocks
net = dual_resnet(game, config)
net.load_state_dict(torch.load(path_to_weights))
return AlphaZeroArgMaxAgent(game, state_encoder, net, config)
config = yaml.safe_load(f)
config['device'] = 'cuda' if torch.cuda.is_available() else 'cpu'
def read_move(player: TicTacToePlayer) -> TicTacToeMove:
x, y = input(f"{player.name} move: ").split()
x, y = int(x), int(y)
return TicTacToeMove(x, y)
if __name__ == '__main__':
game = TicTacToeGame(config['game_size'])
state_encoder = TicTacToeStateEncoder(config['device'])
net = dual_resnet(game, config)
mcts = MonteCarloTreeSearch(game=game,
state_encoder=state_encoder,
nn=net,
config=config)
net.load_state_dict(
torch.load(os.path.join('pretrained', 'ttt_dualres_comp.pth')))
# net.load_state_dict(torch.load(os.path.join(config['log_dir'], 'best.pth')))
net.eval()
agent = AlphaZeroArgMaxAgent(game, state_encoder, net, config)
agent_role = random.choice([TicTacToePlayer.X, TicTacToePlayer.O])
while not game.is_over:
game.show_board()