def test_alpha_zero_player_from_model():
game = OthelloGame(6)
some_net = OthelloNNet(game)
player = AlphaZeroPlayer(game, some_net)
board = game.get_init_board()
action = player.play(board)
assert action
def test_bare_model_player_from_model():
game = OthelloGame(6)
some_net = OthelloNNet(game)
player = BareModelPlayer(game, some_net)
board = game.get_init_board()
action = player.play(board)
assert action
def test_human_player_valid_action():
action_valid = "1,2"
game = OthelloGame(6)
player = HumanPlayer(game)
board = game.get_init_board()
with patch("builtins.input", side_effect=[action_valid]):
action = player.play(board)
assert action == 8
def test_alpha_zero_player_reset():
game = OthelloGame(6)
player = AlphaZeroPlayer(game, OthelloNNet, num_mcts_sims=4)
assert not player.mcts.Qsa
board = game.get_init_board()
player.play(board)
assert player.mcts.Qsa
player.reset()
assert not player.mcts.Qsa
def test_human_player_invalid_action(capsys):
action_valid = "1,2"
action_invalid = "0,0"
game = OthelloGame(6)
player = HumanPlayer(game)
board = game.get_init_board()
with patch("builtins.input", side_effect=[action_invalid, action_valid]):
player.play(board)
out, _err = capsys.readouterr()
assert out == "Invalid action.\n"
def test_bare_model_player_from_checkpoint():
game = OthelloGame(6)
some_net = OthelloNNet(game)
folder, filename = "/tmp/", "checkpoint_bare_model_player"
some_net.save_checkpoint(folder, filename)
del some_net
player = BareModelPlayer(game,
OthelloNNet,
folder=folder,
filename=filename)
board = game.get_init_board()
action = player.play(board)
assert action
def test_alpha_zero_player_from_checkpoint():
game = OthelloGame(6)
some_net = OthelloNNet(game)
folder, filename = "/tmp/", "checkpoint_alpha_zero_player"
some_net.save_checkpoint(folder, filename)
del some_net
player = AlphaZeroPlayer(game,
OthelloNNet,
folder=folder,
filename=filename,
num_mcts_sims=4)
board = game.get_init_board()
action = player.play(board)
assert action
def test_league_lazy_loading_players():
game = OthelloGame(6)
league = League(game, rounds=1, games=2, cache_size=2)
for i in range(10):
league.add_player(f"random_{i}", lambda: RandomPlayer(game))
league.start()
final_ratings = league.ratings()
assert len(final_ratings) == 10
def test_arena():
game = OthelloGame(6)
player1 = RandomPlayer(game)
player2 = RandomPlayer(game)
arena = Arena(player1, player2, game, display=OthelloGame.display)
number_of_games = 10
one_won, two_won, draws = arena.play_games(number_of_games, verbose=True)
assert one_won + two_won + draws == number_of_games
def test_self_play(local_ray, tmpdir):
game = OthelloGame(6)
nnet = OthelloNNet(game)
s = SharedStorage.remote(nnet.get_weights())
r = ReplayBuffer.remote(games_to_play=1, games_to_use=1, folder=tmpdir)
assert ray.get(r.get_number_of_games_played.remote()) == 0
self_play = SelfPlay.remote(r, s, game, nnet.__class__, dict(args))
ray.get(self_play.start.remote())
assert ray.get(r.get_number_of_games_played.remote()) == 1
assert ray.get(r.played_enough.remote()) is True
games = ray.get(ray.get(r.get_examples.remote()))
assert len(games) == 1
examples = games[0]
assert len(examples) > 2
board, policy, winner = examples[0]
assert isinstance(board, type(game.get_init_board()))
assert len(policy) == game.get_action_size()
assert all(0 <= value <= 1 for value in policy)
assert winner in [1, -1]
def test_model_trainer_loop(local_ray, tmpdir):
game = OthelloGame(6)
nnet = OthelloNNet(game)
s = SharedStorage.remote(nnet.get_weights())
assert ray.get(s.get_revision.remote()) == 0
r = MockedReplayBuffer.remote(games_to_play=4,
games_to_use=4,
folder=tmpdir)
r.add_game_examples.remote(mock_example_data(game))
model_trainer = ModelTrainer.options(num_gpus=0).remote(
r, s, game, nnet.__class__, dict(args), selfplay_training_ratio=1)
ray.get(model_trainer.start.remote())
assert ray.get(s.get_revision.remote()) > 0
assert ray.get(s.trained_enough.remote()) is True
def test_model_trainer_pit_reject_model(capsys, local_ray, tmpdir):
game = OthelloGame(6)
nnet = OthelloNNet(game)
s = SharedStorage.remote(nnet.get_weights())
assert ray.get(s.get_revision.remote()) == 0
r = ReplayBuffer.remote(games_to_play=2, games_to_use=2, folder=tmpdir)
r.add_game_examples.remote(mock_example_data(game))
# provoke model rejection by tweaking updateThreshold to fail
custom_args = dict(args, updateThreshold=1.1)
model_trainer = ModelTrainer.options(num_gpus=0).remote(
r, s, game, nnet.__class__, custom_args, pit_against_old_model=True)
ray.get(model_trainer.train.remote())
assert ray.get(s.get_revision.remote()) == 0
out, _err = capsys.readouterr()
assert "PITTING AGAINST PREVIOUS VERSION" in out
assert "REJECTING NEW MODEL" in out
def test_league_incrementally_add_players_to_running():
game = OthelloGame(6)
rounds = 2
league = League(game, initial_rating=1500, rounds=rounds, games=4)
assert league.has_started is False
league.start()
assert league.has_started is True
league.add_player("random_1", RandomPlayer(game))
assert league.ratings() == [(1500, 1, "random_1")]
assert len(league.history) == 0
for i in range(2, 6):
league.add_player(f"random_{i}", RandomPlayer(game))
assert len(league.ratings()) == i
assert len(league.history) == sum(range(i)) * rounds
def test_league_basic_usage():
game = OthelloGame(6)
rounds = 2
league = League(game, initial_rating=1500, rounds=rounds, games=4)
league.add_player("random_1", RandomPlayer(game), initial_rating=2000)
league.add_player("random_2", RandomPlayer(game), initial_rating=1750)
league.add_player("random_3", RandomPlayer(game))
league.add_player("random_4", RandomPlayer(game), initial_rating=1000)
assert league.has_started is False
initial_ratings = league.ratings()
assert initial_ratings == [
(2000, 1, "random_1"),
(1750, 2, "random_2"),
(1500, 3, "random_3"),
(1000, 4, "random_4"),
]
league.start()
assert league.has_started is True
assert len(league.history) == sum(range(4)) * rounds
final_ratings = league.ratings()
assert len(final_ratings) == 4
assert final_ratings != initial_ratings
def test_bare_model_player_invalid_nnet_parameter():
neither_nnet_nor_nnet_class = "something_else"
game = OthelloGame(6)
with pytest.raises(TypeError) as excinfo:
BareModelPlayer(game, neither_nnet_nor_nnet_class)
assert "NeuralNet subclass or instance" in str(excinfo.value)
def test_coach(capsys, tmpdir):
args.checkpoint = tmpdir
game = OthelloGame(6)
nnet = OthelloNNet(game)
coach = Coach(game, nnet, args)
coach.learn()
def test_bare_model_player():
game = OthelloGame(6)
player = BareModelPlayer(game, OthelloNNet)
board = game.get_init_board()
action = player.play(board)
assert action
def test_greedy_player():
game = OthelloGame(6)
player = GreedyPlayer(game)
board = game.get_init_board()
action = player.play(board)
assert action
def test_random_player():
game = OthelloGame(6)
player = RandomPlayer(game)
board = game.get_init_board()
action = player.play(board)
assert action
def test_alpha_zero_player():
game = OthelloGame(6)
player = AlphaZeroPlayer(game, OthelloNNet, num_mcts_sims=4)
board = game.get_init_board()
action = player.play(board)
assert action
