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_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_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_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_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_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_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_coach(capsys, tmpdir):
args.checkpoint = tmpdir
game = OthelloGame(6)
nnet = OthelloNNet(game)
coach = Coach(game, nnet, args)
coach.learn()