def get_default_hyperparams(**overrides):
    """Returns the hyperparams for the neural net.

    In other words, returns a dict whose parameters come from the AGZ
    paper:
      k: number of filters (AlphaGoZero used 256). We use 128 by
        default for a 19x19 go board.
      fc_width: Dimensionality of the fully connected linear layer
      num_shared_layers: number of shared residual blocks.  AGZ used both 19
        and 39. Here we use 19 because it's faster to train.
      l2_strength: The L2 regularization parameter.
      momentum: The momentum parameter for training
    """
    k = _round_power_of_two(go.N ** 2 / 3)  # width of each layer
    hparams = {
        'k': k,  # Width of each conv layer
        'fc_width': 2 * k,  # Width of each fully connected layer
        'num_shared_layers': go.N,  # Number of shared trunk layers
        'l2_strength': 1e-4,  # Regularization strength
        'momentum': 0.9,  # Momentum used in SGD
    }
    hparams.update(**overrides)
    return hparams 

def incorporate_results(self, move_probabilities, value, up_to):
        assert move_probabilities.shape == (go.N * go.N + 1,)
        # A finished game should not be going through this code path - should
        # directly call backup_value() on the result of the game.
        assert not self.position.is_game_over()
        if self.is_expanded:
            self.revert_visits(up_to=up_to)
            return
        self.is_expanded = True
        self.original_prior = self.child_prior = move_probabilities
        # initialize child Q as current node's value, to prevent dynamics where
        # if B is winning, then B will only ever explore 1 move, because the Q
        # estimation will be so much larger than the 0 of the other moves.
        #
        # Conversely, if W is winning, then B will explore all 362 moves before
        # continuing to explore the most favorable move. This is a waste of search.
        #
        # The value seeded here acts as a prior, and gets averaged into Q calculations.
        self.child_W = np.ones([go.N * go.N + 1], dtype=np.float32) * value
        self.backup_value(value, up_to=up_to) 

def stone_features(position):
    # a bit easier to calculate it with axis 0 being the 16 board states,
    # and then roll axis 0 to the end.
    features = np.zeros([16, go.N, go.N], dtype=np.uint8)

    num_deltas_avail = position.board_deltas.shape[0]
    cumulative_deltas = np.cumsum(position.board_deltas, axis=0)
    last_eight = np.tile(position.board, [8, 1, 1])
    # apply deltas to compute previous board states
    last_eight[1:num_deltas_avail + 1] -= cumulative_deltas
    # if no more deltas are available, just repeat oldest board.
    last_eight[num_deltas_avail +
               1:] = last_eight[num_deltas_avail].reshape(1, go.N, go.N)

    features[::2] = last_eight == position.to_play
    features[1::2] = last_eight == -position.to_play
    return np.rollaxis(features, 0, 3) 

def make_tf_example(features, pi, value):
    '''
    Args:
        features: [N, N, FEATURE_DIM] nparray of uint8
        pi: [N * N + 1] nparray of float32
        value: float
    '''
    return tf.train.Example(features=tf.train.Features(feature={
        'x': tf.train.Feature(
            bytes_list=tf.train.BytesList(
                value=[features.tostring()])),
        'pi': tf.train.Feature(
            bytes_list=tf.train.BytesList(
                value=[pi.tostring()])),
        'outcome': tf.train.Feature(
            float_list=tf.train.FloatList(
                value=[value]))}))

# Write tf.Example to files 

def test_pick_moves(self):
        player = initialize_basic_player()
        root = player.root
        root.child_N[coords.to_flat((2, 0))] = 10
        root.child_N[coords.to_flat((1, 0))] = 5
        root.child_N[coords.to_flat((3, 0))] = 1

        root.position.n = go.N ** 2  # move 81, or 361, or... Endgame.

        # Assert we're picking deterministically
        self.assertTrue(root.position.n > player.temp_threshold)
        move = player.pick_move()
        self.assertEqual(move, (2, 0))

        # But if we're in the early part of the game, pick randomly
        root.position.n = 3
        self.assertFalse(player.root.position.n > player.temp_threshold)

        with mock.patch('random.random', lambda: .5):
            move = player.pick_move()
            self.assertEqual(move, (2, 0))

        with mock.patch('random.random', lambda: .99):
            move = player.pick_move()
            self.assertEqual(move, (3, 0)) 

def test_dont_pass_if_losing(self):
        player = initialize_almost_done_player()

        # check -- white is losing.
        self.assertEqual(player.root.position.score(), -0.5)

        for i in range(20):
            player.tree_search()
        # uncomment to debug this test
        # print(player.root.describe())

        # Search should converge on D9 as only winning move.
        flattened = coords.to_flat(coords.from_kgs('D9'))
        best_move = np.argmax(player.root.child_N)
        self.assertEqual(best_move, flattened)
        # D9 should have a positive value
        self.assertGreater(player.root.children[flattened].Q, 0)
        self.assertGreaterEqual(player.root.N, 20)
        # passing should be ineffective.
        self.assertLess(player.root.child_Q[-1], 0)
        # no virtual losses should be pending
        self.assertNoPendingVirtualLosses(player.root)
        # uncomment to debug this test
        # print(player.root.describe()) 

def test_only_check_game_end_once(self):
        # When presented with a situation where the last move was a pass,
        # and we have to decide whether to pass, it should be the first thing
        # we check, but not more than that.

        white_passed_pos = go.Position(
        ).play_move((3, 3)  # b plays
                    ).play_move((3, 4)  # w plays
                                ).play_move((4, 3)  # b plays
                                            ).pass_move()  # w passes - if B passes too, B would lose by komi.

        player = MCTSPlayerMixin(DummyNet())
        player.initialize_game(white_passed_pos)
        # initialize the root
        player.tree_search()
        # explore a child - should be a pass move.
        player.tree_search()
        pass_move = go.N * go.N
        self.assertEqual(player.root.children[pass_move].N, 1)
        self.assertEqual(player.root.child_N[pass_move], 1)
        player.tree_search()
        # check that we didn't visit the pass node any more times.
        self.assertEqual(player.root.child_N[pass_move], 1) 

def test_never_select_illegal_moves(self):
        probs = np.array([0.02] * (go.N * go.N + 1))
        # let's say the NN were to accidentally put a high weight on an illegal move
        probs[1] = 0.99
        root = MCTSNode(SEND_TWO_RETURN_ONE)
        root.incorporate_results(probs, 0, root)
        # and let's say the root were visited a lot of times, which pumps up the
        # action score for unvisited moves...
        root.N = 100000
        root.child_N[root.position.all_legal_moves()] = 10000
        # this should not throw an error...
        leaf = root.select_leaf()
        # the returned leaf should not be the illegal move
        self.assertNotEqual(leaf.fmove, 1)

        # and even after injecting noise, we should still not select an illegal move
        for i in range(10):
            root.inject_noise()
            leaf = root.select_leaf()
            self.assertNotEqual(leaf.fmove, 1) 

def load_board(string):
    reverse_map = {
        'X': go.BLACK,
        'O': go.WHITE,
        '.': go.EMPTY,
        '#': go.FILL,
        '*': go.KO,
        '?': go.UNKNOWN
    }

    string = re.sub(r'[^XO\.#]+', '', string)
    assert len(string) == go.N ** 2, "Board to load didn't have right dimensions"
    board = np.zeros([go.N, go.N], dtype=np.int8)
    for i, char in enumerate(string):
        np.ravel(board)[i] = reverse_map[char]
    return board 

def make_sgf(
    move_history,
    result_string,
    ruleset="Chinese",
    komi=7.5,
    white_name=PROGRAM_IDENTIFIER,
    black_name=PROGRAM_IDENTIFIER,
    comments=[]
):
    '''Turn a game into SGF.

    Doesn't handle handicap games or positions with incomplete history.

    Args:
        move_history: iterable of PlayerMoves
        result_string: "B+R", "W+0.5", etc.
        comments: iterable of string/None. Will be zipped with move_history.
    '''
    boardsize = go.N
    game_moves = ''.join(translate_sgf_move(*z)
                         for z in itertools.zip_longest(move_history, comments))
    result = result_string
    return SGF_TEMPLATE.format(**locals()) 

def get_default_hyperparams(**overrides):
    """Returns the hyperparams for the neural net.

    In other words, returns a dict whose parameters come from the AGZ
    paper:
      k: number of filters (AlphaGoZero used 256). We use 128 by
        default for a 19x19 go board.
      fc_width: Dimensionality of the fully connected linear layer
      num_shared_layers: number of shared residual blocks.  AGZ used both 19
        and 39. Here we use 19 because it's faster to train.
      l2_strength: The L2 regularization parameter.
      momentum: The momentum parameter for training
    """
    k = _round_power_of_two(go.N ** 2 / 3)  # width of each layer
    hparams = {
        'k': k,  # Width of each conv layer
        'fc_width': 2 * k,  # Width of each fully connected layer
        'num_shared_layers': go.N,  # Number of shared trunk layers
        'l2_strength': 1e-4,  # Regularization strength
        'momentum': 0.9,  # Momentum used in SGD
    }
    hparams.update(**overrides)
    return hparams 

def __init__(self, position, fmove=None, parent=None):
        if parent is None:
            parent = DummyNode()
        self.parent = parent
        self.fmove = fmove  # move that led to this position, as flattened coords
        self.position = position
        self.is_expanded = False
        self.losses_applied = 0  # number of virtual losses on this node
        # using child_() allows vectorized computation of action score.
        self.illegal_moves = 1000 * (1 - self.position.all_legal_moves())
        self.child_N = np.zeros([go.N * go.N + 1], dtype=np.float32)
        self.child_W = np.zeros([go.N * go.N + 1], dtype=np.float32)
        # save a copy of the original prior before it gets mutated by d-noise.
        self.original_prior = np.zeros([go.N * go.N + 1], dtype=np.float32)
        self.child_prior = np.zeros([go.N * go.N + 1], dtype=np.float32)
        self.children = {}  # map of flattened moves to resulting MCTSNode 

def select_leaf(self):
        current = self
        pass_move = go.N * go.N
        while True:
            current.N += 1
            # if a node has never been evaluated, we have no basis to select a child.
            if not current.is_expanded:
                break
            # HACK: if last move was a pass, always investigate double-pass first
            # to avoid situations where we auto-lose by passing too early.
            if (current.position.recent
                and current.position.recent[-1].move is None
                    and current.child_N[pass_move] == 0):
                current = current.maybe_add_child(pass_move)
                continue

            best_move = np.argmax(current.child_action_score)
            current = current.maybe_add_child(best_move)
        return current 

def incorporate_results(self, move_probabilities, value, up_to):
        assert move_probabilities.shape == (go.N * go.N + 1,)
        # A finished game should not be going through this code path - should
        # directly call backup_value() on the result of the game.
        assert not self.position.is_game_over()
        if self.is_expanded:
            self.revert_visits(up_to=up_to)
            return
        self.is_expanded = True
        self.original_prior = self.child_prior = move_probabilities
        # initialize child Q as current node's value, to prevent dynamics where
        # if B is winning, then B will only ever explore 1 move, because the Q
        # estimation will be so much larger than the 0 of the other moves.
        #
        # Conversely, if W is winning, then B will explore all 362 moves before
        # continuing to explore the most favorable move. This is a waste of search.
        #
        # The value seeded here acts as a prior, and gets averaged into Q calculations.
        self.child_W = np.ones([go.N * go.N + 1], dtype=np.float32) * value
        self.backup_value(value, up_to=up_to) 

def stone_features(position):
    # a bit easier to calculate it with axis 0 being the 16 board states,
    # and then roll axis 0 to the end.
    features = np.zeros([16, go.N, go.N], dtype=np.uint8)

    num_deltas_avail = position.board_deltas.shape[0]
    cumulative_deltas = np.cumsum(position.board_deltas, axis=0)
    last_eight = np.tile(position.board, [8, 1, 1])
    # apply deltas to compute previous board states
    last_eight[1:num_deltas_avail + 1] -= cumulative_deltas
    # if no more deltas are available, just repeat oldest board.
    last_eight[num_deltas_avail +
               1:] = last_eight[num_deltas_avail].reshape(1, go.N, go.N)

    features[::2] = last_eight == position.to_play
    features[1::2] = last_eight == -position.to_play
    return np.rollaxis(features, 0, 3) 

def make_tf_example(features, pi, value):
    '''
    Args:
        features: [N, N, FEATURE_DIM] nparray of uint8
        pi: [N * N + 1] nparray of float32
        value: float
    '''
    return tf.train.Example(features=tf.train.Features(feature={
        'x': tf.train.Feature(
            bytes_list=tf.train.BytesList(
                value=[features.tostring()])),
        'pi': tf.train.Feature(
            bytes_list=tf.train.BytesList(
                value=[pi.tostring()])),
        'outcome': tf.train.Feature(
            float_list=tf.train.FloatList(
                value=[value]))}))

# Write tf.Example to files 

def batch_parse_tf_example(batch_size, example_batch):
    '''
    Args:
        example_batch: a batch of tf.Example
    Returns:
        A tuple (feature_tensor, dict of output tensors)
    '''
    features = {
        'x': tf.FixedLenFeature([], tf.string),
        'pi': tf.FixedLenFeature([], tf.string),
        'outcome': tf.FixedLenFeature([], tf.float32),
    }
    parsed = tf.parse_example(example_batch, features)
    x = tf.decode_raw(parsed['x'], tf.uint8)
    x = tf.cast(x, tf.float32)
    x = tf.reshape(x, [batch_size, go.N, go.N,
                       features_lib.NEW_FEATURES_PLANES])
    pi = tf.decode_raw(parsed['pi'], tf.float32)
    pi = tf.reshape(pi, [batch_size, go.N * go.N + 1])
    outcome = parsed['outcome']
    outcome.set_shape([batch_size])
    return (x, {'pi_tensor': pi, 'value_tensor': outcome}) 

def generate(model_num):
    if model_num == 0:
        new_name = 'bootstrap'
    elif go.N == 19:
        new_name = random.choice(NAMES)
    else:
        new_name = petname.generate()
    full_name = "%06d-%s" % (model_num, new_name)
    return full_name 

def get_inference_input():
    """Set up placeholders for input features/labels.

    Returns the feature, output tensors that get passed into model_fn."""
    return (tf.placeholder(tf.float32,
                           [None, go.N, go.N, features.NEW_FEATURES_PLANES],
                           name='pos_tensor'),
            {'pi_tensor': tf.placeholder(tf.float32, [None, go.N * go.N + 1]),
             'value_tensor': tf.placeholder(tf.float32, [None])}) 

def apply_symmetry_pi(s, pi):
    pi = np.copy(pi)
    # rotate all moves except for the pass move at end
    pi[:-1] = IMPLS[s](pi[:-1].reshape([go.N, go.N])).ravel()
    return pi 

def D_NOISE_ALPHA(): return 0.03 * 361 / (go.N ** 2) 

def __repr__(self):
        return "<MCTSNode move=%s, N=%s, to_play=%s>" % (
            self.position.recent[-1:], self.N, self.position.to_play) 

def child_U(self):
        return (c_PUCT * math.sqrt(1 + self.N) *
                self.child_prior / (1 + self.child_N)) 

def Q(self):
        return self.W / (1 + self.N) 

def N(self):
        return self.parent.child_N[self.fmove] 

def N(self, value):
        self.parent.child_N[self.fmove] = value 

def revert_visits(self, up_to):
        """Revert visit increments.

        Sometimes, repeated calls to select_leaf return the same node.
        This is rare and we're okay with the wasted computation to evaluate
        the position multiple times by the dual_net. But select_leaf has the
        side effect of incrementing visit counts. Since we want the value to
        only count once for the repeatedly selected node, we also have to
        revert the incremented visit counts.
        """
        self.N -= 1
        if self.parent is None or self is up_to:
            return
        self.parent.revert_visits(up_to) 

def inject_noise(self):
        dirch = np.random.dirichlet([D_NOISE_ALPHA()] * ((go.N * go.N) + 1))
        self.child_prior = self.child_prior * 0.75 + dirch * 0.25 

def most_visited_path(self):
        node = self
        output = []
        while node.children:
            next_kid = np.argmax(node.child_N)
            node = node.children.get(next_kid)
            if node is None:
                output.append("GAME END")
                break
            output.append("%s (%d) ==> " % (coords.to_kgs(
                                            coords.from_flat(node.fmove)),
                                            node.N))
        output.append("Q: {:.5f}\n".format(node.Q))
        return ''.join(output) 

def describe(self):
        sort_order = list(range(go.N * go.N + 1))
        sort_order.sort(key=lambda i: (
            self.child_N[i], self.child_action_score[i]), reverse=True)
        soft_n = self.child_N / sum(self.child_N)
        p_delta = soft_n - self.child_prior
        p_rel = p_delta / self.child_prior
        # Dump out some statistics
        output = []
        output.append("{q:.4f}\n".format(q=self.Q))
        output.append(self.most_visited_path())
        output.append(
            "move:  action      Q      U      P    P-Dir    N  soft-N  p-delta  p-rel\n")
        output.append("\n".join(["{!s:6}: {: .3f}, {: .3f}, {:.3f}, {:.3f}, {:.3f}, {:4d} {:.4f} {: .5f} {: .2f}".format(
            coords.to_kgs(coords.from_flat(key)),
            self.child_action_score[key],
            self.child_Q[key],
            self.child_U[key],
            self.child_prior[key],
            self.original_prior[key],
            int(self.child_N[key]),
            soft_n[key],
            p_delta[key],
            p_rel[key])
            for key in sort_order][:15]))
        return ''.join(output)