コード例 #1
0
def test_board_state():
    """ Tests the proto_to_state_space  """
    game = Game()
    game_map = game.map
    state_proto = state_space.extract_state_proto(game)
    new_game = state_space.build_game_from_state_proto(state_proto)

    # Retrieving board_state
    state_proto_2 = state_space.extract_state_proto(new_game)
    board_state_1 = state_space.proto_to_board_state(state_proto, game_map)
    board_state_2 = state_space.proto_to_board_state(state_proto_2, game_map)

    # Checking
    assert np.allclose(board_state_1, board_state_2)
    assert board_state_1.shape == (state_space.NB_NODES, state_space.NB_FEATURES)
    assert game.get_hash() == new_game.get_hash()
コード例 #2
0
ファイル: base.py プロジェクト: zhanpengfang/research
    def get_feedable_item(locs, state_proto, power_name, phase_history_proto, possible_orders_proto, **kwargs):
        """ Computes and return a feedable item (to be fed into the feedable queue)
            :param locs: A list of locations for which we want orders
            :param state_proto: A `.proto.game.State` representation of the state of the game.
            :param power_name: The power name for which we want the orders and the state values
            :param phase_history_proto: A list of `.proto.game.PhaseHistory`. This represents prev phases.
            :param possible_orders_proto: A `proto.game.PossibleOrders` object representing possible order for each loc.
            :param kwargs: Additional optional kwargs:
                - player_seed: The seed to apply to the player to compute a deterministic mask.
                - noise: The sigma of the additional noise to apply to the intermediate layers (i.e. sigma * epsilon)
                - temperature: The temperature to apply to the logits. (Default to 0. for deterministic/greedy)
                - dropout_rate: The amount of dropout to apply to the inputs/outputs of the decoder.
            :return: A feedable item, with feature names as key and numpy arrays as values
        """
        # pylint: disable=too-many-branches
        # Converting to state space
        map_object = Map(state_proto.map)
        board_state = proto_to_board_state(state_proto, map_object)

        # Building the decoder length
        # For adjustment phase, we restrict the number of builds/disbands to what is allowed by the game engine
        in_adjustment_phase = state_proto.name[-1] == 'A'
        nb_builds = state_proto.builds[power_name].count
        nb_homes = len(state_proto.builds[power_name].homes)

        # If we are in adjustment phase, making sure the locs are the orderable locs (and not the policy locs)
        if in_adjustment_phase:
            orderable_locs, _ = get_orderable_locs_for_powers(state_proto, [power_name])
            if sorted(locs) != sorted(orderable_locs):
                if locs:
                    LOGGER.warning('Adj. phase requires orderable locs. Got %s. Expected %s.', locs, orderable_locs)
                locs = orderable_locs

        # WxxxA - We can build units
        # WxxxA - We can disband units
        # Other phase
        if in_adjustment_phase and nb_builds >= 0:
            decoder_length = min(nb_builds, nb_homes)
        elif in_adjustment_phase and nb_builds < 0:
            decoder_length = abs(nb_builds)
        else:
            decoder_length = len(locs)

        # Computing the candidates for the policy
        if possible_orders_proto:

            # Adjustment Phase - Use all possible orders for each location.
            if in_adjustment_phase:

                # Building a list of all orders for all locations
                adj_orders = []
                for loc in locs:
                    adj_orders += possible_orders_proto[loc].value

                # Computing the candidates
                candidates = [get_order_based_mask(adj_orders)] * decoder_length

            # Regular phase - Compute candidates for each location
            else:
                candidates = []
                for loc in locs:
                    candidates += [get_order_based_mask(possible_orders_proto[loc].value)]

        # We don't have possible orders, so we cannot compute candidates
        # This might be normal if we are only getting the state value or the next message to send
        else:
            candidates = []
            for _ in range(decoder_length):
                candidates.append([])

        # Prev orders state
        prev_orders_state = []
        for phase_proto in reversed(phase_history_proto):
            if len(prev_orders_state) == NB_PREV_ORDERS:
                break
            if phase_proto.name[-1] == 'M':
                prev_orders_state = [proto_to_prev_orders_state(phase_proto, map_object)] + prev_orders_state
        for _ in range(NB_PREV_ORDERS - len(prev_orders_state)):
            prev_orders_state = [np.zeros((NB_NODES, NB_ORDERS_FEATURES), dtype=np.uint8)] + prev_orders_state
        prev_orders_state = np.array(prev_orders_state)

        # Building (order) decoder inputs [GO_ID]
        decoder_inputs = [GO_ID]

        # kwargs
        player_seed = kwargs.get('player_seed', 0)
        noise = kwargs.get('noise', 0.)
        temperature = kwargs.get('temperature', 0.)
        dropout_rate = kwargs.get('dropout_rate', 0.)

        # Building feedable data
        item = {
            'player_seed': player_seed,
            'board_state': board_state,
            'board_alignments': get_board_alignments(locs,
                                                     in_adjustment_phase=in_adjustment_phase,
                                                     tokens_per_loc=1,
                                                     decoder_length=decoder_length),
            'prev_orders_state': prev_orders_state,
            'decoder_inputs': decoder_inputs,
            'decoder_lengths': decoder_length,
            'candidates': candidates,
            'noise': noise,
            'temperature': temperature,
            'dropout_rate': dropout_rate,
            'current_power': POWER_VOCABULARY_KEY_TO_IX[power_name],
            'current_season': get_current_season(state_proto)
        }

        # Return
        return item
コード例 #3
0
ファイル: base.py プロジェクト: zhanpengfang/research
def get_policy_data(saved_game_proto, power_names, top_victors):
    """ Computes the proto to save in tf.train.Example as a training example for the policy network
        :param saved_game_proto: A `.proto.game.SavedGame` object from the dataset.
        :param power_names: The list of powers for which we want the policy data
        :param top_victors: The list of powers that ended with more than 25% of the supply centers
        :return: A dictionary with key: the phase_ix
                              with value: A dict with the power_name as key and a dict with the example fields as value
    """
    nb_phases = len(saved_game_proto.phases)
    policy_data = {phase_ix: {} for phase_ix in range(nb_phases - 1)}
    game_id = saved_game_proto.id
    map_object = Map(saved_game_proto.map)

    # Determining if we have a draw
    nb_sc_to_win = len(map_object.scs) // 2 + 1
    has_solo_winner = max([len(saved_game_proto.phases[-1].state.centers[power_name].value)
                           for power_name in saved_game_proto.phases[-1].state.centers]) >= nb_sc_to_win
    survivors = [power_name for power_name in saved_game_proto.phases[-1].state.centers
                 if saved_game_proto.phases[-1].state.centers[power_name].value]
    has_draw = not has_solo_winner and len(survivors) >= 2

    # Processing all phases (except the last one)
    current_year = 0
    for phase_ix in range(nb_phases - 1):

        # Building a list of orders of previous phases
        previous_orders_states = [np.zeros((NB_NODES, NB_ORDERS_FEATURES), dtype=np.uint8)] * NB_PREV_ORDERS
        for phase_proto in saved_game_proto.phases[max(0, phase_ix - NB_PREV_ORDERS_HISTORY):phase_ix]:
            if phase_proto.name[-1] == 'M':
                previous_orders_states += [proto_to_prev_orders_state(phase_proto, map_object)]
        previous_orders_states = previous_orders_states[-NB_PREV_ORDERS:]
        prev_orders_state = np.array(previous_orders_states)

        # Parsing each requested power in the specified phase
        phase_proto = saved_game_proto.phases[phase_ix]
        phase_name = phase_proto.name
        state_proto = phase_proto.state
        phase_board_state = proto_to_board_state(state_proto, map_object)

        # Increasing year for every spring or when the game is completed
        if phase_proto.name == 'COMPLETED' or (phase_proto.name[0] == 'S' and phase_proto.name[-1] == 'M'):
            current_year += 1

        for power_name in power_names:
            phase_issued_orders = get_issued_orders_for_powers(phase_proto, [power_name])
            phase_possible_orders = get_possible_orders_for_powers(phase_proto, [power_name])
            phase_draw_target = 1. if has_draw and phase_ix == (nb_phases - 2) and power_name in survivors else 0.

            # Data to use when not learning a policy
            blank_policy_data = {'board_state': phase_board_state,
                                 'prev_orders_state': prev_orders_state,
                                 'draw_target': phase_draw_target}

            # Power is not a top victor - We don't want to learn a policy from him
            if power_name not in top_victors:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Finding the orderable locs
            orderable_locations = list(phase_issued_orders[power_name].keys())

            # Skipping power for this phase if we are only issuing Hold
            for order_loc, order in phase_issued_orders[power_name].items():
                order_tokens = get_order_tokens(order)
                if len(order_tokens) >= 2 and order_tokens[1] != 'H':
                    break
            else:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Removing orderable locs where orders are not possible (i.e. NO_CHECK games)
            for order_loc, order in phase_issued_orders[power_name].items():
                if order not in phase_possible_orders[order_loc] and order_loc in orderable_locations:
                    if 'NO_CHECK' not in saved_game_proto.rules:
                        LOGGER.warning('%s not in all possible orders. Phase %s - Game %s.', order, phase_name, game_id)
                    orderable_locations.remove(order_loc)

                # Remove orderable locs where the order is either invalid or not frequent
                if order_to_ix(order) is None and order_loc in orderable_locations:
                    orderable_locations.remove(order_loc)

            # Determining if we are in an adjustment phase
            in_adjustment_phase = state_proto.name[-1] == 'A'
            nb_builds = state_proto.builds[power_name].count
            nb_homes = len(state_proto.builds[power_name].homes)

            # WxxxA - We can build units
            # WxxxA - We can disband units
            # Other phase
            if in_adjustment_phase and nb_builds >= 0:
                decoder_length = min(nb_builds, nb_homes)
            elif in_adjustment_phase and nb_builds < 0:
                decoder_length = abs(nb_builds)
            else:
                decoder_length = len(orderable_locations)

            # Not all units were disbanded - Skipping this power as we can't learn the orders properly
            if in_adjustment_phase and nb_builds < 0 and len(orderable_locations) < abs(nb_builds):
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Not enough orderable locations for this power, skipping
            if not orderable_locations or not decoder_length:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # decoder_inputs [GO, order1, order2, order3]
            decoder_inputs = [GO_ID]
            decoder_inputs += [order_to_ix(phase_issued_orders[power_name][loc]) for loc in orderable_locations]
            if in_adjustment_phase and nb_builds > 0:
                decoder_inputs += [order_to_ix('WAIVE')] * (min(nb_builds, nb_homes) - len(orderable_locations))
            decoder_length = min(decoder_length, NB_SUPPLY_CENTERS)

            # Adjustment Phase - Use all possible orders for each location.
            if in_adjustment_phase:
                build_disband_locs = list(get_possible_orders_for_powers(phase_proto, [power_name]).keys())
                phase_board_alignments = get_board_alignments(build_disband_locs,
                                                              in_adjustment_phase=in_adjustment_phase,
                                                              tokens_per_loc=1,
                                                              decoder_length=decoder_length)

                # Building a list of all orders for all locations
                adj_orders = []
                for loc in build_disband_locs:
                    adj_orders += phase_possible_orders[loc]

                # Not learning builds for BUILD_ANY
                if nb_builds > 0 and 'BUILD_ANY' in state_proto.rules:
                    adj_orders = []

                # No orders found - Skipping
                if not adj_orders:
                    policy_data[phase_ix][power_name] = blank_policy_data
                    continue

                # Computing the candidates
                candidates = [get_order_based_mask(adj_orders)] * decoder_length

            # Regular phase - Compute candidates for each location
            else:
                phase_board_alignments = get_board_alignments(orderable_locations,
                                                              in_adjustment_phase=in_adjustment_phase,
                                                              tokens_per_loc=1,
                                                              decoder_length=decoder_length)
                candidates = []
                for loc in orderable_locations:
                    candidates += [get_order_based_mask(phase_possible_orders[loc])]

            # Saving results
            # No need to return temperature, current_power, current_season
            policy_data[phase_ix][power_name] = {'board_state': phase_board_state,
                                                 'board_alignments': phase_board_alignments,
                                                 'prev_orders_state': prev_orders_state,
                                                 'decoder_inputs': decoder_inputs,
                                                 'decoder_lengths': decoder_length,
                                                 'candidates': candidates,
                                                 'draw_target': phase_draw_target}
    # Returning
    return policy_data
コード例 #4
0
def generate_trajectory(players,
                        reward_fn,
                        advantage_fn,
                        env_constructor=None,
                        hparams=None,
                        power_assignments=None,
                        set_player_seed=None,
                        initial_state_bytes=None,
                        update_interval=0,
                        update_queue=None,
                        output_format='proto'):
    """ Generates a single trajectory (Saved Gamed Proto) for RL (self-play) with the power assigments
        :param players: A list of instantiated players
        :param reward_fn: The reward function to use to calculate rewards
        :param advantage_fn: An instance of `.models.self_play.advantages`
        :param env_constructor: A callable to get the OpenAI gym environment (args: players)
        :param hparams: A dictionary of hyper parameters with their values
        :param power_assignments: Optional. The power name we want to play as. (e.g. 'FRANCE') or a list of powers.
        :param set_player_seed: Boolean that indicates that we want to set the player seed on reset().
        :param initial_state_bytes: A `game.State` proto (in bytes format) representing the initial state of the game.
        :param update_interval: Optional. If set, a partial saved game is put in the update_queue this every seconds.
        :param update_queue: Optional. If update interval is set, partial games will be put in this queue
        :param output_format: The output format. One of 'proto', 'bytes', 'zlib'
        :return: A SavedGameProto representing the game played (with policy details and power assignments)
                 Depending on format, the output might be converted to a byte array, or a compressed byte array.
        :type players: List[diplomacy_research.players.player.Player]
        :type reward_fn: diplomacy_research.models.self_play.reward_functions.AbstractRewardFunction
        :type advantage_fn: diplomacy_research.models.self_play.advantages.base_advantage.BaseAdvantage
        :type update_queue: multiprocessing.Queue
    """
    # pylint: disable=too-many-arguments
    assert output_format in ['proto', 'bytes', 'zlib'
                             ], 'Format should be "proto", "bytes", "zlib"'
    assert len(players) == NB_POWERS

    # Making sure we use the SavedGame wrapper to record the game
    if env_constructor:
        env = env_constructor(players)
    else:
        env = default_env_constructor(players, hparams, power_assignments,
                                      set_player_seed, initial_state_bytes)
    wrapped_env = env
    while not isinstance(wrapped_env, DiplomacyEnv):
        if isinstance(wrapped_env, SaveGame):
            break
        wrapped_env = wrapped_env.env
    else:
        env = SaveGame(env)

    # Detecting if we have a Auto-Draw wrapper
    has_auto_draw = False
    wrapped_env = env
    while not isinstance(wrapped_env, DiplomacyEnv):
        if isinstance(wrapped_env, AutoDraw):
            has_auto_draw = True
            break
        wrapped_env = wrapped_env.env

    # Resetting env
    env.reset()

    # Timing vars for partial updates
    time_last_update = time.time()
    year_last_update = 0
    start_phase_ix = 0
    current_phase_ix = 0
    nb_transitions = 0

    # Cache Variables
    powers = sorted(
        [power_name for power_name in get_map_powers(env.game.map)])
    assigned_powers = env.get_all_powers_name()
    stored_board_state = OrderedDict()  # {phase_name: board_state}
    stored_prev_orders_state = OrderedDict()  # {phase_name: prev_orders_state}
    stored_possible_orders = OrderedDict()  # {phase_name: possible_orders}

    power_variables = {
        power_name: {
            'orders': [],
            'policy_details': [],
            'state_values': [],
            'rewards': [],
            'returns': [],
            'last_state_value': 0.
        }
        for power_name in powers
    }

    new_state_proto = None
    phase_history_proto = []
    map_object = Map(name=env.game.map.name)

    # Generating
    while not env.is_done:
        state_proto = new_state_proto if new_state_proto is not None else extract_state_proto(
            env.game)
        possible_orders_proto = extract_possible_orders_proto(env.game)

        # Computing board_state
        board_state = proto_to_board_state(state_proto,
                                           map_object).flatten().tolist()
        state_proto.board_state.extend(board_state)

        # Storing possible orders for this phase
        current_phase = env.game.get_current_phase()
        stored_board_state[current_phase] = board_state
        stored_possible_orders[current_phase] = possible_orders_proto

        # Getting orders, policy details, and state value
        tasks = [(player, state_proto, pow_name,
                  phase_history_proto[-NB_PREV_ORDERS_HISTORY:],
                  possible_orders_proto)
                 for player, pow_name in zip(env.players, assigned_powers)]
        step_args = yield [get_step_args(*args) for args in tasks]

        # Stepping through env, storing power variables
        for power_name, (orders, policy_details,
                         state_value) in zip(assigned_powers, step_args):
            if orders:
                env.step((power_name, orders))
                nb_transitions += 1
            if has_auto_draw and policy_details is not None:
                env.set_draw_prob(power_name, policy_details['draw_prob'])

        # Processing
        env.process()
        current_phase_ix += 1

        # Retrieving draw action and saving power variables
        for power_name, (orders, policy_details,
                         state_value) in zip(assigned_powers, step_args):
            if has_auto_draw and policy_details is not None:
                policy_details['draw_action'] = env.get_draw_actions(
                )[power_name]
            power_variables[power_name]['orders'] += [orders]
            power_variables[power_name]['policy_details'] += [policy_details]
            power_variables[power_name]['state_values'] += [state_value]

        # Getting new state
        new_state_proto = extract_state_proto(env.game)

        # Storing reward for this transition
        done_reason = DoneReason(env.done_reason) if env.done_reason else None
        for power_name in powers:
            power_variables[power_name]['rewards'] += [
                reward_fn.get_reward(prev_state_proto=state_proto,
                                     state_proto=new_state_proto,
                                     power_name=power_name,
                                     is_terminal_state=done_reason is not None,
                                     done_reason=done_reason)
            ]

        # Computing prev_orders_state for the previous state
        last_phase_proto = extract_phase_history_proto(
            env.game, nb_previous_phases=1)[-1]
        if last_phase_proto.name[-1] == 'M':
            prev_orders_state = proto_to_prev_orders_state(
                last_phase_proto, map_object).flatten().tolist()
            stored_prev_orders_state[last_phase_proto.name] = prev_orders_state
            last_phase_proto.prev_orders_state.extend(prev_orders_state)
            phase_history_proto += [last_phase_proto]

        # Sending partial game if:
        # 1) We have update_interval > 0 with an update queue, and
        # 2a) The game is completed, or 2b) the update time has elapsted and at least 5 years as passed
        has_update_interval = update_interval > 0 and update_queue is not None
        game_is_completed = env.is_done
        min_time_has_passed = time.time() - time_last_update > update_interval
        current_year = 9999 if env.game.get_current_phase(
        ) == 'COMPLETED' else int(env.game.get_current_phase()[1:5])
        min_years_have_passed = current_year - year_last_update >= 5

        if (has_update_interval
                and (game_is_completed or
                     (min_time_has_passed and min_years_have_passed))):

            # Game is completed - last state value is 0
            if game_is_completed:
                for power_name in powers:
                    power_variables[power_name]['last_state_value'] = 0.

            # Otherwise - Querying the model for the value of the last state
            else:
                tasks = [
                    (player, new_state_proto, pow_name,
                     phase_history_proto[-NB_PREV_ORDERS_HISTORY:],
                     possible_orders_proto)
                    for player, pow_name in zip(env.players, assigned_powers)
                ]
                last_state_values = yield [
                    get_state_value(*args) for args in tasks
                ]

                for power_name, last_state_value in zip(
                        assigned_powers, last_state_values):
                    power_variables[power_name][
                        'last_state_value'] = last_state_value

            # Getting partial game and sending it on the update_queue
            saved_game_proto = get_saved_game_proto(
                env=env,
                players=players,
                stored_board_state=stored_board_state,
                stored_prev_orders_state=stored_prev_orders_state,
                stored_possible_orders=stored_possible_orders,
                power_variables=power_variables,
                start_phase_ix=start_phase_ix,
                reward_fn=reward_fn,
                advantage_fn=advantage_fn,
                is_partial_game=True)
            update_queue.put_nowait(
                (False, nb_transitions, proto_to_bytes(saved_game_proto)))

            # Updating stats
            start_phase_ix = current_phase_ix
            nb_transitions = 0
            if not env.is_done:
                year_last_update = int(env.game.get_current_phase()[1:5])

    # Since the environment is done (Completed game) - We can leave the last_state_value at 0.
    for power_name in powers:
        power_variables[power_name]['last_state_value'] = 0.

    # Getting completed game
    saved_game_proto = get_saved_game_proto(
        env=env,
        players=players,
        stored_board_state=stored_board_state,
        stored_prev_orders_state=stored_prev_orders_state,
        stored_possible_orders=stored_possible_orders,
        power_variables=power_variables,
        start_phase_ix=0,
        reward_fn=reward_fn,
        advantage_fn=advantage_fn,
        is_partial_game=False)

    # Converting to correct format
    output = {
        'proto': lambda proto: proto,
        'zlib': proto_to_zlib,
        'bytes': proto_to_bytes
    }[output_format](saved_game_proto)

    # Returning
    return output
コード例 #5
0
def get_policy_data(saved_game_proto, power_names, top_victors):
    """ Computes the proto to save in tf.train.Example as a training example for the policy network
        :param saved_game_proto: A `.proto.game.SavedGame` object from the dataset.
        :param power_names: The list of powers for which we want the policy data
        :param top_victors: The list of powers that ended with more than 25% of the supply centers
        :return: A dictionary with key: the phase_ix
                              with value: A dict with the power_name as key and a dict with the example fields as value
    """
    # pylint: disable=too-many-branches
    nb_phases = len(saved_game_proto.phases)
    policy_data = {phase_ix: {} for phase_ix in range(nb_phases - 1)}
    game_id = saved_game_proto.id
    map_object = Map(saved_game_proto.map)

    # Determining if we have a draw
    nb_sc_to_win = len(map_object.scs) // 2 + 1
    has_solo_winner = max([len(saved_game_proto.phases[-1].state.centers[power_name].value)
                           for power_name in saved_game_proto.phases[-1].state.centers]) >= nb_sc_to_win
    survivors = [power_name for power_name in saved_game_proto.phases[-1].state.centers
                 if saved_game_proto.phases[-1].state.centers[power_name].value]
    has_draw = not has_solo_winner and len(survivors) >= 2

    # Processing all phases (except the last one
    current_year = 0
    for phase_ix in range(nb_phases - 1):

        # Building a list of orders of previous phases
        previous_orders_states = [np.zeros((NB_NODES, NB_ORDERS_FEATURES), dtype=np.uint8)] * NB_PREV_ORDERS
        for phase_proto in saved_game_proto.phases[max(0, phase_ix - NB_PREV_ORDERS_HISTORY):phase_ix]:
            if phase_proto.name[-1] == 'M':
                previous_orders_states += [proto_to_prev_orders_state(phase_proto, map_object)]
        previous_orders_states = previous_orders_states[-NB_PREV_ORDERS:]
        prev_orders_state = np.array(previous_orders_states)

        # Parsing each requested power in the specified phase
        phase_proto = saved_game_proto.phases[phase_ix]
        phase_name = phase_proto.name
        state_proto = phase_proto.state
        phase_board_state = proto_to_board_state(state_proto, map_object)

        # Increasing year for every spring or when the game is completed
        if phase_proto.name == 'COMPLETED' or (phase_proto.name[0] == 'S' and phase_proto.name[-1] == 'M'):
            current_year += 1

        for power_name in power_names:
            phase_issued_orders = get_issued_orders_for_powers(phase_proto, [power_name])
            phase_possible_orders = get_possible_orders_for_powers(phase_proto, [power_name])
            phase_draw_target = 1. if has_draw and phase_ix == (nb_phases - 2) and power_name in survivors else 0.

            # Data to use when not learning a policy
            blank_policy_data = {'board_state': phase_board_state,
                                 'prev_orders_state': prev_orders_state,
                                 'draw_target': phase_draw_target}

            # Power is not a top victor - We don't want to learn a policy from him
            if power_name not in top_victors:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Finding the orderable locs
            orderable_locations = list(phase_issued_orders[power_name].keys())

            # Skipping power for this phase if we are only issuing Hold
            for order_loc, order in phase_issued_orders[power_name].items():
                order_tokens = get_order_tokens(order)
                if len(order_tokens) >= 2 and order_tokens[1] != 'H':
                    break
            else:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Removing orderable locs where orders are not possible (i.e. NO_CHECK games)
            for order_loc, order in phase_issued_orders[power_name].items():
                if order not in phase_possible_orders[order_loc]:
                    if 'NO_CHECK' not in saved_game_proto.rules:
                        LOGGER.warning('%s not in all possible orders. Phase %s - Game %s.', order, phase_name, game_id)
                    orderable_locations.remove(order_loc)

            # Determining if we are in an adjustment phase
            in_adjustment_phase = state_proto.name[-1] == 'A'
            nb_builds = state_proto.builds[power_name].count
            nb_homes = len(state_proto.builds[power_name].homes)

            # WxxxA - We can build units
            # WxxxA - We can disband units
            # Other phase
            if in_adjustment_phase and nb_builds >= 0:
                decoder_length = TOKENS_PER_ORDER * min(nb_builds, nb_homes)
            elif in_adjustment_phase and nb_builds < 0:
                decoder_length = TOKENS_PER_ORDER * abs(nb_builds)
            else:
                decoder_length = TOKENS_PER_ORDER * len(orderable_locations)

            # Not all units were disbanded - Skipping this power as we can't learn the orders properly
            if in_adjustment_phase and nb_builds < 0 and len(orderable_locations) < abs(nb_builds):
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Not enough orderable locations for this power, skipping
            if not orderable_locations or not decoder_length:
                policy_data[phase_ix][power_name] = blank_policy_data
                continue

            # Encoding decoder inputs - Padding each order to 6 tokens
            # The decoder length should be a multiple of 6, since each order is padded to 6 tokens
            decoder_inputs = [GO_ID]
            for loc in orderable_locations[:]:
                order = phase_issued_orders[power_name][loc]
                try:
                    tokens = [token_to_ix(order_token) for order_token in get_order_tokens(order)] + [EOS_ID]
                    tokens += [PAD_ID] * (TOKENS_PER_ORDER - len(tokens))
                    decoder_inputs += tokens
                except KeyError:
                    LOGGER.warning('[data_generator] Order "%s" is not valid. Skipping location.', order)
                    orderable_locations.remove(loc)

            # Adding WAIVE orders
            if in_adjustment_phase and nb_builds > 0:
                waive_tokens = [token_to_ix('WAIVE'), EOS_ID] + [PAD_ID] * (TOKENS_PER_ORDER - 2)
                decoder_inputs += waive_tokens * (min(nb_builds, nb_homes) - len(orderable_locations))
            decoder_length = min(decoder_length, TOKENS_PER_ORDER * NB_SUPPLY_CENTERS)

            # Getting decoder mask
            coords = set()

            # Adjustment phase, we allow all builds / disbands in all positions
            if in_adjustment_phase:
                build_disband_locs = list(get_possible_orders_for_powers(phase_proto, [power_name]).keys())
                phase_board_alignments = get_board_alignments(build_disband_locs,
                                                              in_adjustment_phase=in_adjustment_phase,
                                                              tokens_per_loc=TOKENS_PER_ORDER,
                                                              decoder_length=decoder_length)

                # Building a list of all orders for all locations
                adj_orders = []
                for loc in build_disband_locs:
                    adj_orders += phase_possible_orders[loc]

                # Not learning builds for BUILD_ANY
                if nb_builds > 0 and 'BUILD_ANY' in state_proto.rules:
                    adj_orders = []

                # No orders found - Skipping
                if not adj_orders:
                    policy_data[phase_ix][power_name] = blank_policy_data
                    continue

                # Building a list of coordinates for the decoder mask matrix
                for loc_ix in range(decoder_length):
                    coords = get_token_based_mask(adj_orders, offset=loc_ix * TOKENS_PER_ORDER, coords=coords)

            # Regular phase, we mask for each location
            else:
                phase_board_alignments = get_board_alignments(orderable_locations,
                                                              in_adjustment_phase=in_adjustment_phase,
                                                              tokens_per_loc=TOKENS_PER_ORDER,
                                                              decoder_length=decoder_length)
                for loc_ix, loc in enumerate(orderable_locations):
                    coords = get_token_based_mask(phase_possible_orders[loc] or [''],
                                                  offset=loc_ix * TOKENS_PER_ORDER,
                                                  coords=coords)

            # Saving results
            # No need to return temperature, current_power, current_season
            policy_data[phase_ix][power_name] = {'board_state': phase_board_state,
                                                 'board_alignments': phase_board_alignments,
                                                 'prev_orders_state': prev_orders_state,
                                                 'decoder_inputs': decoder_inputs,
                                                 'decoder_lengths': decoder_length,
                                                 'decoder_mask_indices': list(sorted(coords)),
                                                 'draw_target': phase_draw_target}

    # Returning
    return policy_data