def load_raw_data(path, i, training_datasets):
# for i, path in enumerate(replay_files):
print("file idx :", i)
game = utils.HaliteV2(path)
print("loading file from ", path)
if game.game_play_list is not None:
"""
Four features as training input:
1) halite available
2) my ship
3) cargo on my ship
4) my shipyard
5) other players' ships
Sequence is a string that record the actions of ships/shipyards
"""
training_input, _, sequence = game.prepare_vae_encoder_input()
"""
Target ship actions:
"""
pad_offset = 6
board_size = game.config["size"]
# target actions
assert (inference_decoder.dictionary_size == 450)
decoder_input_sequence = []
decoder_target_sequence = []
# TODO: validate max sequence
for step, each_sequence in enumerate(sequence):
# Add ( and ) for teacher forcing
input_sequence = '( ' + each_sequence
output_sequence = each_sequence + ')'
decoder_input_sequence.append(input_sequence)
decoder_target_sequence.append(output_sequence)
assert (len(decoder_target_sequence) == len(decoder_input_sequence) ==
400)
train_dataset = [
training_input, decoder_input_sequence, decoder_target_sequence
]
training_datasets.append(train_dataset)
def test_build_gameplay_list(self):
file = "replayJson.json"
game = utils.HaliteV2(file)
self.assertIsNotNone(game.game_play_list)
self.assertEqual(game.total_turns - 1, len(game.game_play_list))
def test_convert_to_gameplay(self):
file = "replayJson.json"
game = utils.HaliteV2(file)
gameplay = game.convert_to_game_play(2)
self.assertIsNotNone(gameplay)
def test_load_total_turns(self):
file = "replayJson.json"
game = utils.HaliteV2(file)
self.assertGreater(game.total_turns, 1)
self.assertLessEqual(game.total_turns, 400)
def test_load_replay_v2(self):
file = "replayJson.json"
game = utils.HaliteV2(file)
self.assertIsNotNone(game.replay)
self.assertIsNotNone(game.config)
"""Data Extraction"""
PATH = 'train/top_replay/'
replay_files = []
# r=root, d=directories, f = files
for r, d, f in os.walk(PATH):
for file in f:
if '.json' in file:
replay_files.append(os.path.join(r, file))
for f in replay_files:
print(f)
game = None
seq_list = []
training_datasets = []
for i, path in enumerate(replay_files):
game = utils.HaliteV2(path)
print("index", i)
if i == 2:
break
if game.game_play_list is not None and game.winner_id == 0:
game.prepare_data_for_vae()
"""
Four features as training input:
1) halite available
2) my ship
3) cargo on my ship
4) my shipyard
"""
training_input = np.zeros((400, 32, 32, 4), dtype=np.float32)
my_ship_positions = game.ship_position
def save_training_data():
from numpy import save
PATH = 'top_replay/'
replay_files = []
# r=root, d=directories, f = files
for r, d, f in os.walk(PATH):
for file in f:
if '.json' in file:
replay_files.append(os.path.join(r, file))
for f in replay_files:
print(f)
game = None
# tables.open_file
for i, path in enumerate(replay_files):
game = utils.HaliteV2(path)
print("index", i)
# if i == 3:
# break
if game.game_play_list is not None and game.winner_id == 0:
game.prepare_data_for_vae()
training_input = np.zeros((400, 32, 32, 4), dtype='float32')
my_ship_positions = game.ship_position
target_ship_actions = game.ship_actions
halite_available = game.halite
my_shipyard = game.shipyard_position
my_cargo = game.cargo
"""
Target ship actions:
"""
training_label = np.zeros((400, 32, 32, 6), dtype='float32')
pad_offset = 6
# 1) halite available
for i, halite_map in enumerate(zip(halite_available)):
# print("halite_map", halite_map)
for row_indx, row in enumerate(halite_map[0]):
row = np.squeeze(row)
for col_indx, item in enumerate(row):
# print(item)
training_input[i, row_indx + pad_offset,
col_indx + pad_offset, 0] = item * 10
# 2) my ship position
for i, my_ship_position in enumerate(my_ship_positions):
for row_indx, row in enumerate(my_ship_position):
for col_indx, item in enumerate(row):
training_input[i, row_indx + pad_offset,
col_indx + pad_offset, 1] = item * 10
# 3) cargo on my ship
for i, cargo_map in enumerate(my_cargo):
for row_indx, row in enumerate(cargo_map):
for col_indx, item in enumerate(row):
training_input[i, row_indx + pad_offset,
col_indx + pad_offset, 2] = item * 10
# 4) my ship yard position
for i, shipyard_map in enumerate(my_shipyard):
for row_indx, row in enumerate(shipyard_map):
for col_indx, item in enumerate(row):
training_input[i, row_indx + pad_offset,
col_indx + pad_offset, 3] = item * 10
# target actions
for i, target_ship_action in enumerate(target_ship_actions):
for row_indx, row in enumerate(target_ship_action):
for col_indx, item in enumerate(row):
training_label[i, row_indx + pad_offset,
col_indx + pad_offset,
int(item)] = 1.
print("training input shape", training_input.shape)
# Do word embedding
board_size = game.config["size"]
vocab_dict = {}
num_dict = {}
for i in range(board_size**2):
vocab_dict[str(i)] = i
num_dict[i] = str(i)
vocab_idx = board_size**2
move_option = [
"EAST", "WEST", "SOUTH", "NORTH", "CONVERT", "SPAWN", "NO",
"(", ")"
]
for option in move_option:
vocab_dict[option] = vocab_idx
num_dict[vocab_idx] = option
vocab_idx += 1
# target actions
decoder_input_data = np.zeros((400, 50, len(vocab_dict)),
dtype='float32')
decoder_target_data = np.zeros((400, 50, len(vocab_dict)),
dtype='float32')
sequence = game.move_sequence
sequence.append(sequence[-1])
# TODO: validate max sequence
for step, each_sequence in enumerate(sequence):
each_sequence_list = each_sequence.split()
idx = 0
last_word = ""
for each_word in each_sequence_list:
assert (each_word in vocab_dict)
# TODO: Bug index > 50
if idx == 49:
break
assert (idx < 50)
if idx == 0:
decoder_input_data[step][idx][-2] = 1.
decoder_target_data[step][idx][
vocab_dict[each_word]] = 1.
else:
decoder_input_data[step][idx][
vocab_dict[last_word]] = 1.
decoder_target_data[step][idx][
vocab_dict[each_word]] = 1.
idx += 1
last_word = each_word
decoder_input_data[step][idx][vocab_dict[last_word]] = 1.
decoder_target_data[step][idx][-1] = 1.
print("target action shape", decoder_target_data.shape)
# data_tensor = tf.convert_to_tensor(decoder_input_data)
# data_tensor2 = tf.convert_to_tensor(decoder_target_data)
# train_dataset = tf.data.Dataset.from_tensor_slices(([training_input, decoder_input_data], decoder_target_data))
# save('encoder_input.npy', training_input)
# save('decoder_input.npy', decoder_input_data)
# save('decoder_target.npy', decoder_target_data)
filename = str(path.split('.')[0])
p = Path(filename)
np.savez(
filename,
encoder_input=training_input,
decoder_input=decoder_input_data,
decoder_output=decoder_target_data
) #[training_input, decoder_input_data, decoder_target_data])