def evaluate(model: str,
output_csv: str,
bat_speed: int,
shape_file: str = SHAPE_FILE):
for i in range(N_EPISODE):
random.seed(random.randint(0, 1000))
print("Episode : {}".format(i))
print("S : {}".format(bat_speed))
brain = Dqn(5, 2 * ANGLE_RANGE + 1, GAMMA)
if path.exists(model):
brain.load(model_file=model)
if path.exists(output_csv):
data = pd.read_csv(output_csv)
experiment_number = data['experiment'].max() + 1
else:
columns = [
'experiment', 'time', 'speed', 'gamma', 'signal1', 'signal2',
'signal3', 'distance_to_goal', 'action', 'orientation',
'reward'
]
if ADDING_OBS:
for i in range(-ANGLE_RANGE, ANGLE_RANGE + 1):
columns.append('angle_' + str(i))
data = pd.DataFrame(columns=columns)
experiment_number = 1
if ADDING_OBS:
game = Game(model=brain,
experiment_number=experiment_number,
bat_speed=bat_speed,
training_mode=False,
shape_file=shape_file)
else:
game = Game(model=brain,
experiment_number=experiment_number,
bat_speed=bat_speed,
training_mode=True,
shape_file=shape_file)
print("Size : {}".format(game.state.sand.shape))
done = False
n_actions = 0
while not done:
game.update()
result_of_moves = game.last_action()
# print(result_of_moves)
if result_of_moves == 'GOAL' or result_of_moves == 'HIT_TREE':
done = True
n_actions += 1
if n_actions >= N_MOVES:
done = True
print("Save data")
data = pd.concat([data, pd.DataFrame(game.state.sample)])
data.to_csv(output_csv, index=False)
def __init__(self, model: str, bat_speed: int):
super(Game, self).__init__()
self.height = 500
self.width = 500
self.action2rotation = [
i for i in range(-ANGLE_RANGE, ANGLE_RANGE + 1, 1)
]
self.state = State()
self.state.bat_speed = bat_speed
self.state.brain = Dqn(5, 2 * ANGLE_RANGE + 1, GAMMA)
if path.exists(model):
print("Loading brain")
self.state.brain.load(model)
self.state.experiment = 1
def __init__(self, **kwargs):
"""
Args:
**kwargs:
"""
super(Game, self).__init__()
self.height = 500
self.width = 500
self.action2rotation = [
i for i in range(-ANGLE_RANGE, ANGLE_RANGE + 1, 1)]
self.state = State()
self.state.brain = Dqn(5, 2 * ANGLE_RANGE + 1, GAMMA)
if path.exists(MODEL_FILE):
self.state.brain.load(MODEL_FILE)
self.state.experiment = 1
MODLE_FILE = 'base_brain.pth'
OUTPUT_CSV = 'eval.csv'
N_EPISODE = 300
N_MOVES = 1500
if __name__ == '__main__':
for i in range(N_EPISODE):
random.seed(random.randint(0, 1000))
print("Episode : {}".format(i))
# INIT ENV FOR TRAINING
# 1 DEFINE MODEL
brain = Dqn(5, 2 * ANGLE_RANGE + 1, GAMMA)
if path.exists(MODLE_FILE):
brain.load(MODLE_FILE)
if path.exists(OUTPUT_CSV):
data = pd.read_csv(OUTPUT_CSV)
experiment_number = data['experiment'].max() + 1
else:
columns = [
'experiment', 'time', 'speed', 'gamma', 'signal1', 'signal2',
'signal3', 'distance_to_goal', 'action', 'orientation',
'reward'
]
data = pd.DataFrame(columns=columns)
def training(model_file: str,
output_csv: str,
bat_speed: int,
num_episodes: int = N_EPISODE,
moves_per_episode: int = N_MOVES,
update_only_better_reward: bool = True,
shap_file: str = 'shape'):
for i in range(num_episodes):
print("Episode : {}".format(i))
# INIT ENV FOR TRAINING
# 1 define model
brain = Dqn(5, 2 * ANGLE_RANGE + 1, GAMMA)
# 2 loading and init the new data frame if needed
if path.exists(model_file):
brain.load(model_file)
if path.exists(output_csv):
data = pd.read_csv(output_csv)
experiment_number = data['experiment'].max() + 1
else:
columns = [
'experiment', 'time', 'speed', 'gamma', 'signal1', 'signal2',
'signal3', 'distance_to_goal', 'action', 'orientation',
'reward'
]
data = pd.DataFrame(columns=columns)
experiment_number = 1
game = Game(model=brain,
experiment_number=experiment_number,
bat_speed=bat_speed,
training_mode=True,
shape_file=shap_file)
for _ in range(moves_per_episode):
# Each update called would result an additional one row of data store in game.state.sample
game.update()
if update_only_better_reward:
if data.shape[0] == 0:
current_max_cumulative_reward = -10000000
else:
tmp_df = data[data['experiment'] == (experiment_number - 1)]
current_max_cumulative_reward = tmp_df.reward.sum()
new_df = pd.DataFrame(game.state.sample)
sum_reward = new_df['reward'].sum()
print('Sum reward {}'.format(sum_reward))
print("Current max {}".format(current_max_cumulative_reward))
if sum_reward > current_max_cumulative_reward:
print("Save data")
data = pd.concat([data, new_df])
data.to_csv(output_csv, index=False)
print('Save model')
game.state.brain.save(model_file)
else:
print("Save data")
data = pd.concat([data, pd.DataFrame(game.state.sample)])
data.to_csv(output_csv, index=False)
print('Save model')
game.state.brain.save(model_file)