# Okay, now it's time to learn something! We capture the interrupt exception so that training
# can be prematurely aborted. Notice that you can use the built-in Keras callbacks!
os.chdir(DIR)
os.chdir('runs')
if not args.run_id:
run_id = '{}_{}'.format(args.env_name, time())
else:
run_id = args.run_id
os.mkdir(run_id)
weights_filename = '{}/weights.h5f'.format(run_id)
checkpoint_weights_filename = weights_filename + '_{step}.h5f'
# model_filename = 'dqn_micropolis_model.hdf5'
log_filename = '{}/log.json'.format(run_id)
callbacks = [ModelIntervalCheckpoint(checkpoint_weights_filename, interval=250000)]
# callbacks += [ModelCheckpoint(model_filename)]
callbacks += [FileLogger(log_filename, interval=250000)]
callbacks += [TensorBoard(log_dir=run_id)]
# class TestCallback(Callback):
# def on_epoch_end(self, epoch, logs=None):
# test_env = gym.make(args.env_name)
# test_env.setMapSize(MAP_X,MAP_Y)
# dqn.test(test_env, nb_episodes=1, visualize=True, nb_max_start_steps=100)
# test_env.win1.destroy()
# test_env.close()
# del(test_env)
# callbacks += [TestCallback()]
# if args.loadmodel:
# dqn.model.load(args.loadmodel)
if args.weights:
dqn.load_weights(args.weights)
def training_game():
env = Environment(
map_name="HallucinIce",
visualize=True,
game_steps_per_episode=150,
agent_interface_format=features.AgentInterfaceFormat(
feature_dimensions=features.Dimensions(screen=64, minimap=32)))
input_shape = (_SIZE, _SIZE, 1)
nb_actions = 12 # Number of actions
model = neural_network_model(input_shape, nb_actions)
memory = SequentialMemory(limit=5000, window_length=_WINDOW_LENGTH)
processor = SC2Proc()
# Policy
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(),
attr="eps",
value_max=1,
value_min=0.7,
value_test=.0,
nb_steps=1e6)
# Agent
dqn = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
enable_double_dqn=False,
nb_steps_warmup=500,
target_model_update=1e-2,
policy=policy,
batch_size=150,
processor=processor)
dqn.compile(Adam(lr=.001), metrics=["mae"])
# Save the parameters and upload them when needed
name = "HallucinIce"
w_file = "dqn_{}_weights.h5f".format(name)
check_w_file = "train_w" + name + "_weights.h5f"
if SAVE_MODEL:
check_w_file = "train_w" + name + "_weights_{step}.h5f"
log_file = "training_w_{}_log.json".format(name)
callbacks = [ModelIntervalCheckpoint(check_w_file, interval=1000)]
callbacks += [FileLogger(log_file, interval=100)]
if LOAD_MODEL:
dqn.load_weights(w_file)
dqn.fit(env,
callbacks=callbacks,
nb_steps=1e7,
action_repetition=2,
log_interval=1e4,
verbose=2)
dqn.save_weights(w_file, overwrite=True)
dqn.test(env, action_repetition=2, nb_episodes=30, visualize=False)
