kernel_regularizer=l2(1e-4),
data_format='channels_first')(cv2)
dense = Flatten()(cv3)
dense = Dense(512, activation='relu', kernel_regularizer=l2(1e-4))(dense)
buttons = Dense(nb_actions, activation='linear',
kernel_regularizer=l2(1e-4))(dense)
model = Model(inputs=frame, outputs=buttons)
model.summary()
processor = AtariDQfDProcessor()
# record_demo_data('HeroDeterministic-v4', steps=50000, data_filepath='hero_expert.npy', frame_delay=0.03)
# Load and process the demonstration data.
expert_demo_data = processor.process_demo_data(
load_demo_data_from_file('hero_expert.npy'))
memory = PartitionedMemory(limit=1000000,
pre_load_data=expert_demo_data,
alpha=.4,
start_beta=.6,
end_beta=.6,
window_length=WINDOW_LENGTH)
policy = EpsGreedyQPolicy(.01)
dqfd = DQfDAgent(model=model,
nb_actions=nb_actions,
policy=policy,
memory=memory,
processor=processor,
enable_double_dqn=True,
student_model = Model(inputs=sensors, outputs=s_actions)
# "Expert" (regular dqn) model architecture
e_dense = Flatten()(sensors)
e_dense = Dense(64, activation='relu')(e_dense)
e_dense2 = Dense(128, activation='relu')(e_dense)
e_dense3 = Dense(64, activation='relu')(e_dense2)
e_actions = Dense(nb_actions, activation='linear')(e_dense3)
expert_model = Model(inputs=sensors, outputs=e_actions)
processor = RocketProcessor()
model_saves = './model_saves/'
if __name__ == "__main__":
if args.model == 'student':
# load expert data
expert_demo_data = load_demo_data_from_file(model_saves + 'demos.npy')
expert_demo_data = reward_threshold_subset(expert_demo_data, 0)
print(expert_demo_data.shape)
expert_demo_data = processor.process_demo_data(expert_demo_data)
# memory
memory = PartitionedMemory(limit=500000,
pre_load_data=expert_demo_data,
alpha=.6,
start_beta=.4,
end_beta=.4,
window_length=WINDOW_LENGTH)
# policy
policy = EpsGreedyQPolicy(.01)
# agent
dqfd = DQfDAgent(model=student_model,
nb_actions=nb_actions,
plot_model(curiosity_inverse_model,
show_shapes=True,
to_file=plot_file_prefix + 'curiosity_inverse_model.png')
######## END INVERSE MODEL ################
model_saves = './demonstrations/'
now = datetime.now()
datestr = now.strftime("%m%d_%H%M%S")
filename_append = args.filename_append
environment_name = args.env
if __name__ == "__main__":
if args.model == 'student':
# load expert data
expert_demo_data = load_demo_data_from_file(model_saves +
demonstrations_file)
expert_demo_data = reward_threshold_subset(expert_demo_data, 0)
print(expert_demo_data.shape)
expert_demo_data = processor.process_demo_data(expert_demo_data)
# memory
memory = PartitionedMemory(limit=500000,
pre_load_data=expert_demo_data,
alpha=.6,
start_beta=.4,
end_beta=.4,
window_length=WINDOW_LENGTH)
# policy
policy = EpsGreedyQPolicy(.01)
# agent
dqfd = DQfDAgent(model=student_model,
nb_actions=nb_actions,
episode_start = 0
for i, transition in enumerate(demo_array):
reward = transition[-2]
reward = np.sign(reward) * np.log(1 + abs(reward))
episode_total += reward
if transition[-1]: #terminal
episode_rs[episode_start:i + 1] = episode_total
episode_total = 0
episode_start = i + 1
return episode_rs
def reward_threshold_subset(demo_array, reward_min):
rs = calc_ep_rs(demo_array)
cropped_demos = []
for i, transition in enumerate(demo_array):
if rs[i] > reward_min:
cropped_demos.append(transition)
return np.array(cropped_demos)
def demo_avg(demo_array):
demo_array = np.array(list(set(calc_ep_rs(demo_array))))
return np.mean(demo_array)
if __name__ == "__main__":
record_demo_data('RocketLander-v0', steps=50000)
print(len(load_demo_data_from_file()))