# observation_size = sensor.hp.winx*sensor.hp.winy*2
observation_size = 530#2*(hp.resolution//2)**2+2
rising_beta_schedule = [[hp.beta_t1 // hp.steps_between_learnings, hp.beta_b1], [hp.beta_t2 // hp.steps_between_learnings, hp.beta_b2]]
flat_beta_schedule = [[hp.beta_t1 // hp.steps_between_learnings, hp.beta_b2], [hp.beta_t2 // hp.steps_between_learnings, hp.beta_b2]]
# rising_beta_schedule = [[400000 // hp.steps_between_learnings, 0.1], [700000 // hp.steps_between_learnings, 1]]
# flat_beta_schedule = [[400000 // hp.steps_between_learnings, 1.0], [700000 // hp.steps_between_learnings, 1]]
RL = DeepQNetwork(len(agent.hp.action_space), observation_size*hp.mem_depth,#sensor.frame_size+2,
reward_decay=0.99,
e_greedy=0.95,
e_greedy0=0.8,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=hp.syclop_learning_rate,
double_q=True,
dqn_mode=True,
state_table=np.zeros([1,observation_size*hp.mem_depth]),
soft_q_type='boltzmann',
beta_schedule=rising_beta_schedule if hp.rising_beta_schedule else flat_beta_schedule,
arch='mlp')
keras.backend.set_session(RL.dqn.sess)
if hp.decoder_initial_network is None:
decoder = rnn_model_102(lr=hp.decoder_learning_rate,ignore_input_B=hp.decoder_ignore_position,dropout=hp.decoder_dropout,rnn_type=hp.decoder_rnn_type,
input_size=(hp.resolution//2,hp.resolution//2, 1))
else:
decoder = keras.models.load_model(hp.decoder_initial_network) #for example: 'ref_nets/keras_decoder_5stp_101.model'
keras.backend.set_value(decoder.optimizer.lr, hp.decoder_learning_rate)
if not(hp.dqn_initial_network is None):
RL.dqn.load_nwk_param(hp.dqn_initial_network)
reward = syc.Rewards(
reward_types=['central_rms_intensity', 'speed', 'saccade'],
relative_weights=[1.0, -float(sys.argv[1]), 0])
# observation_size = sensor.hp.winx*sensor.hp.winy*2
observation_size = 64 * 64 + 2
RL = DeepQNetwork(len(agent.hp.action_space),
observation_size,
n_features_shaped=list(np.shape(sensor.dvs_view)) + [1],
shape_fun=None,
reward_decay=0.99,
e_greedy=0.95,
e_greedy0=0.8,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
state_table=np.zeros(
[1, observation_size * hp.mem_depth]),
soft_q_type='boltzmann',
beta=0.1,
beta_schedule=[[4000, 0.1], [7000, 1.0]],
arch='conv_ctrl')
# RL.dqn.load_nwk_param('tempX_1.nwk')
# RL.dqn.save_nwk_param('liron_encircle.nwk')
if not (hp.dqn_initial_network is None):
RL.dqn.load_nwk_param(hp.dqn_initial_network)
hp.scene = scene.hp
hp.sensor = sensor.hp
max_q=[scene.maxx - sensor.hp.winx, scene.maxy - sensor.hp.winy])
reward = syc.Rewards(
reward_types=['central_rms_intensity', 'speed', 'saccade'],
relative_weights=[1.0, hp.speed_penalty, -200])
# observation_size = sensor.hp.winx*sensor.hp.winy*2
observation_size = 256 * 4
RL = DeepQNetwork(
len(agent.hp.action_space),
observation_size * hp.mem_depth, #sensor.frame_size+2,
reward_decay=0.99,
e_greedy=0.95,
e_greedy0=0.8,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
state_table=np.zeros([1, observation_size * hp.mem_depth]),
soft_q_type='boltzmann',
beta_schedule=[[400000 // hp.steps_between_learnings, 0.1],
[700000 // hp.steps_between_learnings, 1]],
arch='mlp')
if not (hp.dqn_initial_network is None):
RL.dqn.load_nwk_param(hp.dqn_initial_network)
hp.scene = scene.hp
hp.sensor = sensor.hp
hp.agent = agent.hp
hp.reward = reward.hp