memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
soft_q_type='boltzmann',
beta_schedule=[[400000//hp.steps_between_learnings, 1], [700000//hp.steps_between_learnings, 10]],
arch='conv_saccades_v1',
n_modulating_features=hp.drift_state_size
)
# at this point drift network is a standalone network taken from some external source (e.g. pretrained)
# in future it will be an action generating network from the drift loop
# drift_net = Stand_alone_net(16*16,10,arch='mlp', layer_size = [None]+[100]+[100]+[ None])
drift_net = Stand_alone_net([32,32,1],10,arch='conv',
layer_size = [None]+[[3,3,32]]+[[2,2,16]]+[200]+[ None],
loss_type='softmax_cross_entropy',
trainable=True,
lr=0.0005,
lambda_reg=0.0) #simple cifar10 classifier
drift_net.assign_session_to_nwk(saccade_RL.dqn.sess)
saccade_RL.dqn.sess.run(tf.global_variables_initializer())
saccade_RL.dqn.reset()
if not(hp.drift_initial_network is None):
drift_net.load_nwk_param(hp.drift_initial_network)
if not(hp.dqn_initial_network is None):
saccade_RL.dqn.load_nwk_param(hp.dqn_initial_network)
# hp.scene = scene.hp
print('debug hp',sensor.hp.centralwinx,sensor.hp.centralwiny)
hp.sensor = sensor.hp
hp.saccade_agent = saccade_agent.hp
hp.reward = reward.hp
hp.saccade_RL = saccade_RL.hp
double_q=True,
dqn_mode=True,
soft_q_type='boltzmann',
beta_schedule=[[400000 // hp.steps_between_learnings, 1],
[700000 // hp.steps_between_learnings, 10]],
arch='conv_saccades_v1',
n_modulating_features=hp.drift_state_size)
# at this point drift network is a standalone network taken from some external source (e.g. pretrained)
# in future it will be an action generating network from the drift loop
# drift_net = Stand_alone_net(16*16,10,arch='mlp', layer_size = [None]+[100]+[100]+[ None])
drift_net = Stand_alone_net(
[32, 32, 3],
10,
arch='conv',
layer_size=[None] + [[5, 5, 96], [5, 5, 80], [5, 5, 64], [5, 5, 64]] +
[200] + [None],
loss_type='softmax_cross_entropy',
trainable=True,
lr=1,
dropout_p_keep=0.99,
lambda_reg=0.0)
drift_net.assign_session_to_nwk(saccade_RL.dqn.sess)
saccade_RL.dqn.sess.run(tf.global_variables_initializer())
saccade_RL.dqn.reset()
if not (hp.drift_initial_network is None):
drift_net.load_nwk_param(hp.drift_initial_network)
if not (hp.dqn_initial_network is None):
saccade_RL.dqn.load_nwk_param(hp.dqn_initial_network)
# hp.scene = scene.hp
print('debug hp', sensor.hp.centralwinx, sensor.hp.centralwiny)
hp.sensor = sensor.hp
reward_decay=0.99,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
soft_q_type='boltzmann',
beta_schedule=[[4000, 1], [7000, 10]],
arch='conv_saccades_v1')
# at this point drift network is a standalone network taken from some external source (e.g. pretrained)
# in future it will be an action generating network from the drift loop
# drift_net = Stand_alone_net(16*16,10,arch='mlp', layer_size = [None]+[100]+[100]+[ None])
drift_net = Stand_alone_net(16 * 16,
16 * 16,
arch='mlp',
layer_size=[None] + [100] + [100] +
[None]) #ae
drift_net.assign_session_to_nwk(saccade_RL.dqn.sess)
saccade_RL.dqn.sess.run(tf.global_variables_initializer())
saccade_RL.dqn.reset()
drift_net.load_nwk_param(hp.drift_initial_network)
if not (hp.dqn_initial_network is None):
saccade_RL.dqn.load_nwk_param(hp.dqn_initial_network)
# hp.scene = scene.hp
hp.sensor = sensor.hp
hp.saccade_agent = saccade_agent.hp
hp.reward = reward.hp
hp.saccade_RL = saccade_RL.hp
deploy_logs()
with open(hp.this_run_path + hp_file, 'wb') as f:
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
soft_q_type='boltzmann',
beta_schedule=[[400000 // hp.steps_between_learnings, 1],
[700000 // hp.steps_between_learnings, 10]],
arch='conv_saccades_v1',
n_modulating_features=16)
# at this point drift network is a standalone network taken from some external source (e.g. pretrained)
# in future it will be an action generating network from the drift loop
# drift_net = Stand_alone_net(16*16,10,arch='mlp', layer_size = [None]+[100]+[100]+[ None])
drift_net = Stand_alone_net(16 * 16,
16 * 16,
arch='mlp',
layer_size=[None] + [100, 16, 100] + [None],
loss_type='mean_squared',
trainable=True,
lr=0.0005,
lambda_reg=0.0) #ae
drift_net.assign_session_to_nwk(saccade_RL.dqn.sess)
saccade_RL.dqn.sess.run(tf.global_variables_initializer())
saccade_RL.dqn.reset()
if not (hp.drift_initial_network is None):
drift_net.load_nwk_param(hp.drift_initial_network)
if not (hp.dqn_initial_network is None):
saccade_RL.dqn.load_nwk_param(hp.dqn_initial_network)
# hp.scene = scene.hp
hp.sensor = sensor.hp
hp.saccade_agent = saccade_agent.hp
hp.reward = reward.hp
hp.saccade_RL = saccade_RL.hp