def test_prediction_long_term_hard_brake(self):
config = TestConfig()
config.n_global_steps = 20000
config.env_id = 'HeuristicRiskEnv-v0'
config.discount = 0. # 599. / 600
config.max_timesteps = 10000
config.prime_timesteps = 50
config.learning_rate = 1e-3
config.adam_beta1 = .995
config.adam_beta2 = .999
config.dropout_keep_prob = 1.
config.l2_reg = 0.
config.local_steps_per_update = 20
config.hidden_layer_sizes = [32, 16]
config.hard_brake_threshold = -3.
config.hard_brake_n_past_frames = 1
config.target_loss_index = 3
config.loss_type = 'mse'
env = build_envs.create_env(config)
test_state = env.reset()
summary_writer = tf.summary.FileWriter('/tmp/test')
with tf.Session() as sess:
trainer = async_td.AsyncTD(env, 0, config)
sess.run(tf.global_variables_initializer())
sess.run(trainer.sync)
trainer.start(sess, summary_writer)
global_step = sess.run(trainer.global_step)
c, h = trainer.network.get_initial_features()
while global_step < config.n_global_steps:
trainer.process(sess)
global_step = sess.run(trainer.global_step)
value = trainer.network.value(test_state, c, h)
print(value)
def test_heuristic_deterministic_case(self):
config = TestConfig()
config.n_global_steps = 50000
config.max_timesteps = 50
config.env_id = 'BayesNetRiskEnv-v0'
config.discount = 1. # 49. / 50
config.value_dim = 5
config.adam_beta1 = .9
config.local_steps_per_update = 100
config.hidden_layer_sizes = [128]
config.learning_rate = 1e-3
config.learning_rate_end = 5e-6
config.loss_type = 'mse'
config.target_loss_index = 1
config.validation_dataset_filepath = '/Users/wulfebw/Dropbox/School/Stanford/research/risk/risk_prediction/data/experiments/heuristic_determinstic_1_lane_5_sec/data/subselect_proposal_prediction_data.h5'
config.max_validation_samples = 1
config.validate_every = 1000
config.visualize_every = 10000
config.summarize_features = True
validation.transfer_dataset_settings_to_config(
config.validation_dataset_filepath, config)
config.base_bn_filepath = '/Users/wulfebw/Dropbox/School/Stanford/research/risk/risk_prediction/data/experiments/heuristic_determinstic_1_lane_5_sec/data/base_bn_filepath.h5'
config.base_prop_filepath = '/Users/wulfebw/Dropbox/School/Stanford/research/risk/risk_prediction/data/experiments/heuristic_determinstic_1_lane_5_sec/data/prop_bn_filepath.h5'
config.max_validation_samples = 1000
# config.roadway_radius = 400.
# config.roadway_length = 100.
# config.lon_accel_std_dev = 0.
# config.lat_accel_std_dev = 0.
# config.overall_response_time = .2
# config.lon_response_time = .0
# config.err_p_a_to_i = .15
# config.err_p_i_to_a = .3
# config.max_num_vehicles = 50
# config.min_num_vehicles = 50
# config.hard_brake_threshold = -3.
# config.hard_brake_n_past_frames = 2
# config.min_base_speed = 30.
# config.max_base_speed = 30.
# config.min_vehicle_length = 5.
# config.max_vehicle_length = 5.
# config.min_vehicle_width = 2.5
# config.max_vehicle_width = 2.5
# config.min_init_dist = 10.
# config.heuristic_behavior_type = "normal"
# build env
env = build_envs.create_env(config)
dataset = validation.build_dataset(config, env)
print('mean validation targets: {}'.format(np.mean(dataset.y, axis=0)))
# run it
summary_writer = tf.summary.FileWriter('/tmp/test')
avg_loss = -1
last_global_step_val = 0
with tf.Session() as sess:
trainer = async_td.AsyncTD(env, 0, config)
sess.run(tf.global_variables_initializer())
sess.run(trainer.sync)
trainer.start(sess, summary_writer)
global_step = sess.run(trainer.global_step)
while global_step < config.n_global_steps:
trainer.process(sess)
if (global_step -
last_global_step_val) > config.validate_every:
avg_loss = trainer.validate(sess, dataset)
last_global_step_val = global_step
global_step = sess.run(trainer.global_step)