def main(main_args):
### Set all arguments
## Target maker
target_maker_args = {}
target_maker_args['future_steps'] = [1,2,4,8,16,32]
target_maker_args['meas_to_predict'] = [0,1,2]
target_maker_args['min_num_targs'] = 3
target_maker_args['rwrd_schedule_type'] = 'exp'
target_maker_args['gammas'] = []
target_maker_args['invalid_targets_replacement'] = 'nan'
## Simulator
simulator_args = {}
simulator_args['config'] = '../../maps/D3_battle.cfg'
simulator_args['resolution'] = (84,84)
simulator_args['frame_skip'] = 4
simulator_args['color_mode'] = 'GRAY'
simulator_args['use_shaping_reward'] = False
simulator_args['maps'] = ['MAP01']
simulator_args['switch_maps'] = False
#train
simulator_args['num_simulators'] = 8
## Experience
# Train experience
train_experience_args = {}
train_experience_args['memory_capacity'] = 20000
train_experience_args['history_length'] = 1
train_experience_args['history_step'] = 1
train_experience_args['action_format'] = 'enumerate'
train_experience_args['shared'] = False
# Test prediction experience
test_prediction_experience_args = train_experience_args.copy()
test_prediction_experience_args['memory_capacity'] = 1
# Test policy experience
test_policy_experience_args = train_experience_args.copy()
test_policy_experience_args['memory_capacity'] = 55000
## Agent
agent_args = {}
# agent type
agent_args['agent_type'] = 'advantage'
# preprocessing
agent_args['preprocess_input_images'] = lambda x: x / 255. - 0.5
agent_args['preprocess_input_measurements'] = lambda x: x / 100. - 0.5
targ_scale_coeffs = np.expand_dims((np.expand_dims(np.array([7.5,30.,1.]),1) * np.ones((1,len(target_maker_args['future_steps'])))).flatten(),0)
agent_args['preprocess_input_targets'] = lambda x: x / targ_scale_coeffs
agent_args['postprocess_predictions'] = lambda x: x * targ_scale_coeffs
# agent properties
agent_args['objective_coeffs_temporal'] = [0., 0. ,0. ,0.5, 0.5, 1.]
agent_args['objective_coeffs_meas'] = [0.5, 0.5, 1.]
agent_args['random_exploration_schedule'] = lambda step: (0.02 + 145000. / (float(step) + 150000.))
agent_args['new_memories_per_batch'] = 8
# net parameters
agent_args['conv_params'] = np.array([(32,8,4), (64,4,2), (64,3,1)],
dtype = [('out_channels',int), ('kernel',int), ('stride',int)])
agent_args['fc_img_params'] = np.array([(512,)], dtype = [('out_dims',int)])
agent_args['fc_meas_params'] = np.array([(128,), (128,), (128,)], dtype = [('out_dims',int)])
agent_args['fc_joint_params'] = np.array([(512,), (-1,)], dtype = [('out_dims',int)]) # we put -1 here because it will be automatically replaced when creating the net
agent_args['weight_decay'] = 0.00000
# optimization parameters
agent_args['batch_size'] = 64
agent_args['init_learning_rate'] = 0.0001
agent_args['lr_step_size'] = 250000
agent_args['lr_decay_factor'] = 0.3
agent_args['adam_beta1'] = 0.95
agent_args['adam_epsilon'] = 1e-4
agent_args['optimizer'] = 'Adam'
agent_args['reset_iter_count'] = False
# directories
agent_args['checkpoint_dir'] = 'checkpoints'
agent_args['log_dir'] = 'logs'
agent_args['init_model'] = ''
agent_args['model_name'] = "predictor.model"
agent_args['model_dir'] = time.strftime("%Y_%m_%d_%H_%M_%S")
# logging and testing
agent_args['print_err_every'] = 50
agent_args['detailed_summary_every'] = 1000
agent_args['test_pred_every'] = 0
agent_args['test_policy_every'] = 7812
agent_args['num_batches_per_pred_test'] = 0
agent_args['num_steps_per_policy_test'] = test_policy_experience_args['memory_capacity'] / simulator_args['num_simulators']
agent_args['checkpoint_every'] = 10000
agent_args['save_param_histograms_every'] = 5000
agent_args['test_policy_in_the_beginning'] = True
# experiment arguments
experiment_args = {}
experiment_args['num_train_iterations'] = 820000
experiment_args['test_objective_coeffs_temporal'] = np.array([0., 0., 0., 0.5, 0.5, 1.])
experiment_args['test_objective_coeffs_meas'] = np.array([0.5,0.5,1.])
experiment_args['test_random_prob'] = 0.
experiment_args['test_checkpoint'] = 'checkpoints/2017_04_08_10_44_20'
experiment_args['test_policy_num_steps'] = 2000
experiment_args['show_predictions'] = False
experiment_args['multiplayer'] = False
# Create and run the experiment
experiment = MultiExperiment(target_maker_args=target_maker_args,
simulator_args=simulator_args,
train_experience_args=train_experience_args,
test_policy_experience_args=test_policy_experience_args,
agent_args=agent_args,
experiment_args=experiment_args)
experiment.run(main_args[0])
def main(main_args):
### Set all arguments
## Target maker
target_maker_args = {}
target_maker_args['future_steps'] = [
1, 2, 4, 8, 16, 32
] # Offsets with which to predict the measurements
target_maker_args['meas_to_predict'] = [
0, 1, 2, 3
] # Measurements that we aim to predict
target_maker_args[
'min_num_targs'] = 3 # Defines the minimum nb of available measurements needed to try evaluate a frame
target_maker_args[
'rwrd_schedule_type'] = 'exp' # Possible discount for future rewards
target_maker_args['gammas'] = []
target_maker_args[
'invalid_targets_replacement'] = 'nan' # Replacement for unavailable targets (when close to the end of experiment)
## Simulator
simulator_args = {}
simulator_args[
'config'] = '../../maps/D5.cfg' # defines some parameter (available buttons, game variables etc)
simulator_args['resolution'] = (84, 84
) # Resolution used for the images (?)
simulator_args[
'frame_skip'] = 4 # Take decisions and predict every n frames (?)
simulator_args['color_mode'] = 'GRAY'
simulator_args['maps'] = ['MAP01'] # Map on which to play
simulator_args['switch_maps'] = False
#train
simulator_args[
'num_simulators'] = 8 # Number of simulations to run in parallel (?)
## Experience
# Train experience
train_experience_args = {}
train_experience_args[
'memory_capacity'] = 20000 # Number of observations to retain
train_experience_args[
'history_length'] = 1 # Number of frames taken as input when making a prediction
train_experience_args['history_step'] = 1
train_experience_args['action_format'] = 'enumerate'
train_experience_args['shared'] = False
# Test prediction experience
test_prediction_experience_args = train_experience_args.copy()
test_prediction_experience_args['memory_capacity'] = 1
# Test policy experience
test_policy_experience_args = train_experience_args.copy()
test_policy_experience_args['memory_capacity'] = 55000
## Agent
agent_args = {}
# agent type
agent_args[
'agent_type'] = 'advantage' # Defines the kind of network used for the prediction
# preprocessing
agent_args[
'preprocess_input_images'] = lambda x: x / 255. - 0.5 # Preprocessing to apply to the images
agent_args[
'preprocess_input_measurements'] = lambda x: x / 100. - 0.5 # Preprocessing to apply to measurements
targ_scale_coeffs = np.expand_dims(
(np.expand_dims(np.array([30., 100., 100., 100.]), 1) * np.ones(
(1, len(target_maker_args['future_steps'])))).flatten(),
0) # addapt dim to meas?
# scaling at 100 works for VELOCITY meas
agent_args[
'preprocess_input_targets'] = lambda x: x / targ_scale_coeffs # targ_scale_coeffs is a simple array of 1
agent_args['postprocess_predictions'] = lambda x: x * targ_scale_coeffs
# agent properties
agent_args['objective_coeffs_temporal'] = [
0., 0., 0., 0.5, 0.5, 1.
] # Multiplicative factors for rewards accross predicted time steps
agent_args['objective_coeffs_meas'] = [1., 0., 0.,
0.] # Weights of means for reward
agent_args['random_exploration_schedule'] = lambda step: (
0.02 + 145000. /
(float(step) + 150000.)) # epsilon for epsilon-greedy policy (?)
agent_args['new_memories_per_batch'] = 8
# net parameters
agent_args['conv_params'] = np.array([(32, 8, 4), (64, 4, 2), (64, 3, 1)],
dtype=[('out_channels', int),
('kernel', int),
('stride', int)])
agent_args['fc_img_params'] = np.array([(512, )],
dtype=[('out_dims', int)])
agent_args['fc_meas_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_joint_params'] = np.array(
[(512, ), (-1, )], dtype=[('out_dims', int)]
) # we put -1 here because it will be automatically replaced when creating the net
agent_args['weight_decay'] = 0.00000
# optimization parameters
agent_args['batch_size'] = 64
agent_args['init_learning_rate'] = 0.0001
agent_args['lr_step_size'] = 250000
agent_args['lr_decay_factor'] = 0.3
agent_args['adam_beta1'] = 0.95
agent_args['adam_epsilon'] = 1e-4
agent_args['optimizer'] = 'Adam'
agent_args['reset_iter_count'] = False
# directories
agent_args['checkpoint_dir'] = 'checkpoints'
agent_args['log_dir'] = 'logs'
agent_args['init_model'] = ''
agent_args['model_name'] = "predictor.model"
agent_args['model_dir'] = time.strftime("%Y_%m_%d_%H_%M_%S")
# logging and testing
agent_args['print_err_every'] = 50
agent_args['detailed_summary_every'] = 1000
agent_args['test_pred_every'] = 0
agent_args['test_policy_every'] = 7812
agent_args['num_batches_per_pred_test'] = 0
agent_args['num_steps_per_policy_test'] = test_policy_experience_args[
'memory_capacity'] / simulator_args['num_simulators']
agent_args['checkpoint_every'] = 10000
agent_args['save_param_histograms_every'] = 5000
agent_args['test_policy_in_the_beginning'] = True
# experiment arguments
experiment_args = {}
experiment_args['num_train_iterations'] = 820000
experiment_args['test_objective_coeffs_temporal'] = np.array(
[0., 0., 0., 0.5, 0.5, 1.])
experiment_args['test_objective_coeffs_meas'] = np.array([1., 0., 0., 0.])
experiment_args['test_random_prob'] = 0.
experiment_args['test_checkpoint'] = 'checkpoints/2017_04_09_09_11_48'
experiment_args['test_policy_num_steps'] = 2000
experiment_args['show_predictions'] = False
experiment_args['multiplayer'] = False
# Create and run the experiment
experiment = MultiExperiment(
target_maker_args=target_maker_args,
simulator_args=simulator_args,
train_experience_args=train_experience_args,
test_policy_experience_args=test_policy_experience_args,
agent_args=agent_args,
experiment_args=experiment_args)
experiment.run(main_args[0])
def main(mode, doom_config_file):
### Set all arguments
## Target maker
target_maker_args = {}
target_maker_args['future_steps'] = [1, 2, 4, 8, 16, 32]
target_maker_args['meas_to_predict'] = [0, 1, 2]
target_maker_args['min_num_targs'] = 3
target_maker_args['rwrd_schedule_type'] = 'exp'
target_maker_args['gammas'] = []
target_maker_args['invalid_targets_replacement'] = 'nan'
## Simulator
simulator_args = {}
simulator_args[
'config'] = '../../maps/' + doom_config_file #D3_battle.cfg'
simulator_args['resolution'] = (84, 84)
simulator_args[
'frame_skip'] = 1 #4 #TODO Change back to 4 for experiements. 1 helps get nicer videos though.
simulator_args['color_mode'] = 'GRAY'
simulator_args['maps'] = ['MAP01']
simulator_args['switch_maps'] = False
#train
simulator_args[
'num_simulators'] = 8 #TODO Set back to 8, just testing something. 8
## Experience
# Train experience
train_experience_args = {}
train_experience_args['memory_capacity'] = 20000
train_experience_args['history_length'] = 1
train_experience_args['history_step'] = 1
train_experience_args['action_format'] = 'enumerate'
train_experience_args['shared'] = False
# Test prediction experience
test_prediction_experience_args = train_experience_args.copy()
test_prediction_experience_args['memory_capacity'] = 1
# Test policy experience
test_policy_experience_args = train_experience_args.copy()
test_policy_experience_args['memory_capacity'] = 55000
## Agent
agent_args = {}
# agent type
agent_args['agent_type'] = 'advantage'
# preprocessing
agent_args['preprocess_input_images'] = lambda x: x / 255. - 0.5
agent_args['preprocess_input_measurements'] = lambda x: x / 100. - 0.5
targ_scale_coeffs = np.expand_dims(
(np.expand_dims(np.array([7.5, 30., 1.]), 1) * np.ones(
(1, len(target_maker_args['future_steps'])))).flatten(), 0)
agent_args['preprocess_input_targets'] = lambda x: x / targ_scale_coeffs
agent_args['postprocess_predictions'] = lambda x: x * targ_scale_coeffs
# agent properties
agent_args['objective_coeffs_temporal'] = [0., 0., 0., 0.5, 0.5, 1.]
agent_args['objective_coeffs_meas'] = [
0.5, 0.5, 1.
] #KOE: Note these are vals for training (?) for testing is below.
agent_args['random_exploration_schedule'] = lambda step: (
0.02 + 145000. / (float(step) + 150000.))
agent_args['new_memories_per_batch'] = 8
agent_args['random_objective_coeffs'] = True
agent_args['objective_coeffs_distribution'] = 'uniform_pos_neg'
# net parameters
agent_args['conv_params'] = np.array([(32, 8, 4), (64, 4, 2), (64, 3, 1)],
dtype=[('out_channels', int),
('kernel', int),
('stride', int)])
agent_args['fc_img_params'] = np.array([(512, )],
dtype=[('out_dims', int)])
agent_args['fc_meas_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_obj_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_joint_params'] = np.array(
[(512, ), (-1, )], dtype=[('out_dims', int)]
) # we put -1 here because it will be automatically replaced when creating the net
agent_args['weight_decay'] = 0.00000
# optimization parameters
agent_args['batch_size'] = 64
agent_args['init_learning_rate'] = 0.0001
agent_args['lr_step_size'] = 250000
agent_args['lr_decay_factor'] = 0.3
agent_args['adam_beta1'] = 0.95
agent_args['adam_epsilon'] = 1e-4
agent_args['optimizer'] = 'Adam'
agent_args['reset_iter_count'] = False
# directories
agent_args['checkpoint_dir'] = 'checkpoints'
agent_args['log_dir'] = 'logs'
agent_args['init_model'] = ''
agent_args['model_name'] = "predictor.model"
agent_args['model_dir'] = time.strftime("%Y_%m_%d_%H_%M_%S")
# logging and testing
agent_args['print_err_every'] = 50
agent_args['detailed_summary_every'] = 1000
agent_args['test_pred_every'] = 0
agent_args['test_policy_every'] = 7812
agent_args['num_batches_per_pred_test'] = 0
agent_args['num_steps_per_policy_test'] = test_policy_experience_args[
'memory_capacity'] / simulator_args['num_simulators']
agent_args['checkpoint_every'] = 10000
agent_args['save_param_histograms_every'] = 5000
agent_args['test_policy_in_the_beginning'] = True
# experiment arguments
experiment_args = {}
experiment_args['num_train_iterations'] = 820000
# KOE: This defines the weights and temporal weights of objectives.
experiment_args['test_objective_coeffs_temporal'] = np.array(
[0., 0., 0., 0.5, 0.5, 1.])
experiment_args['test_objective_coeffs_meas'] = np.array([0.5, 0.5, 1.])
#KOETODO Revert to old values above.
#Health, ammo, frags
#experiment_args['test_objective_coeffs_meas'] = np.array([-1,-1,-1]) KOE Opposite objectives, just for testing.
experiment_args['test_random_prob'] = 0.
experiment_args[
'test_checkpoint'] = 'checkpoints/2017_04_09_09_13_17' #KOE: This defines the weights to load
experiment_args[
'test_policy_num_steps'] = 2000 #KOE: How many steps to run the test agent.
experiment_args['show_predictions'] = False
experiment_args['multiplayer'] = False
# Create and run the experiment
experiment = MultiExperiment(
target_maker_args=target_maker_args,
simulator_args=simulator_args,
train_experience_args=train_experience_args,
test_policy_experience_args=test_policy_experience_args,
agent_args=agent_args,
experiment_args=experiment_args)
return experiment.run(mode)
def main(main_args):
### Set all arguments
## Target maker
target_maker_args = {}
target_maker_args['future_steps'] = [1, 2, 4, 8, 16, 32]
target_maker_args['meas_to_predict'] = [0]
target_maker_args['min_num_targs'] = 3
target_maker_args['rwrd_schedule_type'] = 'exp'
target_maker_args['gammas'] = []
target_maker_args['invalid_targets_replacement'] = 'nan'
## Simulator
simulator_args = {}
simulator_args['config'] = '../../maps/D1_basic.cfg'
simulator_args['resolution'] = (84, 84)
simulator_args['frame_skip'] = 4
simulator_args['color_mode'] = 'GRAY'
# simulator_args['color_mode'] = 'RGB'
simulator_args['maps'] = ['MAP01']
simulator_args['switch_maps'] = False
#train
simulator_args['num_simulators'] = 8
## Experience
# Train experience
train_experience_args = {}
train_experience_args['memory_capacity'] = 20000
train_experience_args['history_length'] = 1
train_experience_args['history_step'] = 1
train_experience_args['action_format'] = 'enumerate'
train_experience_args['shared'] = False
# Test prediction experience
test_prediction_experience_args = train_experience_args.copy()
test_prediction_experience_args['memory_capacity'] = 1
# Test policy experience
test_policy_experience_args = train_experience_args.copy()
test_policy_experience_args['memory_capacity'] = 55000
## Agent
agent_args = {}
# agent type
# agent_args['agent_type'] = 'advantage'
# agent_args['agent_type'] = 'advantage_objects'
agent_args['agent_type'] = 'advantage_objects_depth'
# agent_args['agent_type'] = 'advantage_objects_no_image'
# preprocessing
agent_args['preprocess_input_images'] = lambda x: x / 255. - 0.5
agent_args['preprocess_input_measurements'] = lambda x: x / 100. - 0.5
targ_scale_coeffs = np.expand_dims(
(np.expand_dims(np.array([30.]), 1) * np.ones(
(1, len(target_maker_args['future_steps'])))).flatten(), 0)
agent_args['preprocess_input_targets'] = lambda x: x / targ_scale_coeffs
agent_args['postprocess_predictions'] = lambda x: x * targ_scale_coeffs
# agent properties
agent_args['objective_coeffs_temporal'] = [0., 0., 0., 0.5, 0.5, 1.]
agent_args['objective_coeffs_meas'] = [1.]
agent_args['random_exploration_schedule'] = lambda step: (
0.02 + 145000. / (float(step) + 150000.))
agent_args['new_memories_per_batch'] = 8
# net parameters
agent_args['conv_params'] = np.array([(32, 8, 4), (64, 4, 2), (64, 3, 1)],
dtype=[('out_channels', int),
('kernel', int),
('stride', int)])
agent_args['fc_img_params'] = np.array([(512, )],
dtype=[('out_dims', int)])
agent_args['fc_meas_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_objects_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_joint_params'] = np.array(
[(512, ), (-1, )], dtype=[('out_dims', int)]
) # we put -1 here because it will be automatically replaced when creating the net
agent_args['weight_decay'] = 0.00000
# optimization parameters
agent_args['batch_size'] = 64
agent_args['init_learning_rate'] = 0.0001
agent_args['lr_step_size'] = 250000
agent_args['lr_decay_factor'] = 0.3
agent_args['adam_beta1'] = 0.95
agent_args['adam_epsilon'] = 1e-4
agent_args['optimizer'] = 'Adam'
agent_args['reset_iter_count'] = False
# directories
agent_args['checkpoint_dir'] = 'checkpoints'
agent_args['log_dir'] = 'logs'
agent_args['init_model'] = ''
agent_args['model_name'] = "predictor.model"
agent_args['model_dir'] = time.strftime("%Y_%m_%d_%H_%M_%S")
# logging and testing
agent_args['print_err_every'] = 50
agent_args['detailed_summary_every'] = 1000
agent_args['test_pred_every'] = 0
agent_args['test_policy_every'] = 7812
agent_args['num_batches_per_pred_test'] = 0
agent_args['num_steps_per_policy_test'] = test_policy_experience_args[
'memory_capacity'] / simulator_args['num_simulators']
agent_args['checkpoint_every'] = 10000
agent_args['save_param_histograms_every'] = 5000
agent_args['test_policy_in_the_beginning'] = True
# experiment arguments
experiment_args = {}
experiment_args['num_train_iterations'] = 820000
experiment_args['test_objective_coeffs_temporal'] = np.array(
[0., 0., 0., 0.5, 0.5, 1.])
experiment_args['test_objective_coeffs_meas'] = np.array([1.])
experiment_args['test_random_prob'] = 0.
experiment_args['test_checkpoint'] = 'checkpoints/2017_04_09_09_07_45'
experiment_args['test_policy_num_steps'] = 2000
experiment_args['show_predictions'] = False
experiment_args['multiplayer'] = False
# Create and run the experiment
experiment = MultiExperiment(
target_maker_args=target_maker_args,
simulator_args=simulator_args,
train_experience_args=train_experience_args,
test_policy_experience_args=test_policy_experience_args,
agent_args=agent_args,
experiment_args=experiment_args)
experiment.run(main_args[0])
def main(main_args):
### Set all arguments
## Target maker
target_maker_args = {}
target_maker_args['future_steps'] = [1, 2, 4, 8, 16, 32]
target_maker_args['meas_to_predict'] = [0, 2]
target_maker_args['min_num_targs'] = 3
target_maker_args['rwrd_schedule_type'] = 'exp'
target_maker_args['gammas'] = []
target_maker_args['invalid_targets_replacement'] = 'nan'
## Simulator
simulator_args = {}
simulator_args['config'] = '../../maps/cartpole.cfg'
simulator_args['resolution'] = (84, 84)
simulator_args['frame_skip'] = 0
simulator_args['env_name'] = 'CartPole-v1'
simulator_args['color_mode'] = 'RGB'
simulator_args['environnement'] = 'gym'
simulator_args['num_meas'] = 4
simulator_args['gym'] = True
#train
simulator_args['num_simulators'] = 8
## Experience
# Train experience
train_experience_args = {}
train_experience_args['memory_capacity'] = 20000
train_experience_args['history_length'] = 3
train_experience_args['history_step'] = 1
train_experience_args['action_format'] = 'enumerate'
train_experience_args['shared'] = False
# Test prediction experience
test_prediction_experience_args = train_experience_args.copy()
test_prediction_experience_args['memory_capacity'] = 1
# Test policy experience
test_policy_experience_args = train_experience_args.copy()
test_policy_experience_args['memory_capacity'] = 3000
## Agent
agent_args = {}
# agent type
agent_args['agent_type'] = 'advantage'
# preprocessing
agent_args['preprocess_input_images'] = lambda x: x / 255. - 0.5
agent_args['preprocess_input_measurements'] = lambda x: x
targ_scale_coeffs = np.expand_dims(
(np.expand_dims(np.array([1, 12.]), 1) * np.ones(
(1, len(target_maker_args['future_steps'])))).flatten(), 0)
agent_args['preprocess_input_targets'] = lambda x: x / targ_scale_coeffs
agent_args['postprocess_predictions'] = lambda x: x * targ_scale_coeffs
agent_args['discrete_controls_manual'] = []
agent_args['opposite_button_pairs'] = [[0, 1]]
# agent properties
agent_args['objective_coeffs_temporal'] = [0.1, 0.1, 0.1, 1, 1, 1]
agent_args['objective_coeffs_meas'] = [
-0.1, -1
] # position and angle pos (-4.8, 4.8) , angle (-24, 24 )
def f1(x):
return x**2
def f2(x):
return np.abs(x)
agent_args['objective_function'] = [f1, f2] # do not deviate from center
agent_args['random_exploration_schedule'] = lambda step: (0.02 + 14500. / (
float(step) + 15000.))
agent_args['new_memories_per_batch'] = 8
agent_args['gym'] = True # in gym only one button is allowed
# net parameters
agent_args['conv_params'] = np.array([(32, 8, 4), (64, 4, 2), (64, 3, 1)],
dtype=[('out_channels', int),
('kernel', int),
('stride', int)])
agent_args['fc_img_params'] = np.array([(512, )],
dtype=[('out_dims', int)])
agent_args['fc_meas_params'] = np.array([(128, ), (128, ), (128, )],
dtype=[('out_dims', int)])
agent_args['fc_joint_params'] = np.array(
[(512, ), (-1, )], dtype=[('out_dims', int)]
) # we put -1 here because it will be automatically replaced when creating the net
agent_args['weight_decay'] = 0.00001
# optimization parameters
agent_args['batch_size'] = 64
agent_args['init_learning_rate'] = 0.0001
agent_args['lr_step_size'] = 250000
agent_args['lr_decay_factor'] = 0.3
agent_args['adam_beta1'] = 0.95
agent_args['adam_epsilon'] = 1e-4
agent_args['optimizer'] = 'Adam'
agent_args['reset_iter_count'] = False
# directories
agent_args['checkpoint_dir'] = 'checkpoints'
agent_args['log_dir'] = 'logs'
agent_args['init_model'] = ''
agent_args['model_name'] = "predictor.model"
agent_args['model_dir'] = time.strftime("%Y_%m_%d_%H_%M_%S")
# logging and testing
agent_args['print_err_every'] = 50
agent_args['detailed_summary_every'] = 1000
agent_args['test_pred_every'] = 0
agent_args['test_policy_every'] = 1000
agent_args['num_batches_per_pred_test'] = 0
agent_args['num_steps_per_policy_test'] = test_policy_experience_args[
'memory_capacity'] / simulator_args['num_simulators']
agent_args['checkpoint_every'] = 10000
agent_args['save_param_histograms_every'] = 5000
agent_args['test_policy_in_the_beginning'] = True
# experiment arguments
experiment_args = {}
experiment_args['num_train_iterations'] = 820000
experiment_args['test_objective_coeffs_temporal'] = np.array(
[0.1, 0.1, 0.1, 1, 1, 1])
experiment_args['test_objective_coeffs_meas'] = np.array([-0.1, -1])
experiment_args['test_random_prob'] = 0.
experiment_args['test_checkpoint'] = 'checkpoints/2020_01_11_23_06_26'
experiment_args['test_policy_num_steps'] = 2000
experiment_args['show_predictions'] = False
experiment_args['multiplayer'] = False
# Create and run the experiment
experiment = MultiExperiment(
target_maker_args=target_maker_args,
simulator_args=simulator_args,
train_experience_args=train_experience_args,
test_policy_experience_args=test_policy_experience_args,
agent_args=agent_args,
experiment_args=experiment_args)
experiment.run(main_args[0])
