def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['conditions']
self._cond_idx = range(self.M)
self._hyperparams['train_conditions'] = self._cond_idx
self._hyperparams['test_conditions'] = self._cond_idx
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
if self._hyperparams['fit_dynamics']:
dynamics = self._hyperparams['dynamics']
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
if self._hyperparams['fit_dynamics']:
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], self._cond_idx[m]
)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt']
)
if type(hyperparams['cost']) == list:
self.cost = [
hyperparams['cost'][i]['type'](hyperparams['cost'][i])
for i in range(self.M)
]
else:
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
self.mpc = [] # For initialize
def __init__(self, hyperparams, cond):
config = deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.M = self._hyperparams['M'] = config['init_mpc']['T']
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
del self._hyperparams['agent'] # Don't want to pickle this
self.N = int(ceil(self.T / (self.M - 1.)))
# It will update wwith different X_t in update function
# Note that is different from M in last n-th MPC
self.T_mpc = self.M
# Setup policy
init_mpc = config['init_mpc']
init_mpc['x0'] = agent.x0
init_mpc['dX'] = agent.dX
init_mpc['dU'] = agent.dU
self.mpc_pol = []
for n in range(self.N):
init_mpc = extract_condition(config['init_mpc'], cond)
self.mpc_pol.append(init_mpc['type'](init_mpc))
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['conditions']
self._cond_idx = range(self.M)
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
self.traj_distr = {self.iteration_count: []}
self.traj_info = {self.iteration_count: []}
self.kl_div = {self.iteration_count:[]}
self.dists_to_target = {self.iteration_count:[]}
self.sample_list = {i: SampleList([]) for i in range(self.M)}
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
dynamics = self._hyperparams['dynamics']
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], self._cond_idx[m]
)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_distr[self.iteration_count].append(self.cur[m].traj_distr)
self.traj_info[self.iteration_count].append(self.cur[m].traj_info)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt']
)
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
if self._hyperparams['ioc']:
self.gt_cost = [
hyperparams['gt_cost']['type'](hyperparams['gt_cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['conditions']
self._cond_idx = range(self.M)
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
self.traj_distr = {self.iteration_count: []}
self.traj_info = {self.iteration_count: []}
self.kl_div = {self.iteration_count: []}
self.dists_to_target = {self.iteration_count: []}
self.sample_list = {i: SampleList([]) for i in range(self.M)}
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
dynamics = self._hyperparams['dynamics']
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], self._cond_idx[m])
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_distr[self.iteration_count].append(
self.cur[m].traj_distr)
self.traj_info[self.iteration_count].append(self.cur[m].traj_info)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt'])
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
if self._hyperparams['ioc']:
self.gt_cost = [
hyperparams['gt_cost']['type'](hyperparams['gt_cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
self.timers = OrderedDict()
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['tac'][
'clusters'] if 'tac' in hyperparams else hyperparams[
'conditions']
self._cond_idx = range(self.M)
self._hyperparams['train_conditions'] = self._cond_idx
self._hyperparams['test_conditions'] = self._cond_idx
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
#self.dX = self._hyperparams['agent']['dtgtX'] + self.dX
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
self.new_mu = [None] * self.M
self.new_sigma = [None] * self.M
dynamics = self._hyperparams['dynamics']
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
if dynamics is not None:
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'],
self._cond_idx[0] # TODO Global x0
)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
#self.traj_opt = hyperparams['traj_opt']['type'](
# hyperparams['traj_opt']
#)
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
self.M = hyperparams['conditions']
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
dynamics = self._hyperparams['dynamics']
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], m
)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt']
)
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
self.M = hyperparams['conditions']
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
dynamics = self._hyperparams['dynamics']
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], m)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt'])
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def reset_alg(self):
"""
reset the algorithm to initial state at the beginning.
reset the traj_distr and traj_info but keep policy_opt
"""
last_pol = list()
for m in range(self.M):
last_pol.append(copy.deepcopy(self.cur[m].last_pol))
self.cur = [IterationData() for _ in range(self.M)]
if self._hyperparams['fit_dynamics']:
dynamics = self._hyperparams['dynamics']
for m in range(self.M):
self.cur[m].last_pol = copy.deepcopy(last_pol[m])
self.cur[m].traj_info = TrajectoryInfo()
if self._hyperparams['fit_dynamics']:
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], self._cond_idx[m]
)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
self.traj_opt = self._hyperparams['traj_opt']['type'](
self._hyperparams['traj_opt']
)
if type(self._hyperparams['cost']) == list:
self.cost = [
self._hyperparams['cost'][i]['type'](self._hyperparams['cost'][i])
for i in range(self.M)
]
else:
self.cost = [
self._hyperparams['cost']['type'](self._hyperparams['cost'])
for _ in range(self.M)
]
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
# False
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['conditions'] # 2
self._cond_idx = range(self.M) # [0, 1]
self._hyperparams['train_conditions'] = self._cond_idx
self._hyperparams['test_conditions'] = self._cond_idx
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
# class IterationData(BundleType):
# """ Collection of iteration variables. """
# def __init__(self):
# variables = {
# 'sample_list': None, # List of samples for the current iteration.
# 'traj_info': None, # Current TrajectoryInfo object.
# 'pol_info': None, # Current PolicyInfo object.
# 'traj_distr': None, # Initial trajectory distribution.
# 'new_traj_distr': None, # Updated trajectory distribution.
# 'cs': None, # Sample costs of the current iteration.
# 'step_mult': 1.0, # KL step multiplier for the current iteration.
# 'eta': 1.0, # Dual variable used in LQR backward pass.
# }
# BundleType.__init__(self, variables)
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
# False
if self._hyperparams['fit_dynamics']:
dynamics = self._hyperparams['dynamics']
for m in range(self.M):
# class TrajectoryInfo(BundleType):
# """ Collection of trajectory-related variables. """
# def __init__(self):
# variables = {
# 'dynamics': None, # Dynamics object for the current iteration.
# 'x0mu': None, # Mean for the initial state, used by the dynamics.
# 'x0sigma': None, # Covariance for the initial state distribution.
# 'cc': None, # Cost estimate constant term.
# 'cv': None, # Cost estimate vector term.
# 'Cm': None, # Cost estimate matrix term.
# 'last_kl_step': float('inf'), # KL step of the previous iteration.
# }
# BundleType.__init__(self, variables)
self.cur[m].traj_info = TrajectoryInfo()
# False
if self._hyperparams['fit_dynamics']:
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
# algorithm['init_traj_distr'] = {
# 'type': init_lqr,
# 'init_gains': 1.0 / PR2_GAINS,
# 'init_acc': np.zeros(SENSOR_DIMS[ACTION]),
# 'init_var': 1.0,
# 'stiffness': 0.5,
# 'stiffness_vel': 0.25,
# 'final_weight': 50,
# 'dt': agent['dt'],
# 'T': agent['T'],
# }
# self._cond_idx = [0, 1]
init_traj_distr = extract_condition(
self._hyperparams['init_traj_distr'], self._cond_idx[m])
# def extract_condition(hyperparams, m):
# """
# Pull the relevant hyperparameters corresponding to the specified
# condition, and return a new hyperparameter dictionary.
# """
# return {var: val[m] if isinstance(val, list) else val
# for var, val in hyperparams.items()}
# If list, extract index m, otherwise extract whole value.
# Save to traj_distr (Initial trajectory distribution)
self.cur[m].traj_distr = init_traj_distr['type'](init_traj_distr)
# algorithm['traj_opt'] = 'type': TrajOptLQRPython
# self.traj_opt = TrajOptLQRPython(hyperparams)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt'])
# algorithm['cost'] = {
# 'type': CostSum,
# 'costs': [torque_cost, fk_cost1, fk_cost2],
# 'weights': [1.0, 1.0, 1.0],
# }
if type(hyperparams['cost']) == list: # False
self.cost = [
hyperparams['cost'][i]['type'](hyperparams['cost'][i])
for i in range(self.M)
]
else: # self.cost = [CostSum(cost_param), CostSum(cost_param)]
self.cost = [
hyperparams['cost']['type'](hyperparams['cost'])
for _ in range(self.M)
]
# 'kl_step': 5.0
self.base_kl_step = self._hyperparams['kl_step']
def __init__(self, hyperparams):
config = copy.deepcopy(ALG)
config.update(hyperparams)
self._hyperparams = config
if 'train_conditions' in hyperparams:
self._cond_idx = hyperparams['train_conditions']
self.M = len(self._cond_idx)
else:
self.M = hyperparams['conditions']
self._cond_idx = range(self.M)
self._hyperparams['train_conditions'] = self._cond_idx
self._hyperparams['test_conditions'] = self._cond_idx
self.iteration_count = 0
# Grab a few values from the agent.
agent = self._hyperparams['agent']
#print(agent)
self.agent = agent
self.T = self._hyperparams['T'] = agent.T
self.dU = self._hyperparams['dU'] = agent.dU
self.dX = self._hyperparams['dX'] = agent.dX
self.dO = self._hyperparams['dO'] = agent.dO
init_traj_distr = config['init_traj_distr']
init_traj_distr['x0'] = agent.x0
init_traj_distr['dX'] = agent.dX
init_traj_distr['dU'] = agent.dU
del self._hyperparams['agent'] # Don't want to pickle this.
# IterationData objects for each condition.
self.cur = [IterationData() for _ in range(self.M)]
self.prev = [IterationData() for _ in range(self.M)]
dynamics = self._hyperparams['dynamics']
for m in range(self.M):
self.cur[m].traj_info = TrajectoryInfo()
self.cur[m].traj_info.dynamics = dynamics['type'](dynamics)
cur_init_traj_distr = extract_condition(init_traj_distr,
self._cond_idx[m])
cur_init_traj_distr['cur_cond_idx'] = self._cond_idx[m]
#print(cur_init_traj_distr)
self.cur[m].traj_distr = cur_init_traj_distr['type'](
cur_init_traj_distr, agent)
self.traj_opt = hyperparams['traj_opt']['type'](
hyperparams['traj_opt'])
self.cost = []
for m in range(self.M):
cost_hyperparams = hyperparams['cost'].copy()
cost_hyperparams['cur_cond_idx'] = self._cond_idx[m]
self.cost.append(hyperparams['cost']['type'](cost_hyperparams))
self.base_kl_step = self._hyperparams['kl_step']
policy_prior = self._hyperparams['policy_prior']
for m in range(self.M):
self.cur[m].pol_info = PolicyInfo(self._hyperparams)
self.cur[m].pol_info.policy_prior = \
policy_prior['type'](policy_prior)
self.policy_opt = self._hyperparams['policy_opt']['type'](
self._hyperparams['policy_opt'], self.dO, self.dU)
