curobo.opt.newton.lbfgs module

jit_lbfgs_compute_step_direction(
alpha_buffer: Tensor,
rho_buffer: Tensor,
y_buffer: Tensor,
s_buffer: Tensor,
grad_q: Tensor,
m: int,
epsilon: float,
stable_mode: bool = True,
)
jit_lbfgs_update_buffers(
q,
grad_q,
s_buffer,
y_buffer,
rho_buffer,
x_0,
grad_0,
stable_mode: bool,
)
class LBFGSOptConfig(
d_action: 'int',
action_lows: 'List[float]',
action_highs: 'List[float]',
action_horizon: 'int',
horizon: 'int',
n_iters: 'int',
cold_start_n_iters: 'Union[int, None]',
rollout_fn: 'RolloutBase',
tensor_args: 'TensorDeviceType',
use_cuda_graph: 'bool',
store_debug: 'bool',
debug_info: 'Any',
n_problems: 'int',
num_particles: 'Union[int, None]',
sync_cuda_time: 'bool',
use_coo_sparse: 'bool',
line_search_scale: List[int],
cost_convergence: float,
cost_delta_threshold: float,
fixed_iters: bool,
inner_iters: int,
line_search_type: curobo.opt.newton.newton_base.LineSearchType,
use_cuda_line_search_kernel: bool,
use_cuda_update_best_kernel: bool,
min_iters: int,
step_scale: float,
last_best: float = 0,
use_temporal_smooth: bool = False,
cost_relative_threshold: float = 0.999,
fix_terminal_action: bool = False,
history: int = 10,
epsilon: float = 0.01,
use_cuda_kernel: bool = True,
stable_mode: bool = True,
use_shared_buffers_kernel: bool = True,
)

Bases: NewtonOptConfig

history: int = 10
epsilon: float = 0.01
use_cuda_kernel: bool = True
stable_mode: bool = True
use_shared_buffers_kernel: bool = True
cost_relative_threshold: float = 0.999
static create_data_dict(
data_dict: Dict,
rollout_fn: RolloutBase,
tensor_args: TensorDeviceType = TensorDeviceType(device=device(type='cuda', index=0), dtype=torch.float32, collision_geometry_dtype=torch.float32, collision_gradient_dtype=torch.float32, collision_distance_dtype=torch.float32),
child_dict: Dict | None = None,
)

Helper function to create dictionary from optimizer parameters and rollout class.

Parameters:

  • data_dict – optimizer parameters dictionary.

  • rollout_fn – rollout function.

  • tensor_args – tensor cuda device.

  • child_dict – new dictionary where parameters will be stored.

Returns:

Dictionary with parameters to create a OptimizerConfig

fix_terminal_action: bool = False
last_best: float = 0
use_temporal_smooth: bool = False
line_search_scale: List[int]
cost_convergence: float
cost_delta_threshold: float
fixed_iters: bool
inner_iters: int
line_search_type: LineSearchType
use_cuda_line_search_kernel: bool
use_cuda_update_best_kernel: bool
min_iters: int
step_scale: float
d_action: int

Number of optimization variables per timestep.

action_lows: List[float]

Lower bound for optimization variables.

action_highs: List[float]

Higher bound for optimization variables

action_horizon: int
horizon: int

Number of timesteps in optimization state, total variables = d_action * horizon

n_iters: int

Number of iterations to run optimization

cold_start_n_iters: int | None

Number of iterations to run optimization during the first instance. Setting to None will use n_iters. This parameter is useful in MPC like settings where we need to run many iterations during initialization (cold start) and then run only few iterations (warm start).

rollout_fn: RolloutBase

Rollout function to use for computing cost, given optimization variables.

tensor_args: TensorDeviceType

Tensor device to use for optimization.

use_cuda_graph: bool

Capture optimization iteration in a cuda graph and replay graph instead of eager execution. Enabling this can make optimization 10x faster. But changing control flow, tensor shapes, or problem type is not allowed.

store_debug: bool

Record debugging data such as optimization variables, and cost at every iteration. Enabling this will disable cuda graph.

debug_info: Any

Use this to record additional attributes from rollouts.

n_problems: int

Number of parallel problems to optimize.

num_particles: int | None

Number of particles to use per problem. Common optimization solvers use many particles to optimize a single problem. E.g., MPPI rolls out many parallel samples and computes a weighted mean. In cuRobo, Quasi-Newton solvers use particles to run many line search magnitudes. Total optimization batch size = n_problems * num_particles.

sync_cuda_time: bool

Synchronize device before computing solver time.

use_coo_sparse: bool

Matmul with a Sparse tensor is used to create particles for each problem index to save memory and compute. Some versions of pytorch do not support coo sparse, specifically during torch profile runs. Set this to False to use a standard tensor.

class LBFGSOpt(
config: LBFGSOptConfig | None = None,
)

Bases: NewtonOptBase, LBFGSOptConfig

Base optimization solver class

Parameters:

config – Initialized with parameters from a dataclass.

use_cuda_kernel: bool = True
history: int = 10
reset()

Reset optimizer.

update_nproblems(n_problems)

Update the number of problems that need to be optimized.

Parameters:

n_problems – number of problems.

init_hessian(b=1)
_get_step_direction(
cost,
q,
grad_q,
)

Reimplement this function in derived class. Gradient Descent is implemented here.

_update_q(grad_q)
_update_r(r_init)
_update_buffers(q, grad_q)
_shift(shift_steps=0)

Shift the optimizer by shift_steps * d_opt

Parameters:

shift_steps (int, optional) – _description_. Defaults to 0.

_shift_buffer(
buffer,
shift_d,
shift_steps: int = 1,
)
_call_opt_iters_graph(
q,
grad_q,
)
_compute_cost_gradient(x)
Parameters:

line_search_scale (_type_) – should have n values

_create_opt_iters_graph(
q,
grad_q,
shift_steps,
)
_initialize_opt_iters_graph(
q,
grad_q,
shift_steps,
)
_opt_iters(
q,
grad_q,
shift_steps=0,
)
_opt_step(q, grad_q)
_optimize(
q: Tensor,
shift_steps=0,
n_iters=None,
)

Implement this function in a derived class containing the solver.

Parameters:

  • opt_tensor – Initial value of optimization variables. Shape: [n_problems, action_horizon, d_action]

  • shift_steps – Shift variables along action_horizon. Useful in mpc warm-start setting.

  • n_iters – Override number of iterations to run optimization.

Returns:

Optimized variables in tensor shape [action_horizon, d_action].

_update_best(q, grad_q, cost)
_update_problem_kernel(
n_problems: int,
num_particles: int,
)

Update matrix used to map problem index to number of particles.

Parameters:

  • n_problems – Number of optimization problems.

  • num_particles – Number of particles per problem.

check_convergence(cost)
clip_bounds(x)
cost_relative_threshold: float = 0.999
static create_data_dict(
data_dict: Dict,
rollout_fn: RolloutBase,
tensor_args: TensorDeviceType = TensorDeviceType(device=device(type='cuda', index=0), dtype=torch.float32, collision_geometry_dtype=torch.float32, collision_gradient_dtype=torch.float32, collision_distance_dtype=torch.float32),
child_dict: Dict | None = None,
)

Helper function to create dictionary from optimizer parameters and rollout class.

Parameters:

  • data_dict – optimizer parameters dictionary.

  • rollout_fn – rollout function.

  • tensor_args – tensor cuda device.

  • child_dict – new dictionary where parameters will be stored.

Returns:

Dictionary with parameters to create a OptimizerConfig

epsilon: float = 0.01
fix_terminal_action: bool = False
get_all_rollout_instances() → List[RolloutBase]

Get all instances of Rollout class in the optimizer.

get_nproblem_tensor(x)

This function takes an input tensor of shape (n_problem,….) and converts it into (n_particles,…).

last_best: float = 0
optimize(
opt_tensor: Tensor,
shift_steps=0,
n_iters=None,
) → Tensor

Find a solution through optimization given the initial values for variables.

Parameters:

  • opt_tensor – Initial value of optimization variables. Shape: [n_problems, action_horizon, d_action]

  • shift_steps – Shift variables along action_horizon. Useful in mpc warm-start setting.

  • n_iters – Override number of iterations to run optimization.

Returns:

Optimized values returned as a tensor of shape [n_problems, action_horizon, d_action].

project_bounds(x)
reset_cuda_graph()

Reset CUDA Graphs. This does not work, workaround is to create a new instance.

reset_seed()

Reset seeds.

reset_shape()

Reset any flags in rollout class. Useful to reinitialize tensors for a new shape.

scale_step_direction(dx)
stable_mode: bool = True
update_params(
goal: Goal,
)

Update parameters in the curobo.rollout.rollout_base.RolloutBase instance.

Parameters:

goal – parameters to update rollout instance.

use_shared_buffers_kernel: bool = True
use_temporal_smooth: bool = False
line_search_scale: List[int]
cost_convergence: float
cost_delta_threshold: float
fixed_iters: bool
inner_iters: int
line_search_type: LineSearchType
use_cuda_line_search_kernel: bool
use_cuda_update_best_kernel: bool
min_iters: int
step_scale: float
d_action: int

Number of optimization variables per timestep.

action_lows: List[float]

Lower bound for optimization variables.

action_highs: List[float]

Higher bound for optimization variables

action_horizon: int
horizon: int

Number of timesteps in optimization state, total variables = d_action * horizon

n_iters: int

Number of iterations to run optimization

cold_start_n_iters: int | None

Number of iterations to run optimization during the first instance. Setting to None will use n_iters. This parameter is useful in MPC like settings where we need to run many iterations during initialization (cold start) and then run only few iterations (warm start).

rollout_fn: RolloutBase

Rollout function to use for computing cost, given optimization variables.

tensor_args: TensorDeviceType

Tensor device to use for optimization.

use_cuda_graph: bool

Capture optimization iteration in a cuda graph and replay graph instead of eager execution. Enabling this can make optimization 10x faster. But changing control flow, tensor shapes, or problem type is not allowed.

store_debug: bool

Record debugging data such as optimization variables, and cost at every iteration. Enabling this will disable cuda graph.

debug_info: Any

Use this to record additional attributes from rollouts.

n_problems: int

Number of parallel problems to optimize.

num_particles: int | None

Number of particles to use per problem. Common optimization solvers use many particles to optimize a single problem. E.g., MPPI rolls out many parallel samples and computes a weighted mean. In cuRobo, Quasi-Newton solvers use particles to run many line search magnitudes. Total optimization batch size = n_problems * num_particles.

sync_cuda_time: bool

Synchronize device before computing solver time.

use_coo_sparse: bool

Matmul with a Sparse tensor is used to create particles for each problem index to save memory and compute. Some versions of pytorch do not support coo sparse, specifically during torch profile runs. Set this to False to use a standard tensor.