Language

unilab.tasks.motion_tracking.common.tensor_sampler¶

Device-resident motion sampling for tensor Manager owners.

This owner owns the cold metadata derivation and hot tensor lifecycle: adaptive failure statistics, frame advance, reset sampling, and sampling diagnostics stay on one Torch device and consume the Manager-owned Torch generator.

Classes

`MotionSampler`	Torch-native adaptive/mixed motion sampler for `MotionCommand`.
`MotionSamplerDiagnostics`	Count observed compatibility paths that would defeat tensor-only resets.

class unilab.tasks.motion_tracking.common.tensor_sampler.MotionSampler[source]¶

Bases: object

Torch-native adaptive/mixed motion sampler for MotionCommand.

Parameters:

mode (str)
num_envs (int)
num_frames (int)
clip_offsets (Tensor)
clip_end_frames (Tensor)
bin_count (int)
adaptive_lambda (float)
adaptive_kernel_size (int)
adaptive_uniform_ratio (float)
adaptive_alpha (float)
start_ratio (float)
initial_frames (Tensor)
initial_clip_end_frames (Tensor)
device (device)

__init__(*, mode, num_envs, num_frames, clip_offsets, clip_end_frames, bin_count, adaptive_lambda, adaptive_kernel_size, adaptive_uniform_ratio, adaptive_alpha, start_ratio, initial_frames, initial_clip_end_frames, device)[source]¶

Parameters:

mode (str)
num_envs (int)
num_frames (int)
clip_offsets (Tensor)
clip_end_frames (Tensor)
bin_count (int)
adaptive_lambda (float)
adaptive_kernel_size (int)
adaptive_uniform_ratio (float)
adaptive_alpha (float)
start_ratio (float)
initial_frames (Tensor)
initial_clip_end_frames (Tensor)
device (device)

property bin_failed_count: Tensor¶

sample_frames(reset_mask, torch_rng)[source]¶

Sample full-width frames and publish only selected rows.

Parameters:

reset_mask (Tensor)
torch_rng (Generator | None)

Return type:

Tensor

update_failure_stats(terminated)[source]¶

Parameters:: terminated (Tensor)
Return type:: None

step_full(active, time_steps)[source]¶

Advance and publish full-width frame mirrors without row gathers.

Parameters:

active (Tensor)
time_steps (Tensor)

Return type:

Tensor

class unilab.tasks.motion_tracking.common.tensor_sampler.MotionSamplerDiagnostics[source]¶

Bases: object

Count observed compatibility paths that would defeat tensor-only resets.

The sampler itself no longer transfers rows to the host. We retain a real (not synthesized) counter for reset sampling calls so diagnostics can assert that reset dispatch happened while host-transfer counts remain structurally zero.

Parameters:

reset_dispatches (int)
reset_host_row_transfers (int)
step_host_mirror_transfers (int)

reset_dispatches: int = 0¶

reset_host_row_transfers: int = 0¶

step_host_mirror_transfers: int = 0¶

record_reset_dispatch()[source]¶

Return type:: None

property total: int¶

__init__(reset_dispatches=0, reset_host_row_transfers=0, step_host_mirror_transfers=0)¶

Parameters:

reset_dispatches (int)
reset_host_row_transfers (int)
step_host_mirror_transfers (int)