Public FeatDepth CPU/CUDA inference evidence¶
The external FeatDepth runner now supports FP32 batches and bounded parallel PNG
loading, and reads/hashes the checkpoint once from the same bytes that PyTorch
decodes. Its checkpoint reference is reused in the input/provider manifests.
--device cuda already existed; the new batch/prefetch options and checkpoint
reuse improve that execution path. CPU remains the default. The large CPU/GPU
ratios below describe moving depth inference to GPU, rather than the incremental
effect of batching alone.
The measured implementation commit is
0b67865a52fbc17651ee02e96905a83bb54e416e. Runner SHA256:
df13e2f39b9711a4112cc0d830b52b8fa2181f8e8028a5b97526c2306ae4b552.
Actual measured bytes match the committed LF source. The pinned external
FeatDepth commit is 550420b3fb51a027549716b74c6fbce41651d3a5; checkpoint SHA256
is 559f168c762d01af7c745b32ced8930a12423a0e9b62af791f7d894bd902af27.
Images, weights, LiDAR scans and generated NumPy arrays remain external.
Public GPU model labels and hardware-specific output-directory names are anonymized. Published file bindings in the evidence manifest were updated; original input, result and measurement-launcher digests still identify the original run bytes. Numerical values and timings are unchanged. Archived validation launchers therefore have separate original and published digests.
Workload and execution¶
KITTI raw 2011_09_30_drive_0018_sync/image_02, the same 25 uniformly selected
frames as the public six-DoF demo, 1226 x 370 source images, 1024 x 320 network
inputs. Raw inverse-depth/disparity is returned to each original image grid.
The i7-9750H / NVIDIA CUDA GPU 6 GiB / 32 GiB Windows host was shared with unrelated
user work. Environment: Python 3.12.10, NumPy 2.3.4, Pillow 12.3.0,
PyTorch 2.11.0+cu128, CUDA runtime 12.8. GPU FP32 settings retain the existing
cuDNN defaults, including benchmark=false. Training overlap with KITTI raw
is recorded as possible; no online refinement is performed.
python tools/benchmark_featdepth_provider.py `
E:/datasets/calibrex_camera_lidar_v06/kitti_raw/2011_09_30/2011_09_30_drive_0018_sync `
--featdepth-repository E:/datasets/calibrex_camera_lidar_v06/providers/FeatDepth-git `
--checkpoint E:/datasets/calibrex_camera_lidar_v06/providers/checkpoints/fm_depth.pth `
--output-dir outputs/public-kitti-featdepth-batching-v3 --cpu-threads 2 --repetitions 3
All nine subprocesses are fresh, with no separate warmup. Order is CPU, serial CUDA, batched CUDA on odd repetitions and reversed on even repetitions. OS file caches persist across runs. Every image/checkpoint/depth reference is verified after each subprocess. Startup, model setup, inference, depth export, primary manifest and optional timing-manifest writes are included in subprocess wall; parent verification/comparison and benchmark aggregation are excluded.
| Repetition | CPU, 2 threads (s) | Serial CUDA (s) | CUDA, batch 2 + 2 PNG workers (s) |
|---|---|---|---|
| 1 | 117.347018 | 13.176117 | 12.458591 |
| 2 | 123.940307 | 13.372894 | 13.078349 |
| 3 | 132.531623 | 22.416443 | 18.169170 |
| Median | 123.940307 | 13.372894 | 13.078349 |
The CPU-2-thread / batched-CUDA wall median ratio is 9.48x. Within the already accelerated CUDA path, the wall median ratio is only 1.02x here. Inference-plus-export medians are 113.640095 / 2.730953 / 2.345925 s for CPU / serial CUDA / batched CUDA: 48.44x CPU-to-batched-CUDA and 1.16x serial-to-batched-CUDA. Short clips are dominated by startup/model preparation. The prediction-batch timer is a subset of inference/export and must not be added again. Main-entry timing excludes interpreter/import startup and the timing manifest write; subprocess wall includes both.
A separate one-repetition execution omits --cpu-threads, giving PyTorch's
default six CPU threads on this host. It measures 89.481593 / 18.528891 /
17.824680 s for CPU / serial CUDA / batched CUDA, an observed 5.02x
CPU-to-batched-CUDA ratio. See default_cpu.definition.json and
default_cpu.benchmark.json. This single pair has different concurrent load;
do not mix it into the three-repetition median or infer CPU-thread scaling.
Three repeated batched-CUDA runs have identical per-frame depth-file hashes. Compared with the same-repetition CPU output, all predictions are finite; maximum absolute disparity difference is 6.139279e-6, maximum relative difference 3.037236e-5 (denominator floor 1e-12), and pixel-weighted RMSE 1.436477e-7. Those differences do not alone establish calibration equality.
The batched run records peak Torch GPU allocated/reserved memory of 675,414,528 / 1,027,604,480 bytes, versus serial CUDA 406,979,072 / 599,785,472 bytes. These process counters exclude host memory and other GPU users. At batch 2/workers 2, at most six image tasks are queued alongside the current two-image batch. Decode buffers, outputs and model memory are additional. Checkpoint loading retains one approximately 630 MiB byte buffer alongside deserialized weights, then releases it.
Six fixed initializations¶
validation.definition.json and validation.benchmark.json record twenty-four
fresh six-DoF estimates: two-thread CPU, default six-thread CPU, serial CUDA and
batched CUDA depth for each frozen x+/x-/y+/y-/z+/z- problem. Initial rotation and
translation perturbations remain +/-0.5 degrees and +/-0.5 m on the same axis.
Bounds remain +/-2 degrees / +/-1 m and the evaluation cap is 800. Twenty fit
frames resolve histogram ranges; the same five holdout frames never enter
fitting or range selection. The D2D solver/backend is unchanged.
All twenty-four estimates converge. Every GPU-depth estimate has exactly the same final translation/quaternion components and evaluation count as the same-version CPU-depth estimate. Full numerical payloads are not exactly equal: learned disparity differences change some histogram counts, objective/trace scores and resolved ranges. Final holdout MI differs only in y-positive, by 6.2001394e-7 nat for both GPU-depth modes. Final poses and reference errors remain identical across the four depth-generation modes in all six cases.
The initial eighteen trials compare two-thread CPU with both GPU modes. Six default-six-thread CPU-depth estimates were added after detecting thread-count differences in saved depth bytes. Their maximum absolute disparity difference from two-thread CPU is 1.192093e-7, RMSE 1.820104e-10. Their complete calibration numerical payloads match the two-thread CPU anchor in all six cases. Priors, split, bounds and stopping rules are unchanged; no model or holdout tuning was performed for this supplementary check.
| Case | Evaluations | Rotation reference error (deg) | Translation reference error (m) | CPU final holdout MI (nat) | GPU holdout MI absolute difference (nat) |
|---|---|---|---|---|---|
| x-positive | 229 | 0.390923 | 0.160078 | 1.293494542 | 0 |
| x-negative | 313 | 0.410874 | 0.104583 | 1.287191168 | 0 |
| y-positive | 277 | 0.527379 | 0.167705 | 1.294100526 | 6.2001394e-7 |
| y-negative | 241 | 0.266233 | 0.091001 | 1.281462561 | 0 |
| z-positive | 265 | 0.258640 | 0.087052 | 1.283677673 | 0 |
| z-negative | 241 | 0.285129 | 0.089486 | 1.281674508 | 0 |
Prior provider artifacts used another PyTorch version, so the comparison anchor is newly generated CPU FP32 depth from the same environment as the GPU modes. The six historical priors are preserved exactly in their transform fields. Reference values are never optimized against. The y-positive reference rotation error remains greater than its initial 0.5 degrees. Convergence and unchanged poses do not establish calibration accuracy: covariance/observability are unknown, scans are not motion compensated, and training overlap is possible. All twenty-four full result traces and their problem templates are published.
Published artifacts and provenance¶
inference.definition.json and inference.benchmark.json retain every timing,
command, execution configuration, source/provider/timing-file digest and map
difference. First-repetition provider/timing manifests retain all 25 input/output
references, environment versions, checkpoint identity and execution metadata.
Published paths use /external/calibrex-public-inputs and /workspace/Calibrex.
Parsed path strings and embedded JSON trial notes are normalized; numerical
payloads, original source/output SHA256 values and timings are retained.
publication.depth-provider.json is a schema-valid manifest containing original
and published byte hashes for every evidence artifact. Trial output hashes bind
original local files, not their normalized publications. Published problems are
evidence templates: restore external paths and bind the chosen provider's actual
file hash before executing them. External arrays/checkpoints are not embedded.
The manual validation and publication launchers are archived byte-for-byte as
validation-launcher.py.txt, default-cpu-validation-launcher.py.txt and
publication-launcher.py.txt, with their SHA256
values in the publication manifest. They use the frozen local source-problem
directories and new inference outputs named above; restore those inputs and use
new/empty output directories when reproducing the experiment. The reusable
inference benchmark tool is tracked under tools/.
These observations describe this workload and shared host. Repeated frames do not provide independent accuracy trials, confidence intervals or portable latency guarantees. See the usage guide for GPU/CPU commands and timing scope.