A Darknet/YOLO model exported by hank-ai's darknet_onnx_export (v5.1+) has
two outputs: confs [B,N,nc], already objectness x class probability, and
boxes [B,N,1,4], corners x1,y1,x2,y2 normalized to the input. It trains and
predicts on a stretched square and has no NMS in the graph. ONNXYOLO took
the 4-D boxes tensor for segmentation proto masks and letterboxed the input,
so such a model decoded as garbage; a sealed package named as YOLOV4 went to
YOLOOD, which cannot open an .ansmodel at all.
- ONNXYOLO / RTYOLO: recognise the layout once at load (package family
"darknet", output names confs/boxes, or their shapes), before the seg
check; stretch-resize the input (single, batch, and RTYOLO's NV12 path);
new postprocessDarknet: best class per candidate, threshold, stretch map
to the original image, clip, class-aware NMS.
- dllmain: a sealed package passed as modelType 1 (YOLOV4) is routed as
detection (3), i.e. to RTYOLO on NVIDIA and ONNXYOLO elsewhere, in
CreateANSODHandle, CreateANSODHandleEx and OptimizeModelStr.
- Every other model is unchanged: all new branches are gated on the layout.
Verified on macOS only: ANSONNXYOLO.cpp passes a syntax check, and the new
decode functions, extracted into a standalone harness running the model
through onnxruntime, reproduce Darknet's own detections on dog.jpg. NOT yet
built on Windows, and the TensorRT path has not run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The gateway's product table already runs to 1009 ANSMIND and 1010 ANSGW.
The licensing library stopped at 1008, so a real ANSMIND licence could not
be named, deactivated or listed here, while every product name it did not
recognise quietly became product 1009. With 1009 now a real product, that
fallback would mint an ANSMIND trial for a typo.
ANSPRODUCTS gains ANSMIND = 1009 and ANSGW = 1010, and both are added
wherever the product table is spelled out: ActivateTrialLicense,
DeactivateLicense, ActivateLicense, ActivateLicenseWithCustomHWID and
ListLicenses, which now reports an ANSMIND and an ANSGW row like the
gateway does.
Unknown names no longer map to any id. ActivateTrialLicense returns 0, its
existing failure value, instead of generating a key. ListLicenses skips
the row. ActivateLicense and ActivateLicenseWithCustomHWID keep passing
the name through as the registration name: they never assigned an id, the
key's own validation data rejects a name it was not issued for, and the
gateway resolves an unknown product the same way.
The ANSVIS master-licence rule in LicenseVerification is untouched.
Separately, the fallback loops in ANSUtilities, ANSAWSS3 and ANSLLM asked
for product 1008 under the registration name "ANSSRT". The product is
ANSTS, so a genuine ANSTS licence could never satisfy them. Only that
string changes; the loops still stop at 1008.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A .ansmodel reaching the GPU path went to ModelOptimizer, which knows one
way to open a model — unzip it — and failed on the first byte. Even past
that, the TensorRT engine only ever parsed an ONNX file, so the one thing
the format exists to prevent, a decrypted model on disk, was the only way
it could have run.
The builder now takes its bytes from memory when it has them. The path it
is given still names the plan that gets cached beside it, because a plan
is a built artefact and carries no model: the second run loads it and
never needs the ONNX at all. For a package the plan lands in the folder
the zip flow used, so a model converted from a zip reuses the plan that
zip already built.
OptimizeModelStr recognises a package and skips the unzip-and-optimise
leg entirely: the handle below it opens the package in memory and builds
the plan from the decrypted bytes, which is the same work by a shorter
road.
Still on the file path: the other TensorRT handlers (segmentation, pose,
classification). Only the YOLO detector reads a package today.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every test in that file names a model and a video from whichever
workstation it was written on, so anywhere else the harness prints
"Model zip file is not exist" and exits before it reaches the GPU. That
is what happened on the machine that is meant to be testing CUDA and
TensorRT.
Given a model and a video on the command line it runs those instead,
with an optional engine type and detection type. With no arguments it
behaves exactly as before, so nobody's workflow changes.
The model may be a .zip or a sealed .ansmodel, because the engine now
opens either — and that is the only way to watch a sealed model run on
a GPU, which is the last thing in this format that has never been seen
working end to end.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The package library's Windows branch pointed at vcpkg alone. On a real
Windows build machine the header lives in the ANSLibs tree, so anything
linking this library failed with "Cannot open include file:
'nlohmann/json.hpp'" while the engines compiled fine — they add ANSLibs
to their own include path and never noticed.
Now it is searched for by name, with the ANSLibs tree, vcpkg and the
usual system paths as hints, and the configure says where it found it.
One code path for every platform instead of an assumption per platform.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every .cu file failed on CUDA 13 with "nvcc fatal: Unknown option
'--use-fast-math'". The flag is --use_fast_math, with underscores; older
toolkits accepted the hyphenated form as well and this one does not, so
the engine layer could not compile a single kernel on a current CUDA.
Found on the Windows machine with CUDA 13.0, where it stopped the build
at the first two kernels.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Line two set CMP0167 unconditionally. That policy arrived in CMake 3.30
and this project asks for 3.28, so every CMake between the two failed
the configure at line two with "Policy CMP0167 is not known to this
version" — before it had done anything at all. Found on the Windows
machine that builds the CUDA engines, where it stopped the build dead.
Guarded, so a newer CMake still gets the quiet FindBoost the policy was
set for and an older one configures instead of refusing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This engine took the feature count from outputShape[1] unconditionally.
For a yolo26 export with end2end on — [1, 300, 6], rows of
(x1 y1 x2 y2 score class) with the suppression already inside the graph
— that means 296 classes, coordinates read as class scores, and boxes
covering the frame. No error, just nonsense, which is exactly what the
Labelling Tool produced from the same model before its decoder was
fixed.
Its sibling ANSONNXYOLO has recognised that head from the tensor shape
for a while; this one had not, and the backend marker cannot tell them
apart because it names the runtime rather than the head.
The branch fires only where the tensor cannot be a grid: the 6 must be
the last axis with a detection count before it, hundreds rather than the
thousands of anchors a two-class [1, 6, 8400] export carries.
The decoder reuses this engine's own scaleCoords rather than repeating
the letterbox arithmetic, so both heads undo the padding the same way
and agree about where a box is, and the rounding into the integral box
type is spelled out instead of left to a narrowing conversion.
NOT COMPILED. The engine layer needs CUDA and the ANSLibs tree, neither
of which exists on this machine, so this is verified by reading: the
shape rule matches the one the other three products now share, every
symbol it uses is declared in this translation unit, and the arithmetic
is delegated to the function the grid path already trusts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Whether the model-package key is present was an __has_include test
inside the header, and that test leaves no dependency for the build
system. A key emitted after a build had already compiled the file would
not take effect until someone cleaned, and the resulting engine looks
exactly like one built without a key — which on a release machine means
shipping an engine that silently cannot open a package.
CMake now decides it and says which case it is, so emitting the key and
re-configuring is what installs it, visibly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A package whose family says "yolo" may hold an end-to-end export. Boxes
took the family from the backend marker, and the marker names the
runtime, not the head — so a fast yolo26 export arrived declared as
plain yolo while its graph emits [1, N, 6] with the NMS inside it.
Decoding that as a [1, 4+nc, N] grid reads coordinates as confidences:
detections everywhere, and nothing that looks like a failure.
OpenModelPackage now reads the first output's dims from the model it has
just decrypted, and corrects the family with the reason in the warning
rather than swapping it silently. It costs nothing there: the bytes are
in hand. The legacy marker is left alone, so anything converting back to
a zip still writes what ANSCORE expects.
A two-class v8 export is [1, 6, 8400] and must not be mistaken for one,
which is why the 6 has to be the last axis with a detection count before
it. Both cases are tested, and the package this Mac exported before the
writer was fixed now reads as yolo10 here too. 46 checks, no failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Only cmake-build-*/ and out/ were ignored, so configuring into build/ —
the directory the ANSCORE_DEBUGVIEW_STRIP usage note in CMakeLists.txt
shows — left an untracked build/ in git status. build-mac/ is untracked
as well and stays that way; this pattern matches build/ alone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
tests/CMakeLists.txt added every test directory unconditionally, so a
machine with no CUDA toolkit — every macOS one — died at generate time on
"links to CUDA::cudart but the target was not found", taking the
cross-platform targets down with it.
Gate each test the way its subject is already gated: the five that need
the toolkit behind ANSCORE_HAS_CUDA, the ANSIO one behind WIN32, and the
ANSONVIF one behind both. ANSONVIF itself uses no CUDA; its test names
CUDA::cudart and CUDA::nvjpeg only because it drives capture through
ANSCV and was copied from the ANSCV test.
Two non-test targets had the same problem. ANSCV genuinely needs the
toolkit — ANSGpuFrameOps.h includes <cuda_runtime.h> unconditionally and
calls cudaMemcpy2D/cudaMallocPitch — so it moves into the existing CUDA
block beside the engine modules. ANSLibsLoader binds cudart at runtime
and ships its own stubs, but still compiles against the runtime header
for the types and links CUDA::cudart_static, so it is gated too.
Also call enable_testing(), which the project never did: without it the
add_test() under tests/ reached nothing and ctest found no tests. Plain
enable_testing(), not include(CTest), to avoid its extra targets.
A CUDA machine is unchanged: every gate holds there, ANSCV keeps its
place in the subdirectory order, and the only new output is
CTestTestfile.cmake.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The reader has been able to open packages for a while and has held no
key, so it refused all of them. This is the seam that hands it one.
The base private key is not in this repository and never will be. It is
written by the ceremony (ANSMINDCORE's ansmodel-keys) into a generated
header that is gitignored and produced from the key vault on the machine
that builds a release. ANSModelKeys.h uses that file when it is there
and says "no keys" when it is not, so a developer's checkout builds and
runs with no secret in it: every zip still loads, and a package is
refused with a sentence that says why.
Keys are settled once — what the build carries, plus a developer's
environment — unless a caller sets a keyring outright, which now means
exactly that rather than depending on which call ran first.
Checked with the real keys against the packages ANSMINDCORE's side
produced: a package sealed by a box that holds only the public half
opens here byte for byte, and one signed by the new release key verifies
against the built-in root. If those two products disagreed, a model that
runs in one would not run in the other, which is the failure this format
exists to avoid. 40 checks in the package test, no failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A package now carries what its model expects on the input, but every
package written before that does not — and neither does the zip, which
is most of what is in the field. The engine was then left to assume
640x640, which is wrong for a classification export at 224 and for
anything trained at a rectangular size.
The model itself always knows. ANSOnnxShape reads the first input's
shape straight out of the ONNX protobuf, from memory or from a file,
walking the top level and SEEKING past the nodes and the initializers
rather than reading them — the fields it wants sit after all of the
weight on the wire, and a 324 MB model answers in ten milliseconds.
OpenModelPackage uses it only when the manifest is silent, which costs
nothing there because the bytes are already decrypted in hand. An axis
the graph leaves open is reported as zero and a file that cannot be
parsed is a failure: both mean "this model does not say", and the caller
keeps the default it had. Being wrong about an input size would be worse
than being silent about it.
Checked against python-onnx on all 46 models on this machine, static and
dynamic, four-dimensional and not, with no disagreements. The test
writes ONNX graphs byte by byte, with node and initializer ahead of the
input as a real export has them, and packs one into a package whose
manifest says nothing — which is the case this exists for. 40 checks,
no failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An engine's model has always been a path to a password-protected zip:
ANSODBase inflates it into a folder, and every engine then picks out
train_last.onnx, classes.names and model_config.json. A .ansmodel
package replaces all four with one sealed file, and this is the part
that turns it back into what the engines already understand.
ANSODBase::LoadModel and ::Initialize now recognise a .ansmodel path and
take a second route: the package is opened with the process keyring, the
model is decrypted into memory, and the classes, task and input geometry
come out of the manifest instead of loose files. Nothing is extracted,
so the unprotected model never touches the disk — which is the reason
the format exists, and it is one function's job to keep.
The engines follow it down. BasicOrtHandler gains a constructor that
takes the bytes rather than a path; it keeps a view of them, not a copy,
because the weights are in RAM either way once a session exists and a
copy would only double what every loaded model costs. The owner of the
plaintext is ANSODBase, which holds it for as long as the engine object
lives — that is what lets the DirectML recovery path rebuild a session
after a driver fault — and wipes it on teardown.
Wired through ANSONNXYOLO (and so the classification facade), ANSONNXOBB,
ANSONNXPOSE, ANSONNXSEG and YOLO12OD. YOLOOD is left alone: it loads
darknet weights and OpenCV DNN graphs, neither of which the writer
produces. A zip still loads exactly as before, on every engine.
Keys come from SetProcessModelKeys, as in ANSMINDCORE, with the same
development fallback from the environment. With none set, a package is
refused with a sentence saying so and every zip still works, so a build
made before the key ceremony runs everything except packages.
The engine layer needs CUDA to compile, so the contract it depends on is
checked where it can be: ANSModelPackage-UnitTest now opens the fixture
ANSMINDCORE wrote through the same glue ANSODBase uses, and covers the
refusal a keyless build gives. 32 checks, no failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The second reader. A package is one encrypted file — cleartext header,
sealed manifest, model bytes sealed in chunks — and this opens it: the
content key unwrapped with X25519 against the base key, the header
verified as Ed25519 signed by a release key the root vouches for, the
model decrypted into memory and handed to ANSModelSource, which the
engines already take.
It is a new library rather than code inside the engine layer, and
deliberately free of the heavy dependencies around it: no CUDA, no
OpenCV, no LabVIEW. That is what lets it build and be tested on every
platform, including the ones where the engine layer does not build.
THE TWO IMPLEMENTATIONS MUST AGREE, so this does not merely claim to:
tests/ANSModelPackage-UnitTest/data/interop.ansmodel was written by
ANSMINDCORE's packer, and the test opens it here — manifest, classes and
a chunked payload, byte for byte. A change to either wire format shows
up as this failing. The fixture's keys are derived by published rules,
recorded beside it, so anyone can rebuild it.
The AEAD and the small primitives are copies of ANSMINDCORE's, not
reimplementations, for the same reason. Monocypher 4.0.2 is vendored
here as it is there, including the RFC 8032 pair — its default EdDSA
signs with BLAKE2b and would not interoperate.
Verified by compiling and running the test directly: 19 checks, RFC
8439, RFC 7748 and RFC 8032 vectors, the interop fixture, a round trip,
a flipped bit, an unknown root, a revoked signer, and rubbish bytes.
ANSCORE's own CMake configure fails on this machine for reasons that
predate this change — eight errors with or without it, all missing
dependencies — so the new CMake targets still need a real build box.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every engine built its session from a file path, which is why opening an
encrypted model package means inflating it to a directory first: the
unprotected model lands on disk, is optimised, and is zipped back into a
cache. The new package format cannot avoid that from the outside — the
engines have to be able to take bytes.
ANSModelSource.h introduces a ModelSource (a path, or bytes with a name
for logs) and one MakeSession that builds from either. The six ONNX
engines — classification, OBB, pose, segmentation, YOLO12 and YOLO —
each gain a load overload taking it, with the existing string form
delegating so no caller changes. The Windows wide-path conversion and
the from-memory call now live in one place instead of the eleven copies
they were spread across; YOLOOD, which had only the Windows branch,
becomes portable as a side effect.
The factories are named rather than implicit, so a call with a plain
string literal still resolves to the path overload without ambiguity.
Not compiled here: this layer builds only where CUDA is present, so
verification on this machine was the new header compiled and run, plus a
structural check that every engine declares and defines both overloads,
calls MakeSession exactly once, and has no raw session construction
left. It needs a build on a CUDA box before it ships.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The legacy display remap in FaceAttributeToJsonString rewrote most
emotions to "Attentive", which hid the v1.4 classifier's real classes.
The remap now applies ONLY to the legacy OpenVINO model's lowercase
classes (sad/happy/neutral/anger/surprise - v1.1 output unchanged);
every other name, i.e. the classes read from emotionsrecognition.names
(Angry, Fear, Happy, Neutral, Sad), passes through verbatim. A retrained
classifier with a different .names list needs no code change.
Verified on the classroom video: faces previously labelled "Attentive"
now read "Angry (100%)"; smiling faces read "Happy (100%)". Parity test
updated to check each pack against its own emotion vocabulary.
Note for client code: v1.4 deployments no longer emit "Attentive" -
anything matching that string must handle Angry/Fear/Sad instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
With the heavier v1.4 ONNX age-gender model, interval 20 benchmarks at
throughput parity with the v1.1 pack (16.8 ms / 59 fps vs 16.0 ms / 62
fps on RTX 3070, classroom.mp4, ~3.1 faces/frame) versus 19.9 ms / 50
fps at the old default of 5. Attributes refresh every ~0.8 s on a 25 fps
camera (1.6 s on stable scenes, faster when faces appear), which is
ample for age/gender/emotion; head pose and recognition still run every
frame. Deployments wanting fresher attributes can lower it at runtime
via SetANSRFAttributeInterval. Full test suite passes at the new
default; GetANSRFModelInfo reports "attribute_interval":20.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New API SetANSRFAttributeInterval(Handle, frames) / UpdateAttributeInterval
(1..60, default 5) controls how often the slow attributes (age/gender +
emotions) refresh per camera; head pose and recognition are unaffected.
GetANSRFModelInfo now reports "attribute_interval". The adaptive logic's
stable-scene branch previously capped at a hardcoded 10 - it now doubles
the configured interval instead (default behaviour 10/5/2 unchanged).
Motivation: the v1.4 ViT age-gender model is heavier per attribute frame
than the old OpenVINO IR. Interval sweep on RTX 3070 (classroom.mp4,
~3.1 faces/frame) vs the v1.1 baseline at 16.0 ms / 62 fps:
interval 5 (default): 19.9 ms / 50 fps (+3.9 ms)
interval 10: 17.1 ms / 58 fps (+1.1 ms)
interval 20: 16.5 ms / 61 fps (parity)
Recommendation: default on 8 GB GPUs; interval 10 on small-VRAM machines.
ANSFR-UnitTest: --bench old|new [interval] runs the sweep.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The re-packed ANS_FaceRecognizer_v1.4.zip (ONNX attribute models only, no
OpenVINO agegender/emotion IRs) is now the official pack and lives in the
canonical distribution folder C:\Projects\SharedCore next to v1.1.
Centralize every hardcoded pack path into three file-scope constants
(kFaceRecognizerV14Zip, kFaceRecognizerV11Zip, kFaceDetectorZip) used by
the model-switch test, classroom demo, benchmark, and parity test - a
future pack bump is a one-line edit. Verified against the official path:
adaface + ONNX attribute backends load and bench at 17.9 ms avg. The
slimmer pack was separately validated from a clean extraction (all
TestModelSwitchV14 checks pass without the OpenVINO fallback files).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
SetANSRFRecognizerModel now records the choice in the face database
(SETTINGS key RECOGNIZER_MODEL_PREF), and LoadEngine restores it when no
explicit choice was made in the current session - so a runtime switch to
ArcFace (or AdaFace) survives server restarts without re-calling the API.
Kept separate from EMBEDDING_MODEL on purpose: only an actual API call is
recorded as a preference. An automatic ArcFace fallback (pre-v1.4 zip
without adafacerecognizer.onnx) is not persisted, so upgrading such a
deployment to the v1.4 pack still gets the AdaFace default.
ANSFR-UnitTest: TestModelSwitchV14 (now also runnable via --test) gains a
simulated restart - release handle, recreate on the same database, and
assert the persisted ArcFace choice is restored and still recognizes the
enrolled user. Verified ALL PASSED.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
TestOutputFormatParity (--parity) runs the identical enroll + inference
flow through the v1.1 pack (arcface + OpenVINO attributes) and the v1.4
pack (adaface + ONNX attributes) and asserts the emitted JSON is
format-identical: same result-object fields, same attribute keys
(Gender/Age/AgeYears/Emotion/Pose), same value vocabularies (Male/Female,
the four Age buckets, Happy/Neutral/Attentive emotion labels with (NN%)),
and that both packs recognize the enrolled user. Verified ALL PASSED on
the RTX 3070 test machine.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- ClassroomAttributesVideoTest (default main): plays classroom.mp4 through
the v1.4 pipeline and overlays user name, Gender/Age, and Emotion parsed
from each result's extra_info JSON; saves an annotated frame every 30
frames to C:\Programs\DemoAssets for headless review.
- --bench old|new|both: per-pack speed benchmark (100 measured frames after
10 warmup, full pipeline with attributes on) run one pack per process.
RTX 3070 laptop results: v1.1 arcface+OpenVINO 17.8 ms avg (56.2 fps) vs
v1.4 adaface+ONNX/TRT 18.2 ms avg (55.1 fps) at 3.1 faces/frame — the
new models carry no throughput cost.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Recognizer: adafacerecognizer.onnx (AdaFace, new default) with automatic
fallback to ansfacerecognizer.onnx (ArcFace) for pre-v1.4 zips. AdaFace
consumes BGR: new AdaFace handler in ONNXEngine skips the BGR2RGB swap,
and the TensorRT paths gate their cvtColor on m_swapRB.
- New C API: SetANSRFRecognizerModel (0=AdaFace, 1=ArcFace),
GetANSRFRecognizerModel, GetANSRFModelInfo. A live switch reloads the
recognizer and regenerates all stored embeddings from the enrolled face
crops; the generating model is persisted in a new SETTINGS table
(EMBEDDING_MODEL) so upgrading an ArcFace database to the AdaFace
default migrates automatically on LoadEngine.
- Attributes: prefer agegender.onnx (ViT: gender_logits [male,female] +
age_pred in years; also handles the single-tensor research export) and
emotionsrecognition.onnx (YOLOv11-cls, classes from .names) behind new
IAgeGenderEstimator/IEmotionsClassifier interfaces, falling back to the
legacy OpenVINO IRs. EP follows hardware: NVIDIA uses TensorRT EP with
disk-cached engines, AMD DirectML, Intel OpenVINO.
- Emotion JSON mapping: handle the new Fear class and the never-matched
surprise label (both map to Attentive).
- ANSFR-UnitTest: TestModelSwitchV14 covers default model selection, ONNX
attribute backends, enrollment, and adaface<->arcface switching.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Comments out the unconditional `return 1` added in b5a88b9, so the function
runs its real body again: the argument guards (null Handle / zero-length
buffer -> -1), cv::imdecode, and ANSFacialRecognition::CheckFace with the
detector confidence threshold, isValidFace's geometry gate and the
_minFaceSize filter.
The bypass line is left in place commented rather than deleted, so it stays
easy to flip back on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Repoints the hardcoded fixtures at the installed ANSVIS Server layout rather
than the old E:/C: DemoAssets trees:
- classroom.mp4 now read from C:\ProgramData\ANSCENTER\Shared
- TestCompleteFR1 loads the GPU-specific ServerOptimised zips
(ANS_FaceRecognizer_v1.1_NVIDIAGeForceRTX4070LaptopGPU.zip) and re-enables
the InsertUser call ahead of InsertFace
- main() runs ANSVISTest()
Also comments out the --gpu argument handling in main(), so the harness no
longer sets CUDA_VISIBLE_DEVICES and the single-GPU bisect flag is inert until
that block is restored.
Paths remain machine-specific; this is a developer harness, not a test suite.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CheckFaceEmbedding returns 1 unconditionally, ahead of the JPEG decode and the
CheckFace() call, so every caller is told the submitted image is acceptable.
Deliberate temporary bypass, not a fix. What stops running behind it:
- argument guards -- a null Handle or a zero-length buffer used to return -1
and now returns 1
- cv::imdecode, so an undecodable JPEG is no longer rejected
- ANSFacialRecognition::CheckFace, and with it the detector confidence
threshold, isValidFace's geometry gate (5 landmarks inside the box, roll
<= 27 deg, yaw asymmetry <= 0.5, pitch 0.15-0.75, mouth/eye ratio 0.6-1.6)
and the _minFaceSize filter
Enrollment callers that relied on this as a pre-check now accept any input.
Revert this commit to restore the checks.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
hwinfo::getAllGPUs() is an OpenCL device list -- the prebuilt HWiNFO.lib has no
Win32_VideoController path, only clGetPlatformIDs/clGetDeviceIDs/clGetDeviceInfo.
A machine whose driver never registered an OpenCL ICD therefore reported no GPU
at all, even with a healthy WDDM driver driving the display. Two consequences:
the system-info JSON emitted "GPU":"", and CheckHardwareInformation() silently
fell back to the CPU engine on a box holding a perfectly good CUDA card.
DXGI needs only a WDDM display driver, so it sees the adapter either way.
CheckHardwareInformation() now classifies adapters by PCI vendor id. Exact ids
also retire the substring matching and its "ATI" inside "CORPORATION" workaround.
It falls back to the OpenCL list when DXGI returns nothing.
GetSystemHardwareInformation() keeps its OpenCL loop untouched and merely appends
DXGI adapters that OpenCL did not already report, so the JSON stays byte-identical
-- same keys, order, indices and values -- on any machine that already worked.
Appended entries reuse OpenCL's vendor spelling (OpenClStyleVendor) and read the
driver version from the display-adapter class key, since DXGI exposes neither.
Verified on a dual-adapter box: the GPU node matches the pre-change output
exactly, and forcing the append path yields the same shape with values sourced
from DXGI plus the registry.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
.claude/settings.local.json is machine-specific permission config; it
was tracked by accident and showed as perpetually modified on every
machine. Untracked (kept on disk) and ignored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Enrolling a face while the recognizer engine is down used to store the
user row and silently drop the embedding: InsertUser is a pure DB write,
Feature() returns empty when uninitialized, and InsertFace's size check
misreported that as "Invalid embedding generated" — blaming the face.
The result surfaces hours later as every match returning Unknown/2.0
with nothing in the log saying why (seen on the A400: 45 users, zero
embeddings).
- ANSFRBase gains a public IsInitialized() accessor (flag existed but
was protected).
- InsertFace / InsertMultipleFaces gate on it before computing the
embedding and log the true cause: recognizer not initialized, face
NOT enrolled, fix engine errors then re-enroll.
- The residual invalid-embedding message now reports actual vs expected
size and the remaining causes (unusable crop / recognizer reloading).
- Reload() warns when the database has users but zero embeddings — the
one line that identifies stale half-enrolled databases at startup.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A ServerOptimised zip is a per-GPU cache of a previous engine build. When
the engine profile changes (batch gate, TRT/CUDA upgrade), the cached
engine inside it goes stale: every re-extract (e.g. after Temp cleanup)
restores the old engine and costs a full ONNX rebuild, forever, because
the server only regenerates the zip when the file is missing.
Engine now records rebuiltFromOnnx() + engineFilePath() in
buildLoadNetwork (set only on the generate-from-ONNX branch; cache hits
and alternative-precision loads don't raise it). After a rebuild,
ANSFaceRecognizer::Initialize deletes the stale same-GPU engine variants
from the model folder and the supplied optimised zip, so the server's
next optimize pass recreates it with the fresh engine.
Safety: only zips whose parent folder is ServerOptimised/-ized qualify —
base model zips can never match. Engine purge is scoped to the
"<model>.engine.<GPU>." prefix, leaving other GPUs' engines intact on
multi-GPU machines. All deletes are best-effort (error_code); multi-GPU
pool paths never set the flag, so the purge is a no-op there.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Batch-4 was tuned for 8 GB GPUs; its TRT workspace (~54 MiB per batch
unit) crowds out the other engines on smaller cards. VramCappedBatch
keeps the configured batch on 8 GB-class cards and up (threshold
7.5 GiB — 8 GB cards report ~8188 MiB) and drops to single-face batches
below that: identical results, multi-face frames just run sequentially,
~160 MiB freed on a 4 GB RTX A400. Also caps the legacy 8/32 LoadEngine
profile, whose .b32 engine (~1.7 GB workspace) would be fatal on small
cards. Initialize() and OptimizeModel() share the same call so their
profiles cannot drift.
Small cards rebuild the recognizer engine once (filename loses the .b4
suffix); 8 GB+ machines keep their cached engines unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Field logs showed clusters of uploads taking exactly 20004/20007/20003 ms —
the default _idleTimeoutMs — across several workers at the same instant, then
recurring. Cause: a connection whose request failed was returned to the pool,
so a socket silently dropped by a middlebox was handed to worker after worker,
each paying the full idle timeout to rediscover it. Chilkat's autoReconnect
only covers a cleanly closed connection, not a half-open one.
Connection handling:
- DiscardConnection(): a TRANSPORT failure closes the connection instead of
pooling it (23 sites). An HTTP error status is not a transport failure - the
server answered - so those connections are still reused.
- Pooled connections idle longer than 15 s are dropped at acquire time.
- First request on a reused connection uses a short idle timeout (a quarter of
the configured one, 5 s by default); a fresh connection keeps the full one.
A dead socket now costs ~5 s once, not 20 s repeatedly.
- One upload call is capped at 30 s wall-clock. The attempt count alone was not
a time bound: 20 attempts x 20 s could block the caller for ~400 s.
- RetryLoop no longer gives up after one failed connect. Callers call Connect()
once at startup, so returning left the handle dead for the process lifetime.
Now backs off 3 s -> 30 s until connected or stopped.
- Connections carry the config generation they were built against, so a connect
completing just after SetAuthentication cleared the pool cannot leave an
unauthenticated connection behind (those uploads would 403).
- Stale per-request state is cleared on acquire: multipart uploadId/partNumber
query params, and the gzip Content-Encoding set by UploadTextData. Not
ClearAllHeaders - that wipes SigV4 state (fcc985e).
Return codes - upload exports now say whether a retry is worthwhile:
-20 permanent (bad argument, licence, or HTTP 4xx), -21 not connected yet,
-22 transient. Recorded per thread, since one handle is shared by many workers.
Success stays the only positive value, so rc == 1 and rc != 1 are unchanged.
Handle and exception codes (-1/-3/-4) made consistent across all upload
wrappers; some previously returned 0 for these.
Also fixes a misclassification this surfaced: Chilkat reports failure for an
HTTP 4xx delivered during an Expect: 100-continue exchange, so NoSuchBucket was
treated as a network error and retried 20 times before reporting "transient".
The failure path now checks the response status. Chilkat's ~2 KB error dump per
failure is compressed to one line carrying the actual S3 error code.
Verified against live S3: 14,701 uploads across 15 threads with a 150 s idle
gap, 0 failures, no call over 5 s (avg 90 ms, p95 136 ms), matching the
pre-change baseline; and all six return-code classes confirmed end to end.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The id is built from up to 6 enumerated MACs. On Linux getifaddrs() also
returns virtual interfaces — docker0 and the veth* pairs each container
creates, tun*/wg* from a VPN, br-*/virbr* bridges. Container veths carry
RANDOM MACs and appear/disappear with the containers, so the enumerated set
can differ between calls: the hardware id changes and an activated licence
silently stops matching its machine. Boxes running Docker (the Coder feature)
would hit this routinely.
This matters more on Linux than elsewhere because two of the four id
components are already constant there: the system-UUID slot is a literal and
GetHDDPropertyList is an empty function on non-Windows. The id rests on CPU id
+ MACs, so unstable MACs are most of the instability.
Filter mirrors what the Windows branch already does via WMI (PNPDeviceId LIKE
%PCI% — physical adapters only). The test is /sys/class/net/<name>/device,
the kernel's own record of a hardware-backed interface, so it stays correct as
new virtual-interface flavours appear; a name-prefix denylist is kept only for
environments without /sys. MACs are bucketed during the walk and chosen after
it, so a veth appearing earlier in the list cannot consume a slot and push a
real NIC out.
FAIL-SAFE: if no physical NIC is found, fall back to the unfiltered set. An
empty list would drop all MAC entropy and — with the UUID and disk slots
already constant — leave near-identical ids across different machines.
Windows and macOS/BSD paths are untouched (both #else branches preserve the
existing behaviour exactly). Logic verified against a simulated Docker host:
enp3s0/wlp2s0 accepted; lo, docker0, veth*, br-*, tun0 rejected.
ORDERING: this CHANGES the id on any Linux box that has virtual interfaces, so
it must land BEFORE the first Linux licences are issued — afterwards it would
invalidate them.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GetCurrentHardwareId sized its bit stream at 125 bits (16 bytes) but
writes 147 (3 version + 9 x 16-bit hashes), so the last two hash slots
overflowed the heap block by 3 bytes on every id computation. The
Windows heap absorbed it silently for years; it SIGSEGVs under macOS
malloc — found when ANSGateway enabled this code on POSIX (fixed there
in ANSGateway 3a34c68/6c3cdfd; this is the upstream port).
Two parts, both id-preserving (only the first 16 bytes feed the base32
encode — verified byte-identical ids before/after in the gateway):
- hwid.cpp: size the stream to the 147 bits actually written.
- BitStream3: bounds-safe byteAt/setByte accessors on every Write/Read,
Attach records the real buffer size, default ctor zero-inits — an
under-sized stream can no longer corrupt the heap, and out-of-range
reads return 0 (identical to the base32 padding zeros the Match path
always read there).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- hwid.cpp: BitStream3 buffer 125 -> 160 bits (writes totalled 147 bits
and overflowed the old allocation; encoded HWID bytes unchanged).
Log component availability (uuid/cpu/mac/hdd) once per distinct HWID,
and warn when WMI yields no components at all.
- Matcher widened only, generation untouched: exact-string match
short-circuits, UUID match accepts any WMI list entry (generator
hashes the last valid one), fuzzy-match point breakdown is logged so
field status-12 rejections are diagnosable.
- ANSLicense.cpp: fix inverted delete condition (!p != NULL) at 6 sites;
detect the shared "empty fingerprint" HWID of WMI-broken PCs and warn
with repair steps; mask activation keys in error strings.
- licensingclient.h: explain SetKey/validation/NTP failures on stderr.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The margin was a fixed 768 MiB — 3% of a 24 GiB compute card but 19% of
a 4 GiB card (RTX A400), where free VRAM never clears it, so CanLoad()
refused every engine and UnderPressure() skipped every frame. Now
total/8 clamped to [192, 768] MiB: cards >= 6 GiB keep exactly the old
margin, only small cards change. ANSCENTER_GPU_MARGIN_MB / SetMargin()
still override (0 = auto). Refusal log now reports total VRAM alongside
free, which is what made this diagnosable.
Verified on the A400 test PC: detector + recognizer load and match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Overnight soak revealed the gap: ANSRTSPClient does not auto-reconnect
by design (the application layer polls staleness and calls
ReconnectRTSP - ANSVIS/LabVIEW does this in production), but the unit
test never did, so when the laptop's Modern Standby killed the TCP
session at ~1 AM the client returned the cached frame until morning.
Poll GetRTSPLastFrameAgeMs each loop; at 20 s without a decoded frame,
call ReconnectRTSP (timestamped console line), retrying at most every
15 s. Threshold sits above the corruption keyframe gate's ~12 s
worst-case hold so a reconnect never preempts an in-flight recovery.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
After a sequence gap the H.264/H.265 depacketizers resumed reassembly at
the next NAL/FU start. On multi-slice streams every slice starts a NAL,
so reassembly resumed MID-FRAME and delivered access units missing their
leading slices — manufacturing "First slice in a frame missing" /
"Ignoring POC change between slices" / "Could not find ref" decoder
errors and green/blocky concealment output far beyond the actual loss.
Resume only at a true AU boundary instead:
H.265 - non-VCL NAL (type >= 32) or VCL slice with
first_slice_segment_in_pic_flag set (first bit after the
2-byte NAL header)
H.264 - non-VCL NAL (type >= 6) or VCL slice with
first_mb_in_slice == 0 (ue(v): MSB set after 1-byte header)
Also: stderr visibility for RTP sequence gaps (rtp_rx logs went only to
the sys_log file) and for av_log-bridge latches in video_decoder — the
global av_log level filters decoder errors from the console, which made
keyframe-gate arms look causeless during diagnosis.
Live verification, HEVC camera on lossy Wi-Fi (NVDEC decode):
before: 271 decoder errors / 3 min, delivery collapsed to ~2 fps
after: 3 errors / 2 min, full ~21 fps, zero corrupt frames displayed
Remaining rare errors show POC ref misses with ZERO RTP sequence gaps,
i.e. pre-decode queue eviction under TCP burst clumps - a separate,
much smaller issue absorbed invisibly by the keyframe gate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
log_print writes to the sys_log file (unopened in unit tests / field
debugging), so the gate looked inert next to FFmpeg's stderr error lines.
Emit arm/release one-liners to stderr alongside them.
Live verification during a severe Wi-Fi loss burst (271 HEVC errors /
3 min, NVDEC path): gate armed 19x, released 17x, dropping 29-118
corrupt frames per burst and resuming on the next clean keyframe.
The av_log bridge (273dd44) is confirmed working end-to-end.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The corruption re-arm added in 6344062 keyed off AVFrame::decode_error_flags,
but the HEVC decoder on the hwaccel (NVDEC) path never sets it: 120 s of live
Wi-Fi packet loss produced 10+ "First slice in a frame missing" / "Could not
find ref with POC n" log bursts with decode_error_flags == 0 on every frame,
so the keyframe gate never fired.
Bridge: install a process-wide av_log callback (chains to
av_log_default_callback) that maps the logging AVCodecContext back to its
CVideoDecoder and latches a per-decoder m_bCorruptSeen flag on AV_LOG_ERROR
messages. readFrame() consumes the latch alongside decode_error_flags to
re-arm the keyframe gate. With hwaccel the decoder is single-threaded, so
the logging context is exactly our AVCodecContext; SW frame-threaded paths
log from per-thread clones but set decode_error_flags instead, which the
existing check covers.
Registered after avcodec_open2, unregistered in uninit() before the context
is freed. Map lookup only on error-level messages (rare). Also promote the
gate-arm message to log_print(HT_LOG_WARN) for console visibility.
Clean-path verified live: 6105 frames / 5 min NVDEC RTSP with zero errors
and no regressions. Loss-path verification pending a naturally lossy window.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Restores the previously shelved decoder/player hardening and extends it:
1. Startup keyframe gating (video_decoder): after (re)start the decoder
joins the stream mid-GOP; frames referencing unseen pictures render as
green/blocky concealment garbage. Drop everything until the first IDR
(300-frame safety valve so unmarked streams can't blank forever).
2. NEW - corruption re-arm: on lossy links (Wi-Fi cameras) dropped slices
surface as "First slice in a frame missing" / "Could not find ref with
POC n" and FFmpeg marks the output via AVFrame::decode_error_flags.
Re-arm the keyframe gate and drop the rest of the GOP so the display
holds the last clean frame instead of showing concealment artifacts;
corrupt drops count toward the safety valve.
3. libyuv NV12->BGR conversion (video_player): honors NVDEC/D3D11VA padded
strides (1080->1088 rows) that cv::cvtColorTwoPlane mishandled (green
cast + striped noise), picks BT.601/709/2020 coefficients from the
frame's signalled colorspace, and is ~3-5x faster than sws_scale.
Verified live against an RTSP HEVC camera (NVDEC path): steady frame flow,
no regressions on the clean-stream path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The empty-frame path slept 1 s without calling cv::waitKey, so the
HighGUI window stopped pumping Windows messages and went "Not
Responding" whenever the RTSP stream stalled for a few seconds.
Wait via cv::waitKey(1000) instead (ESC still exits). Same fix in
all three test loops that shared the pattern.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two-pass slot acquisition in runInferenceFromPool: Pass A picks an idle
slot whose deviceIndex matches the input GpuMats' resident device
(zero-copy dispatch); Pass B falls back to plain round-robin, where
runInference migrates the inputs via the pinned-host two-hop copy.
Callers that prepare frames on GPU A no longer pay the ~1 ms/frame PCIe
migration when a same-GPU slot is idle, while spillover to the other GPU
under load still works. Input residency is queried outside the slot
mutex via cudaPointerGetAttributes (cheap driver lookup).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Verified on dual RTX 5080: 355/355 detections, both GPUs in round-robin,
zero CUDA errors. Requires TensorRT >= 10.16.1 in SharedCore (10.15.1 has
a Myelin l2cm crash when creating a context on a second Blackwell device).
1. blobFromGpuMats: thread_local GpuMat caches (tl_blob/tl_floatImg) were
device-blind — a pool thread warming GPU[1] then GPU[0] reused device-1
memory with a device-0 stream, causing an async illegal memory access
that poisoned every CUDA context (sticky 700). Track owning device and
rebuild the cache on device switch.
2. runInference: pool round-robin can dispatch to a slot on a different
GPU than the one the caller prepared inputs on. Detect foreign input
GpuMats via cudaPointerGetAttributes and migrate them through a pinned
host staging buffer (two-hop D2H+H2D; direct cross-device cudaMemcpy2D
is unsupported and GeForce has no P2P).
3. warmUp: now pins its own CUDA device before uploading dummy inputs and
returns success/failure; loadNetwork rejects engines whose warmup fails
so broken slots can never join the pool silently.
4. Blackwell builder workaround: setMaxAuxStreams(0) on SM 12.x plus an
.aux0 engine-cache suffix so pre-workaround caches rebuild.
Also: ANSFR-UnitTest gains --gpu N flag (sets CUDA_VISIBLE_DEVICES) for
single-GPU bisection, and rich diagnostics on input-copy failures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add ConsoleTest.cpp, a headless (no-GUI) smoke test for ANSOPENCV image
ops plus an optional RTSP capture path, built as its own executable
(ANSCV-ConsoleTest) so it does not collide with ANSCV-UnitTest's GUI main.
Make the ans_imshow.hpp include (shipped with ANSMIND, outside the repo)
configurable via the ANSMIND_INCLUDE_DIR cache variable, defaulting to
C:/ProgramData/ANSCENTER/ANSMIND. Warn rather than fail configure when the
header is absent so unrelated targets still build.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1. core/ANSLicensingSystem/ANSLicense.h:268 — added default initializers to ModelConfig POD members (inpWidth = 0, inpHeight = 0, etc.). This was the actual bug: CreateANSODHandle declares a local ModelConfig and only sets a few fields, leaving inpWidth/inpHeight as indeterminate stack memory. On your machine that memory consistently held ~80, slipping past the <=0 → 640 guard and driving the ORT session at 80×80 — which crashes YOLOv11's /model.12/Concat because 2·ceil(80/32)=6 ≠ ceil(80/16)=5.
2. modules/ANSODEngine/ANSONNXYOLO.h:75 — ONNXYOLO::setInputShape() now snaps width/height up to the next multiple of 32 (with a 64-pixel floor). Defense in depth: even if a future caller hands in a non-stride-32 size, the FPN concat alignment is preserved.
3. modules/ANSODEngine/ANSONNXYOLO.cpp — two diagnostic additions:
WarmUpEngine bails with a clear log line if the warm-up dims aren't a multiple of 32, instead of letting the model.12 error fire and mask the real cause.
transform() logs the input shape, canvas shape, source size, and dynamic flag once per session in DebugView — so any future shape-mismatch crash points at the offending HxW directly.