mirror of
https://github.com/azaion/gps-denied-onboard.git
synced 2026-06-21 06:41:12 +00:00
97f5f9793c
AZ-965 ships the NetVLAD .pt checkpoint that clears the AZ-839
empty-c10_provisioning.backbones SKIP gate. Pipeline-integration
scaffold — encoder is real, NetVLAD tail is honestly labelled as
untrained.
Composition:
* Encoder (26 keys, encoder.0..encoder.28): torchvision
vgg16(weights=IMAGENET1K_V1) features [:-2], BSD-3-Clause.
Real ImageNet-pretrained VGG16 conv stack.
* NetVLAD pool + PCA tail (5 keys: pool.conv.{weight,bias},
pool.centroids, pca.{weight,bias}): random-init via
torch.manual_seed(0). NOT trained for visual place recognition.
Total: 149,002,112 params (568.4 MiB fp32, sha256=745c6f29...).
Round-trip verified locally: torch.load(weights_only=True) +
load_state_dict(strict=True) succeed; forward(1,3,480,480) emits
{'vlad_descriptor': (1, 4096) fp32} — matches NetVladStrategy
contract per net_vlad.py:247-251.
Two material discoveries documented in the AZ-965 spec:
1. The NetVLAD-VGG16 architecture already lives in repo at
src/gps_denied_onboard/components/c2_vpr/_net_vlad_architecture.py
— we instantiate it and save a state_dict, NOT externally source.
2. The PyTorch FP16 runtime expects a .pt state_dict (NOT .onnx).
BackboneConfig.onnx_path is a misnomer for NetVLAD: per AZ-321
design + c2_vpr description.md §1, NetVLAD runs on PyTorch FP16
(NOT TRT). compile_engine is a no-op sha256+path wrap;
deserialize_engine does torch.load(weights_only=True) +
load_state_dict(strict=True).
User skipped Option A/B/C/D/E question — judgment call = Option B
(IMAGENET1K_V1 + random tail) per "use judgment, don't block":
* Option A (Nanne translation) was 5-8 SP, above the 5 SP budget.
* Option B is 3 SP, fits the budget, honestly labelled.
* Option C (pure random) was borderline-dishonest per Real Results.
Files:
* scripts/mk_netvlad_checkpoint.py — deterministic generator.
* models/netvlad/netvlad.pt — 568 MiB, via git-lfs (.gitattributes
extended for models/**/*.pt, *.onnx, *.engine).
* configs/operator_replay.yaml — c2_vpr + c10_provisioning blocks
populated; the field literally named onnx_path actually points
at the .pt for NetVLAD per the runtime semantics noted above.
* docker-compose.test.jetson.yml — ./models:/opt/models:ro bind
mount added to e2e-runner.
* _docs/03_ip_attribution/netvlad.md — provenance, licence, how-to-
reproduce, honest scope statement ("NOT a real-retrieval
checkpoint; ESKF divergence under garbage retrievals is the
expected next gate").
* _docs/02_tasks/todo/AZ-965_netvlad_onnx_backbone_provisioning.md
— rewritten to reflect the .pt-not-.onnx + Option B discoveries.
Tier-2 verification follows in a separate commit after the harness
run confirms the empty-backbones SKIP gate clears.
Out of scope (filed as follow-ups):
* Real-retrieval NetVLAD weights (Nanne Pittsburgh-30k translation
or internal team checkpoint) — separate ticket.
* AZ-840 orchestrator PASSing end-to-end (depends on retrieval
quality + ESKF stability).
* AZ-963 60s smoke ESKF divergence (independent chain).
Co-authored-by: Cursor <cursoragent@cursor.com>
258 lines
12 KiB
YAML
258 lines
12 KiB
YAML
# Tier-2 e2e harness — Jetson Orin Nano (JetPack 6.x, L4T R36.x).
|
|
#
|
|
# AZ-615: companion compose file to `docker-compose.test.yml` that runs
|
|
# the full Reality Gate on a CUDA-capable host. Used via `ssh jetson-e2e
|
|
# "docker compose -f docker-compose.test.jetson.yml up ..."` driven by
|
|
# `scripts/run-tests-jetson.sh`.
|
|
#
|
|
# Difference vs. docker-compose.test.yml:
|
|
# * `runtime: nvidia` + `gpus: all` on `e2e-runner` so the SUT can
|
|
# resolve `model.half().cuda()` against the Orin GPU.
|
|
# * `GPS_DENIED_TIER=2` — turns OFF the auto-skip for `@pytest.mark.tier2`
|
|
# ACs (see tests/conftest.py:31-44). The heavy ACs (AC-1, AC-2, AC-3,
|
|
# AC-5, AC-6) actually run.
|
|
# * Builds from `tests/e2e/Dockerfile.jetson` (l4t-pytorch base).
|
|
# * Companion / db / mock-sat continue to come from the root
|
|
# `docker-compose.yml` via `extends:` (same as Colima) — they have ARM64
|
|
# tags via the existing build pipeline.
|
|
#
|
|
# AZ-688 (sibling of AZ-616): the real satellite-provider .NET service is
|
|
# defined inline below (services.satellite-provider + services.satellite-
|
|
# provider-postgres). `run-tests-jetson.sh` rsyncs `../satellite-provider/`
|
|
# to a sibling directory on the Jetson so the build context resolves
|
|
# identically on the workstation and on the Jetson.
|
|
#
|
|
# Why inline instead of `include: ../satellite-provider/docker-compose.yml`:
|
|
# Compose's `include:` rejects same-name service overrides ("conflicts with
|
|
# imported resource"). We need to customize the api service (healthcheck,
|
|
# network alias, internal-only ports) so the upstream compose's verbatim
|
|
# `include:` doesn't work. Inline is cleaner than the multi-`-f` ordering
|
|
# games required to make overlay precedence work.
|
|
#
|
|
# `mock-sat` remains in the graph for now — AZ-692 retires it once the
|
|
# gps-denied client (AZ-691) lands.
|
|
|
|
services:
|
|
# ------------------------------------------------------------------
|
|
# Init services (profiles: [setup]) — NOT started by the default
|
|
# `docker compose up`. They are invoked explicitly as one-shot jobs
|
|
# via `docker compose run --rm --profile setup <service>` before the
|
|
# main harness run:
|
|
#
|
|
# 1. db-migrate — applies Alembic migrations so companion's
|
|
# FreshnessGate / PostgresFilesystemStore find their tables.
|
|
# (AZ-618 ordering gap: build_pre_constructed queries the DB
|
|
# before the composition root can call apply_migrations.)
|
|
#
|
|
# 2. tile-init — writes a minimal valid HNSW32 FAISS descriptor
|
|
# index into the tile-data volume so FaissDescriptorIndex._load()
|
|
# succeeds during build_pre_constructed.
|
|
# (AZ-618 gap: the production provisioning pipeline normally
|
|
# writes the index; in the test harness it must pre-exist.)
|
|
#
|
|
# They are in profile "setup" so they do not participate in the
|
|
# default `docker compose up` and do not trip --abort-on-container-exit.
|
|
# ------------------------------------------------------------------
|
|
db-migrate:
|
|
profiles: ["setup"]
|
|
image: gps-denied-onboard/e2e-runner:jetson
|
|
entrypoint: ["alembic"]
|
|
command: ["upgrade", "head"]
|
|
working_dir: /opt/project
|
|
volumes:
|
|
- .:/opt/project:ro
|
|
depends_on:
|
|
db:
|
|
condition: service_healthy
|
|
restart: "no"
|
|
|
|
tile-init:
|
|
profiles: ["setup"]
|
|
image: gps-denied-onboard/e2e-runner:jetson
|
|
entrypoint: ["python3"]
|
|
command: ["/opt/project/scripts/mk_test_faiss_fixture.py"]
|
|
volumes:
|
|
- .:/opt/project:ro
|
|
- tile-data:/var/lib/gps-denied/tiles
|
|
restart: "no"
|
|
|
|
# companion and operator-orchestrator are intentionally absent from
|
|
# the Jetson e2e test harness.
|
|
#
|
|
# Every test in tests/e2e/replay/ invokes the ``gps-denied-replay``
|
|
# console-script directly as a subprocess and does not call the
|
|
# companion or operator-orchestrator HTTP APIs. Including either
|
|
# service caused the harness to abort before any test could run:
|
|
#
|
|
# * companion crashes at startup because live-mode requires a
|
|
# production-provisioned C7 inference engine (PyTorch FP16 or
|
|
# TensorRT) that is absent from the test environment. This is the
|
|
# pre-existing AZ-618 gap (build_pre_constructed fails before the
|
|
# composition root can apply_migrations + engine artifacts).
|
|
# * operator-orchestrator crashed for the same C7 inference reason.
|
|
#
|
|
# When the AZ-618 epic ships the full airborne boot-up in a sandboxed
|
|
# environment (Phase E / engine stubs), companion can be re-added here.
|
|
|
|
db:
|
|
extends:
|
|
file: docker-compose.yml
|
|
service: db
|
|
|
|
e2e-runner:
|
|
build:
|
|
context: .
|
|
dockerfile: tests/e2e/Dockerfile.jetson
|
|
image: gps-denied-onboard/e2e-runner:jetson
|
|
# nvidia-container-runtime exposes the Tegra GPU + libcuda mounts.
|
|
# Without this block the container starts but `torch.cuda.is_available()`
|
|
# returns False and every tier2 AC errors at `.cuda()`.
|
|
runtime: nvidia
|
|
deploy:
|
|
resources:
|
|
reservations:
|
|
devices:
|
|
- driver: nvidia
|
|
count: all
|
|
capabilities: [gpu]
|
|
depends_on:
|
|
db:
|
|
condition: service_healthy
|
|
satellite-provider:
|
|
condition: service_healthy
|
|
environment:
|
|
# Same FullSystemConfig env block as Colima — see comments in
|
|
# docker-compose.test.yml for the per-var rationale.
|
|
GPS_DENIED_FC_PROFILE: ardupilot_plane
|
|
# Tier-2 turns OFF the `tier2` / `gpu` auto-skip in tests/conftest.py
|
|
# so the heavy ACs in tests/e2e/replay/test_derkachi_1min.py actually
|
|
# execute. This is the WHOLE POINT of the Jetson harness.
|
|
GPS_DENIED_TIER: "2"
|
|
DB_URL: postgresql://gps_denied:dev@db:5432/gps_denied
|
|
# AZ-777 Phase 1: e2e-runner consumes the real parent-suite
|
|
# satellite-provider .NET service over its compose-DNS name. The
|
|
# dev TLS cert is self-signed against `localhost`, so the suite-
|
|
# internal probe must skip cert verification — see SECURITY note
|
|
# in `.env.test.example`. Production deploys ship a real CA-issued
|
|
# cert and MUST set SATELLITE_PROVIDER_TLS_INSECURE="0" (or omit it).
|
|
SATELLITE_PROVIDER_URL: https://satellite-provider:8080
|
|
SATELLITE_PROVIDER_TLS_INSECURE: "1"
|
|
SATELLITE_PROVIDER_API_KEY: ${SATELLITE_PROVIDER_API_KEY:?SATELLITE_PROVIDER_API_KEY must be set via .env.test — see scripts/mint_dev_jwt.py}
|
|
# AZ-777 Phase 1 also forwards the JWT triple so the smoke test
|
|
# can mint its own dev token in-container as a fallback when
|
|
# SATELLITE_PROVIDER_API_KEY is rotated mid-session.
|
|
JWT_SECRET: ${JWT_SECRET}
|
|
JWT_ISSUER: ${JWT_ISSUER}
|
|
JWT_AUDIENCE: ${JWT_AUDIENCE}
|
|
COMPANION_URL: http://companion:8080
|
|
CAMERA_CALIBRATION_PATH: /opt/tests/fixtures/calibration/adti26.json
|
|
LOG_LEVEL: INFO
|
|
LOG_SINK: console
|
|
INFERENCE_BACKEND: pytorch_fp16
|
|
FDR_PATH: /var/lib/gps-denied/fdr
|
|
TILE_CACHE_PATH: /var/lib/gps-denied/tiles
|
|
MAVLINK_SIGNING_KEY: /opt/tests/fixtures/mavlink_signing/dev_key
|
|
RUN_REPLAY_E2E: "1"
|
|
# Replay-mode build flags (Invariant 9). See identical block in
|
|
# docker-compose.test.yml — all three are required for the
|
|
# composition root to construct the replay strategies. The
|
|
# original harness only set BUILD_REPLAY_SINK_JSONL because
|
|
# every Reality-Gate run died at auto-sync before the other
|
|
# two flags were checked.
|
|
BUILD_VIDEO_FILE_FRAME_SOURCE: "ON"
|
|
BUILD_TLOG_REPLAY_ADAPTER: "ON"
|
|
BUILD_REPLAY_SINK_JSONL: "ON"
|
|
# AZ-894 / AZ-895: the CSV-driven path is now the PRIMARY replay
|
|
# surface (auto-sync was deprecated). `_replay_branch._build_csv_bundle`
|
|
# constructs `CsvReplayFcAdapter`, which fails fast at __init__ when
|
|
# this flag is OFF — every test in tests/e2e/replay/ that runs the
|
|
# `replay_runner` fixture trips that gate without this line.
|
|
BUILD_CSV_REPLAY_ADAPTER: "ON"
|
|
BUILD_FAISS_INDEX: "ON"
|
|
# AZ-964: build_inference_runtime gates pytorch_fp16 behind
|
|
# this flag. The dustynv/l4t-pytorch base image bakes the
|
|
# Tegra-tuned PyTorch wheel, so the strategy module imports
|
|
# cleanly when the flag is ON. build_engine_compiler (called
|
|
# by the AZ-839 fixture) requires c7 inference runtime, so
|
|
# the flag must be ON for the orchestrator test to run.
|
|
BUILD_PYTORCH_FP16_RUNTIME: "ON"
|
|
# AZ-962: the AZ-839 C3 fixture (operator_pre_flight_setup) skips
|
|
# the AZ-840 orchestrator test when this var is missing. The YAML
|
|
# bind-mounted at /opt/configs/operator_replay.yaml declares the
|
|
# four blocks the fixture consumes (c6/c7/c10/c11). c10.backbones
|
|
# is intentionally empty — AZ-964 ships the .onnx + populates it.
|
|
GPS_DENIED_OPERATOR_CONFIG_PATH: /opt/configs/operator_replay.yaml
|
|
volumes:
|
|
- ./tests:/opt/tests:ro
|
|
- ./_docs/00_problem/input_data:/opt/_docs/00_problem/input_data:ro
|
|
- ./configs:/opt/configs:ro
|
|
- ./models:/opt/models:ro
|
|
- fdr-data:/var/lib/gps-denied/fdr
|
|
- tile-data:/var/lib/gps-denied/tiles
|
|
|
|
# AZ-688: real satellite-provider .NET service. Mirrors the upstream
|
|
# compose at ../satellite-provider/docker-compose.yml with three
|
|
# deliberate customizations:
|
|
# * service name = `satellite-provider` (clearer than the upstream's
|
|
# generic `api`) so AZ-692's client uses https://satellite-provider:8080
|
|
# * TCP-level healthcheck via bash /dev/tcp so other services can
|
|
# `depends_on: service_healthy`. The base image
|
|
# (mcr.microsoft.com/dotnet/aspnet:10.0, debian-12-slim) ships
|
|
# bash and /dev/tcp is a bash builtin; no extra package needed.
|
|
# * no host port mappings — internal-only access via compose DNS;
|
|
# keeps host ports free for nested e2e runs.
|
|
satellite-provider:
|
|
build:
|
|
context: ../satellite-provider
|
|
dockerfile: SatelliteProvider.Api/Dockerfile
|
|
image: gps-denied-onboard/satellite-provider:dev
|
|
container_name: gps-denied-e2e-satellite-provider
|
|
environment:
|
|
ASPNETCORE_ENVIRONMENT: Development
|
|
ASPNETCORE_URLS: https://+:8080
|
|
ASPNETCORE_Kestrel__Certificates__Default__Path: /app/certs/api.pfx
|
|
ASPNETCORE_Kestrel__Certificates__Default__Password: satellite-dev-cert
|
|
ConnectionStrings__DefaultConnection: Host=satellite-provider-postgres;Port=5432;Database=satelliteprovider;Username=postgres;Password=postgres
|
|
MapConfig__ApiKey: ${GOOGLE_MAPS_API_KEY:-}
|
|
# Suite JWT contract — see _docs/10_auth.md. Sourced from .env.test
|
|
# via run-tests-jetson.sh; the API fails fast at startup if any of
|
|
# the three are missing or whitespace-only.
|
|
JWT_SECRET: ${JWT_SECRET:?JWT_SECRET must be set via .env.test}
|
|
JWT_ISSUER: ${JWT_ISSUER:?JWT_ISSUER must be set via .env.test}
|
|
JWT_AUDIENCE: ${JWT_AUDIENCE:?JWT_AUDIENCE must be set via .env.test}
|
|
volumes:
|
|
- ../satellite-provider/certs/api.pfx:/app/certs/api.pfx:ro
|
|
- ../satellite-provider/tiles:/app/tiles
|
|
- ../satellite-provider/ready:/app/ready
|
|
- ../satellite-provider/logs:/app/logs
|
|
healthcheck:
|
|
test: ["CMD", "bash", "-c", "exec 3<>/dev/tcp/127.0.0.1/8080"]
|
|
interval: 5s
|
|
timeout: 3s
|
|
retries: 12
|
|
start_period: 30s
|
|
depends_on:
|
|
satellite-provider-postgres:
|
|
condition: service_healthy
|
|
|
|
satellite-provider-postgres:
|
|
image: postgres:16
|
|
container_name: gps-denied-e2e-satellite-provider-postgres
|
|
environment:
|
|
POSTGRES_USER: postgres
|
|
POSTGRES_PASSWORD: postgres
|
|
POSTGRES_DB: satelliteprovider
|
|
volumes:
|
|
- satellite-provider-postgres-data:/var/lib/postgresql/data
|
|
healthcheck:
|
|
test: ["CMD-SHELL", "pg_isready -U postgres"]
|
|
interval: 5s
|
|
timeout: 5s
|
|
retries: 5
|
|
|
|
volumes:
|
|
db-data: {}
|
|
fdr-data: {}
|
|
tile-data: {}
|
|
satellite-provider-postgres-data: {}
|