| name | deepdetect-pytorch-worker |
| description | Use when porting external PyTorch object detection models in DeepDetect through the external PyTorch worker backend. |
External PyTorch Model Import Skill
Use this guide when an LLM agent needs to port, import, adapt, or test an
external PyTorch object detection model in DeepDetect through the external
PyTorch worker backend, the external-pytorch-detector CLI profile,
mllib.entrypoint or service_mllib.entrypoint, connector tensor pull,
generated extern/pytorch_workers/<model_slug>/ adapters, or generic detection
worker hooks.
Ground The Work
Read the local implementation before editing. Start with the PyTorch worker
backend docs, the external worker README, the CLI profile code, and the current
detection worker base. Use rg for:
DetectionTrainingWorkerBase
external-pytorch-detector
service_mllib
mllib.entrypoint
connector_tensor_pull
test_predictions
Inspect the target upstream repository before designing the adapter. Identify
its package requirements, model factory, YAML or config system, training loss
path, postprocessor, checkpoint format, device handling, distributed assumptions,
download or pretrained defaults, and label and box conventions.
Verify whether a needed extension point already exists. If a core change is
required, make it generic for external detection workers rather than naming the
target model.
Keep The Boundary
Keep committed DeepDetect changes reusable:
- Use the generic
external-pytorch-detector profile.
- Do not add a first-slice hard-coded
<target>-detector model profile.
- Load target code through
mllib.entrypoint or service_mllib.entrypoint.
- Preserve the public worker contract:
DeepDetectWorker.configure, train,
and predict.
- Put generated target adapter code in
extern/pytorch_workers/<model_slug>/
unless the user explicitly asks to commit model-specific code.
- Treat
AGENTS.md as operational CLI and monitoring guidance, not the place
for a long model-porting workflow.
Target-specific adapter directories are local workspaces by default. The core
repository should not import them unless the user selects them in YAML or API
parameters.
Build The Adapter
Create this layout for a generated external worker:
extern/pytorch_workers/<model_slug>/
worker.py
config.yaml
manifest.json
README.md
notes.md # optional
The manifest should record upstream repository URL, local checkout path, commit
or tag when known, license, dependencies, entrypoint, class name, expected config
path, checkpoint compatibility, and generation notes.
The README is required. It should include a quickstart with the concrete train
and inference CLI commands for the adapter, required upstream checkout/config
settings, checkpoint expectations, key environment variables, and any
model-specific label or bbox conversion notes.
Prefer subclassing the existing detection training base for object detectors.
Implement only the target-specific pieces:
- Backend import and dependency checks.
- Model and postprocessor construction.
- Optional target config patching, such as class count, disabled teacher models,
disabled pretrained downloads, or single-process defaults.
- DeepDetect batch-to-upstream target conversion.
- Upstream prediction-to-DeepDetect output conversion.
- Checkpoint load/save compatibility.
- Model-specific optimizer construction only when the default optimizer is not
suitable.
Make conversions explicit. DeepDetect detection data normally uses pixel xyxy
boxes and one-based foreground labels, with class 0 reserved for background.
Many DETR-style models expect normalized cxcywh boxes and zero-based labels.
Convert labels and boxes in both directions and keep the mapping visible in code.
When useful, preserve or synthesize generic target metadata: orig_size, size,
area, and iscrowd. Keep existing boxes, labels, and image_id behavior
compatible with torchvision-style detection workers.
Raise typed worker errors:
- Use dependency errors for missing upstream repos, missing Python packages,
invalid config paths, and import failures.
- Use contract errors for invalid worker classes, malformed batches, unsupported
prediction schemas, and impossible target conversions.
- Let training errors carry real model failures after dependency and contract
checks have passed.
Add Generic Core Changes Only
Commit DeepDetect core edits only when they help more than one external detector.
Examples of acceptable generic work:
- Clearer runtime errors for external entrypoint loading.
- Tests that an external worker can be loaded from outside the packaged
deepdetect module.
- Generic hooks for target conversion, prediction conversion, checkpoint formats,
and optimizer construction.
- Dataset metadata enrichment needed by DETR-style detectors.
- CLI support that passes
service_mllib.entrypoint, service_mllib.class, and
target-specific YAML values through normal training flows.
Do not commit a target adapter, target config, or target dependency workaround
as a DeepDetect core change unless the user explicitly requests it.
Validate
For committed generic backend work, add focused tests for the reusable behavior:
- Runtime loading from a temporary external entrypoint path.
- Failure mapping to dependency, contract, or launch errors.
- Preservation of current torchvision detection behavior.
- Prediction, target conversion, checkpoint, and optimizer hooks.
- CLI
external-pytorch-detector config pass-through.
For generated target adapters, add local tests with fake upstream modules when
possible. Gate real-upstream tests behind an environment variable such as
<MODEL>_REPO.
Run Python tests from the repository with the local package on PYTHONPATH, for
example:
PYTHONPATH=bindings/python python3 -m pytest bindings/python/tests/test_pytorch_worker_runtime.py bindings/python/tests/test_cli.py
For manual CLI smoke tests, use the source CLI or project wrapper and select the
external detector profile:
PYTHONPATH=bindings/python python3 -m deepdetect.cli.main train external-pytorch-detector \
--config extern/pytorch_workers/<model_slug>/config.yaml \
--train-data train.txt \
--test-data test.txt \
--repository runs/<model_slug>-smoke \
--nclasses 2 \
--iterations 10 \
--test-interval 5 \
--batch-size 1 \
--terminal verbose \
--output-format jsonl
Add --gpu --gpuid <id> when CUDA is required. Add --visdom --visdom-results
when visual monitoring is needed, and inspect sink_warning, run.json,
metrics.jsonl, and saved files under REPOSITORY/visdom-results/.
Triage Failures
If upstream import fails, install or document the missing dependency and improve
the adapter's dependency error message. If upstream code assumes distributed
training, prefer adapter-local single-process setup or disable the upstream path
that calls distributed APIs before the process group exists.
If evaluation fails after training begins, inspect device placement, cached CPU
tensors, postprocessor size arguments, label offsets, and box units. Prediction
conversion bugs often appear first during test intervals or visual result
generation.
If metrics look plausible but rendered visual results are wrong, inspect the
saved image and JSON pairs before changing training settings. Check test-set
ordering, sample indices, coordinate sizes, confidence thresholds, class ids,
and whether the backend shuffled evaluation data.