getitune-training-a-model
Train a computer-vision model with the getitune library (the Geti training library) using its Python API or CLI. Use when a user wants to train, fine-tune, or evaluate a model with `create_engine(...)` and `engine.train()/engine.test()`, run `getitune train`/`getitune test`, pick or override a recipe under `getitune.recipe.<task>`, choose a device (cpu/gpu/xpu/cuda), warm-start from a checkpoint, or debug a training run. Covers classification, detection, instance/semantic segmentation, and keypoint detection.
How do I install this agent skill?
npx skills add https://github.com/open-edge-platform/geti --skill getitune-training-a-modelIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill provides instructions for training and evaluating computer vision models using the 'getitune' library. It demonstrates standard Python API usage and CLI commands for model training, dataset handling, and hardware acceleration.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Training a model with getitune
getitune is a low-code transfer-learning library. Training is driven by an
Engine created with create_engine(...), which pairs a model/recipe with a
dataset and returns a runnable engine. Recipes (YAML under
library/src/getitune/recipe/<task>/) bundle model + data pipeline + training
config, so a model name alone gives a strong baseline.
There are two equal entry points that share the same objects and recipes:
- Python API —
from getitune.engine import create_engine, thenengine.train()/engine.test(). Preferred for notebooks, scripts, tests, and library integration. Seelibrary/README.md("Quick Start") andlibrary/docs/source/guide/get_started/api_tutorial.rst. - CLI —
getitune train --data_root <path> --model <name|recipe.yaml>. Preferred for reproducible experiments and shell workflows. Seelibrary/docs/source/guide/get_started/cli_commands.rst.
Run everything from library/. Install with the extra that matches your
hardware: uv sync (cpu), uv sync --extra xpu, or uv sync --extra cuda.
Python API workflow
from getitune.engine import create_engine
engine = create_engine(
model="efficientnet_b0", # model name, recipe .yaml path, or model class
data="/path/to/dataset_root", # dataset root (COCO/YOLO/VOC/native), auto-detected
work_dir="./my_workspace", # checkpoints + logs; defaults to ./getitune-workspace
device="auto", # "auto", "cpu", "gpu", "xpu", "cuda", "0", ...
)
engine.train(max_epochs=50)
engine.test()
- Pick the model/recipe. Pass a model name (
"efficientnet_b0"), a recipe path ("src/getitune/recipe/detection/yolox_s.yaml"), or a model class. If a name matches recipes under several tasks, passtask=(e.g.task="DETECTION") to disambiguate. Use thegetitune-discovering-modelsskill to list options.- Done when:
create_engine(...)returns without aValueError/FileNotFoundError.
- Done when:
- Point
data=at the dataset root. Format is auto-detected by Datumaro; see thegetitune-preparing-datasetsskill.- Done when: the engine builds a datamodule without a format/feature error.
- Smoke-test the wiring first with a tiny run (
engine.train(max_epochs=1)or a small subset) before a long run.- Done when: one train + one validation pass complete without shape errors.
- Train, overriding hyperparameters as needed
(
engine.train(max_epochs=50)).- Done when: checkpoints appear under
work_dir.
- Done when: checkpoints appear under
- Evaluate with
engine.test()and confirm the task metric moves, not just loss. Record the model +work_dirthat produced it.
Warm-start from existing weights with
create_engine(..., checkpoint="/path/to/weights.pt").
CLI workflow
# from library/
# 1. Simplest: data only — getitune picks a default model for the task
getitune train --data_root /path/to/dataset
# 2. Choose a model or recipe
getitune train --data_root /path/to/dataset --model yolox_s
# 3. Override hyperparameters
getitune train --data_root /path/to/dataset --model yolox_s \
--max_epochs 200 --checkpoint /path/to/weights.pt
# 4. Run a full, resolved config file
getitune train --data_root /path/to/dataset --config src/getitune/recipe/detection/yolox_s.yaml
getitune test and getitune predict share the same --model / --data_root
shape. Use getitune <cmd> --help -v (and -vv) for the full overridable
argument list.
Choosing a device
device="auto"selects an available accelerator; force with"cpu","gpu","xpu","cuda", or an index like"0".- The device must match the installed extra —
--extra xpufor Intel GPUs,--extra cudafor NVIDIA. Guard nothing yourself; the library handles capability checks.
Debugging a run
- Run one epoch on a small dataset first to isolate construction vs. dataloading vs. training failures.
- Shape/feature mismatches usually mean the dataset's labels or task disagree
with the model — recheck
task=and the dataset format (getitune-preparing-datasets). - Dataset auto-detection failures: confirm the folder matches one supported layout (COCO/YOLO/VOC/native).
Verify
# from library/
just lint
just test-unit -- -k engine # when you changed engine/training code
For API-facing work, add or run a short Python smoke test that calls
create_engine(...) + engine.train(max_epochs=1) on a tiny fixture rather than
a long real run.
Related skills
getitune-discovering-models— list models/recipes and disambiguate by task.getitune-preparing-datasets— thedata=half of the engine.getitune-exporting-a-model— export a trained checkpoint to OpenVINO/ONNX.getitune-running-inference— run predictions with a trained or exported model.geti-library-dev— when the library/model code itself needs changes.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/open-edge-platform/geti/getitune-training-a-model">View getitune-training-a-model on skillZs</a>