ARTI guides
Train with your own model and loss
Define a native task once. Connect its data, parameter ownership and scalar objective to the shared training session.
Choose the training entrance
Use the entrance that matches the object your application already owns. A task recipe adds the complete local run lifecycle; the shared session is also available directly.
| Application | Entrance | What you supply |
|---|---|---|
| Attached host model | model.arti.trainer | Configured layer and host loss |
| Native PyTorch model | arti.trainer | Model and objective callable |
| Complete custom task | TaskRecipe | Factory, data, objective, prediction and evaluation |
Separate task semantics from resource locations
TaskRecipeConfig names the installed factory. Put the task’s mathematical choices in options and relocatable paths in resources. The factory receives the configuration and a mode; train returns TaskRecipe, while predict returns TaskPredictor without building an optimizer.
import arti
config = arti.TaskRecipeConfig(
task="my-task",
factory="my_package.tasks:build_task",
options={"input_shape": [32, 32, 3], "loss_profile": "original-v1"},
resources={"data": "/current/location/dataset.jsonl"},
training=arti.ARTITrainingConfig(device="cpu", steps=100),
)
recipe = arti.TaskRecipe.from_config(config)
result = recipe.fit(run=arti.ARTITrainingRunConfig(output_dir="outputs/run"))Keep the actual task objective
objective(model, batch) returns a differentiable scalar Tensor. The shared engine performs backward, accumulation and optimizer steps. The application retains decoding, collation, labels and metric meaning.
| Contract | Application responsibility |
|---|---|
| objective_id | Version the meaning of the loss |
| training_weight(batch) | Declare normalization units for unequal batches |
| evaluation_weight | Preserve native-loss evaluation units |
| dataset_manifest.sha256 | Compute and verify the actual data identity |
Make execution, optimization and saved state explicit
TrainingBinding(model, parameters=..., state_module=...) separates the executed object, selected trainable parameter owners and the tensor state needed for recovery. Include every supported buffer and resource required to reconstruct the task. An excluded immutable host must be reconstructed by the pinned factory.
Start with the installed file tutorial
The files template includes model code, data, configuration and workflow.py. Generating and checking the task project does not invoke the training factory. Continue with its TUTORIAL.md to adapt the objective and callbacks.
python -m arti task init my-task --name "my task" --template files
cd my-task
python -m arti task check project.json
python -m arti task status project.json
python -m arti task explain project.jsonThis guide targets arti-fit 3.3.0 · Public source dcb321d6