Training Factories

Status: Supported runtime training surface

Modules: artifex.generative_models.training.optimizers.factory, artifex.generative_models.training.schedulers.factory

Sources: src/artifex/generative_models/training/optimizers/factory.py, src/artifex/generative_models/training/schedulers/factory.py

Artifex owns one shared optimizer factory and one shared scheduler factory. The public contract is typed: pass OptimizerConfig and SchedulerConfig, then feed the resulting Optax objects into Trainer or a family-specific loop.

Optimizer Factory

from artifex.generative_models.core.configuration import OptimizerConfig
from artifex.generative_models.training import create_optimizer

optimizer = create_optimizer(
    OptimizerConfig(
        name="adamw",
        optimizer_type="adamw",
        learning_rate=1e-3,
        weight_decay=0.01,
    )
)

Supported optimizer types today: adam, adamw, sgd, rmsprop, adagrad, lamb, radam, and nadam.

Scheduler Factory

from artifex.generative_models.core.configuration import SchedulerConfig
from artifex.generative_models.training import create_scheduler

scheduler = create_scheduler(
    SchedulerConfig(
        name="cosine",
        scheduler_type="cosine",
        warmup_steps=1_000,
        cycle_length=100_000,
        min_lr_ratio=0.1,
    ),
    base_lr=1e-3,
)

Supported scheduler types today: constant, linear, cosine, exponential, polynomial, step, multistep, cyclic, one_cycle, and none.

Shared Trainer Integration

from artifex.generative_models.core.configuration import (
    OptimizerConfig,
    SchedulerConfig,
    TrainingConfig,
)
from artifex.generative_models.training import Trainer, create_optimizer, create_scheduler

optimizer_config = OptimizerConfig(
    name="adamw",
    optimizer_type="adamw",
    learning_rate=1e-3,
    weight_decay=0.01,
)
scheduler_config = SchedulerConfig(
    name="cosine",
    scheduler_type="cosine",
    warmup_steps=1_000,
    cycle_length=100_000,
    min_lr_ratio=0.1,
)
training_config = TrainingConfig(
    name="factory-demo",
    optimizer=optimizer_config,
    scheduler=scheduler_config,
    batch_size=32,
    num_epochs=10,
)

scheduler = create_scheduler(
    SchedulerConfig(
        name="cosine",
        scheduler_type="cosine",
        warmup_steps=1_000,
        cycle_length=100_000,
        min_lr_ratio=0.1,
    ),
    base_lr=optimizer_config.learning_rate,
)
optimizer = create_optimizer(
    OptimizerConfig(
        name="adamw",
        optimizer_type="adamw",
        learning_rate=1e-3,
        weight_decay=0.01,
    ),
    schedule=scheduler,
)

trainer = Trainer(
    model=model,
    training_config=training_config,
    optimizer=optimizer,
    loss_fn=loss_fn,
)

Standalone pages like adamw.md, cosine.md, and scheduler.md stay in the coming-soon bucket because Artifex does not ship those as standalone runtime training modules today.