Core¤
The retained artifex.generative_models.core package is a narrower shared-runtime layer.
Overview¤
from artifex.generative_models import core
configuration = core.configuration
evaluation = core.evaluation
layers = core.layers
losses = core.losses
protocols = core.protocols
sampling = core.sampling
core stays lazy at import time and should be read through its surviving
child-module boundaries.
Sampling¤
Sampling helpers live under core.sampling.
Evaluation Metrics¤
Evaluation metrics live under core.evaluation.metrics.
from artifex.generative_models.core.evaluation.metrics import (
EvaluationPipeline,
FrechetInceptionDistance,
InceptionScore,
PrecisionRecall,
)
Layers¤
Shared layers live under core.layers.
from artifex.generative_models.core.layers import (
FlashMultiHeadAttention,
ResNetBlock,
TransformerEncoder,
)
Protocols¤
Core protocol types live under core.protocols.
from artifex.generative_models.core.protocols import (
BatchableDatasetProtocol,
MetricBase,
NoiseScheduleProtocol,
)
protocol_types = (BatchableDatasetProtocol, MetricBase, NoiseScheduleProtocol)
The remaining model-facing evaluation protocol surface includes
BenchmarkModelProtocol and DatasetProtocol.
Benchmark runtime types now live under artifex.benchmarks.core, not under
core.protocols or core.evaluation. Scaling and sharding APIs live under
artifex.generative_models.scaling, not under core.
Evaluation Protocols | Benchmark Runtime | Scaling Package
Configuration, Losses, And Distributions¤
The remaining core child modules stay package-owned and should be read through their real module boundaries:
core.configurationfor the surviving typed configuration surfacecore.lossesfor functional loss primitives and family-local objective helperscore.distributionsfor shared distribution implementations
See the owner pages in this section for the detailed module-level contracts.
Surface Map¤
| Surface | Owner |
|---|---|
core.configuration |
shared configuration package and typed config helpers |
core.distributions |
distribution implementations and transforms |
core.evaluation.metrics |
evaluation metrics and explicit dependency pipelines |
core.layers |
shared layers and architectural building blocks |
core.losses |
loss primitives and objective helpers |
core.protocols |
evaluation, metric, and training protocols |
core.sampling |
BlackJAX, MCMC, and SDE sampling helpers |
artifex.benchmarks.core |
benchmark configs, results, NNX benchmark bases, and runners |
core top-level helpers |
checkpointing, rematerialization, and device helpers |