API reference

deepposekit.models

Model backbones for pose estimation. All models share the same fit / predict / save / load surface.

StackedDenseNet

StackedDenseNet(
    train_generator,
    n_stacks=2,
    n_transitions=-1,
    growth_rate=32,
    bottleneck_factor=1,
    compression_factor=0.5,
    pretrained=True,
    subpixel=True,
)

The toolkit’s recommended default. Combines DenseNet-style feature reuse with stacked, multi-scale supervision.

StackedHourglass

StackedHourglass(
    train_generator,
    n_stacks=2,
    n_transitions=4,
    filters=256,
    bottleneck_factor=2,
    subpixel=True,
)

A re-implementation of the classic stacked hourglass architecture, included as a baseline for comparison.

load_model

from deepposekit.models import load_model

model = load_model('/path/to/saved_model.h5', augmenter=None, generator=None)

Reload a previously saved model. The keypoint skeleton and metadata are restored from the HDF5 file.