vyntri.backbones

Backbone Registry

Backbone	Features	Params	Input Size	Description
mobilenet_v3_small (default)	576	2.5M	224	Fast, lightweight, good for edge
resnet18	512	11.7M	224	Good balance of speed and accuracy
resnet50	2048	25.6M	224	Higher capacity, slower extraction
efficientnet_b0	1280	5.3M	224	Efficient architecture, strong accuracy
convnext_tiny	768	28.6M	224	ConvNeXt architecture, competitive performance

Registry validated at construction. An invalid backbone name produces a clear error listing all supported names.

Automatic Backbone Selection

When backbone="auto", Vyntri scores candidates with LogME on frozen features, validates the top-k candidates with the analytic pipeline, and selects the best performer. Requires validation data.

from vyntri import Vyntri
from vyntri.data import split

s = split("./my_dataset", train=0.7, test=0.2, seed=42)

model = Vyntri(backbone="auto")
model.fit(train=s.train, val=s.val)

# View selection results
sel = model.metadata_["selection"]
print(f"Selected: {sel['selected']}")
print(f"Scores: {sel['scores']}")

Candidate Control

Control which backbones are evaluated during auto-selection. Only the specified candidates are benchmarked — excluded backbones are never initialized or evaluated.

model = Vyntri(
    backbone="auto",
    candidates=[
        "mobilenet_v3_small",
        "resnet18",
        "efficientnet_b0",
    ],
)
model.fit(train=s.train, val=s.val)

Rules:

candidates is only valid when backbone="auto" — combining candidates with an explicit backbone raises an error
Invalid candidate names are rejected with a clear error message
Duplicate candidates are rejected
If no candidates are provided, all supported backbones are evaluated
Selection is deterministic: same candidates + same data + same seed = same result