Use a different backbone¶
For: someone who suspects the bundled encoder is the limit on their task and wants to try another. Assumes the tutorials. Needs the PyTorch extra for anything beyond the bundled backbone.
What a backbone is here¶
The backbone is the frozen network that turns a photograph into a list of numbers. AdaptShot never trains it; everything it learns sits on top. The standard install bundles one, mobilenet_v3_small (4 MB, ImageNet-pretrained), exported to ONNX so no PyTorch is needed to run it.
Two backbones are registered:
| name | size | needs | notes |
|---|---|---|---|
mobilenet_v3_small |
4.0 MB | nothing — bundled | the default; the one every published number was measured with in 0.3.0 |
resnet18 |
44.8 MB | the torch extra | not bundled because of its size; torchvision fetches its weights on first use |
Switch¶
from adaptshot import AdaptShotConfig, FewShotLearner
from adaptshot.data import sample_images
paths, labels = sample_images()
learner = FewShotLearner(config=AdaptShotConfig(backbone="resnet18")) # docs: not run -- downloads resnet18 weights on first use
learner.load_support_images(paths[:-1], labels[:-1])
print(learner.predict(paths[-1]).prediction)
The first use of resnet18 downloads its weights through torchvision (once, cached under ~/.cache/torch), which means it needs the network that first time. Everything after is offline.
On a standard install without torch, the same code raises BackboneError naming both ways out: use a bundled backbone, or install adaptshot[torch]. It does not raise ImportError: torch from four frames inside the library.
Should you expect it to help?¶
On the CIFAR-10 smoke split (25 queries) the two scored 68% and 76% — intervals that overlap heavily, so at that size they are indistinguishable — and the larger backbone was slower. The PlantVillage result was measured on the bundled backbone only. Your task may differ. Measure it: run the benchmarks with --backbone, or hold out photographs as in tutorial 3 and compare.
Saved learners and backbones¶
A saved learner records the backbone its numbers came from and can only be loaded meaningfully on an install that provides it. A resnet18 learner needs torch wherever it is loaded; a mobilenet_v3_small learner loads anywhere. See save, load and migrate.
Adding a backbone that is not registered¶
BackboneRegistry in adaptshot.core.extractor maps a name to a torchvision constructor, and BACKBONE_OUTPUT_DIM records its embedding width. Adding an entry there, and to the Backbone type in adaptshot.config.settings, registers a new torch backbone. To run it without torch, export it to ONNX. Both are contributions rather than configuration; the contributing guide covers the tests that go with them.