Skip to content

Check what this machine can do

For: someone deciding whether a particular computer โ€” a field laptop, a shared server, a Raspberry Pi โ€” can run AdaptShot, and what it would cost there. Ten seconds to run; works offline.

Ask it

import adaptshot
print(adaptshot.check_environment())

Every figure in the report is measured on the machine running the check: the latency is a real inference on the bundled photographs, the memory is this process's own high-water mark. Nothing is quoted from a document or from someone else's benchmark.

Reading the report

AdaptShot 0.3.0 -- environment report (everything below was measured here)
  Python 3.11.9 ยท Linux 6.1 ยท aarch64 ยท 4 cores ยท 1.9 of 3.8 GB RAM free
  numpy 2.2.6   Pillow 11.1.0   torch: not installed   onnxruntime 1.21.0   faiss: not installed   gradio: not installed
  bundled backbones: mobilenet_v3_small
  Available now:
    โœ“ predict, correct, save / load        88.4 ms per image, median of 5, measured here on mobilenet_v3_small
    โœ“ conformal prediction sets            coverage guarantee validated in tests/test_conformal_coverage.py; needs ceil((1-alpha)/alpha) calibration scores to be informative
    โœ“ out-of-distribution flag             leave-one-out-calibrated Mahalanobis; at least 3 support photos per class
  Not available:
    โœ— fine-tuning (CA-EWC) via correct()   needs torch; download size not measured here (requires the network)
                                           needs: pip install "adaptshot[torch]"
    โœ— backbones other than the bundled one only the bundled backbone(s) without torch
                                           needs: pip install "adaptshot[torch]"
    โœ— faster search for support sets over 100 images   numpy search is used; fine below ~100 images
                                           needs: pip install "adaptshot[faiss]"
  Fits the 250 MB target here: yes -- this process peaked at 118 MB
  • Available now is what works with what is installed. If predict is there, the tutorials work.
  • Not available lists each optional capability with the exact pip command that enables it. The download size is deliberately not stated: it cannot be measured without the network, and a number copied from elsewhere is what this report exists to avoid.
  • The latency line is this machine, now. A busy machine reports a slower number โ€” that is the point. Compare with the published figures for the machine they were measured on.
  • The memory line is the process running the check. If you ran this after importing PyTorch, the figure includes PyTorch and the report says so; the 250 MB target describes the standard install.
  • A GPU, if present, is named and not selected. AdaptShot's defaults stay on the CPU on purpose โ€” the explanation says why.

When you cannot afford the second it takes

import adaptshot
report = adaptshot.check_environment(measure=False)
print([c.name for c in report.capabilities if c.available])

measure=False skips the inference and the memory reading and returns in under a millisecond โ€” for a start-up check in an application, where you want to know whether prediction works, not how fast.

Using it from code

The return value is an EnvironmentReport dataclass, so an application can branch on it:

import adaptshot
report = adaptshot.check_environment(measure=False)
needs = {c.name: c.install for c in report.capabilities if not c.available}
if "fine-tuning (CA-EWC) via correct()" in needs:
    print("Fine-tuning is off on this machine. To enable it:", needs["fine-tuning (CA-EWC) via correct()"])

check_environment is in the experimental tier of the API: the report's shape may change in a minor release as people say what they need in it. See API stability.