lepinet · EfficientNetV2-S: moth and butterfly identification, small and fast

Identifies Lepidoptera (moths and butterflies) from a photo at three ranks at once: species (12,041), genus (4,333) and family (102). The three answers are coherent, because genus and family are computed from the species probabilities.

This is the small model of the lepinet project: 37 M parameters (a 20 M backbone plus the 12,041-class classifier), fast enough for a laptop CPU or an edge device. It is about 2 points of species macro-F1 behind the two larger released models (0.765 vs 0.777 and 0.783 on light-trap images; 0.899 vs 0.906 and 0.921 on ordinary photos), and close to the best of them under a back-off policy; see Which lepinet model?. Use a large one when accuracy matters more than speed.

Which lepinet model?

The three released models are about equally accurate on light-trap images (species macro-F1 0.783, 0.777 and 0.765) but answer differently when they are allowed to back off to genus or family. Under the same 95 %-precision back-off policy, with thresholds fitted on half of the held-out light-trap nights and measured on the other half:

• To get the species name as often as possible, the DINOv3 model names it most often at 95 % precision, and answers "unknown" more often.
• If a genus or family answer is useful, the BioCLIP-2 model gives some correct answer most often, because it backs off more.
• The small model is close to the best on both and runs on a laptop CPU.
• On ordinary photos the BioCLIP-2 model is the most accurate (species macro-F1 0.921, against 0.906 and 0.899 over the full GBIF test fold).

Differences of about a point are within the variation between the two halves of the nights. An earlier comparison reported a 17-point advantage for the BioCLIP-2 model; it counted a family answer like a species answer, fitted thresholds in-sample and used the DINOv3 model's uncalibrated probabilities (analysis).

Quick start

pip install onnxruntime pillow numpy huggingface_hub

import json, numpy as np, onnxruntime as ort
from huggingface_hub import snapshot_download
from PIL import Image

repo = snapshot_download("gmougeot/lepinet-effnetv2s", allow_patterns=["model.onnx", "*.json"])
session = ort.InferenceSession(f"{repo}/model.onnx")
names = json.load(open(f"{repo}/names.json"))["names"]

def load(path, size=256):  # shorter side -> size, centre crop, RGB in [0, 1]
    img = Image.open(path).convert("RGB")
    s = size / min(img.size)
    img = img.resize((max(size, round(img.width * s)), max(size, round(img.height * s))), Image.BICUBIC)
    l, t = (img.width - size) // 2, (img.height - size) // 2
    return np.asarray(img.crop((l, t, l + size, t + size)), np.float32).transpose(2, 0, 1)[None] / 255

out = dict(zip([o.name for o in session.get_outputs()], session.run(None, {"image": load("moth.jpg")})))
for rank in ("species", "genus", "family"):
    p = out[f"prob_{rank}"][0]
    print(f"{rank:8s} {names[rank][p.argmax()]}  {p.max():.2f}")

Or use the bundled command-line script, which applies the confidence thresholds for you:

hf download gmougeot/lepinet-effnetv2s model.onnx predict.py config.json taxonomy.json names.json thresholds.json --local-dir lepinet
python lepinet/predict.py moth1.jpg moth2.jpg --top 3

Which file to use

Measured with onnxruntime 1.30 on an RTX 5090 and a 24-core Intel Core Ultra 9 285K (a shared server under some background load, so read the ratios rather than the absolute CPU numbers). fp16 keeps the classifier in fp32 and agrees with fp32 on 99.9 % of species predictions on the probe set; on the GPU, predictions match the CPU on every image tested.

Why there is no int8 file. Plain dynamic int8 made this network 4× slower, because onnxruntime has no fast kernel for its int8 convolutions. Calibrated static int8 does run 2.2× faster (374 img/s), but it lost 0.6–1.2 points of species macro-F1 on light-trap images in every variant tried: different calibration data, keeping the depthwise convolutions in fp32. At 10 ms per image on a CPU, that trade is not worth it. The large model, for which it is, has an int8 file.

Running on a GPU

pip install "onnxruntime-gpu[cuda,cudnn]"   # instead of onnxruntime; do not install both

ort.preload_dlls()   # loads the CUDA/cuDNN libraries that pip installed
session = ort.InferenceSession(f"{repo}/model.onnx", providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

Name the providers explicitly. ort.get_available_providers() puts TensorRT first, and without TensorRT installed onnxruntime then falls back to the CPU without saying so. predict.py picks CUDA automatically.

The three lepinet releases share identical taxonomy.json and names.json files, so they are drop-in interchangeable.

Files

Outputs

The model has a single species head; genus and family exist only as sums over species.

Calibration. The ArcFace margin makes the raw softmax over-confident: 71 % of validation images would get a species probability of exactly 1.0 in float32. The graph therefore divides the species logits by a temperature T = 1.70, fitted by likelihood on 10,000 in-distribution validation images. This lowers the negative log-likelihood from 0.626 to 0.484, and the share of saturated predictions to 4.7 %. Rankings, and so accuracy, are unchanged.

Using the confidence

Every image gets a species prediction, including images of species outside the label set and images that are not moths at all. thresholds.json implements the back-off policy the project evaluates:

answer species  if prob_species.max() >= t_species
else genus      if prob_genus.max()   >= t_genus
else family     if prob_family.max()  >= t_family
else "unknown"

In code, continuing the quick start above (out, names and repo come from there):

thresholds = {k: v["threshold"] for k, v in json.load(open(f"{repo}/thresholds.json"))["levels"].items()}

def answer(out, i=0):
    """Deepest rank the model is confident about for image i, or None ("unknown")."""
    for rank in ("species", "genus", "family"):
        p = out[f"prob_{rank}"][i]
        if p.max() >= thresholds[rank]:          # a threshold above 1 means "never answer at this rank"
            return rank, names[rank][p.argmax()], float(p.max())
    return None

print(answer(out))    # ('species', 'Macaria notata', 0.99), ('family', 'Erebidae', 0.97) or None

p = out["prob_species"][0]
novelty = float(-(p * np.log(p + 1e-12)).sum())   # entropy: higher = less familiar image

predict.py does exactly this.

The thresholds were fitted on half of the light-trap capture nights and verified on the other half:

Overall, on the held-out nights: 90.0 % of images get an answer, 95.3 % of those answers are correct, so 85.8 % of all images get a correct answer. The thresholds shipped in thresholds.json (levels) are refitted on all nights: species 0.93, genus never, family 0.505. The split fit and its held-out result are stored alongside them.

Species the model has never seen

Any closed-set classifier gives a species to a species it does not know. What the back-off buys is that it usually does not commit to one. Measured with the file above on 3,171 GBIF photos of 591 species outside the label set (the test-fold species below the training floor):

The entropy novelty score separates the two sets with AUROC 0.908. These thresholds were fitted on light-trap images, so on ordinary photos they are a little conservative.

"Unknown" is not proof of a novel species, and a species answer is not proof that the species is in the label set: the same rule catches hard photos of known species. Use novelty for a graded score, for instance to send the most unfamiliar images to an expert.

Evaluation

Species macro-F1 averages F1 over species, so the rare species count as much as the common ones.

• GBIF test fold (in-distribution): held-out images from the same GBIF sources as training. The full fold is 629,742 images of all 12,041 species; the fp16 file was checked on a random 10,000-image sample.
• Probe: 15,200 light-trap images of 368 Danish species, from (trap, night) groups never used in training. This is a real domain shift: night-time camera crops, not curated photos.
• Probe, held-out species: 2,455 images of 58 species for which no trap images were used in any form during training.

Genus and family macro-F1 on probe, from the summed probabilities: 0.816 and 0.798.

The two columns differ by under 1 point, which is within the run-to-run noise of these evaluation sets. The network is numerically identical to the PyTorch model; the difference is image resizing. Trap crops are small (median shorter side 157 px), so they are up-sampled, and the resampling details change 3–5 % of individual predictions. For small crops, resize consistently.

For comparison, the project's EfficientNetV2-S baseline, trained without the trap-domain steps, scores 0.6270 on probe.

Training

1. Data: 6.3 M GBIF occurrence images of Lepidoptera; species with at least 50 images; at most about 2,000 images per species; a held-out test fold. Square-root class-balanced oversampling.
2. Model: EfficientNetV2-S with a cosine classifier trained under an ArcFace angular margin (0.3, z-score logits); species head only. 5 epochs, end to end, Muon optimiser, one-cycle schedule.
3. Trap robustness, both at training time:
   • domain augmentation that mimics light-trap imagery (blur, low light, JPEG);
   • self-training: about 2 % of each epoch is unlabelled light-trap crops, pseudo-labelled by an earlier model. This is the single largest robustness gain in the project. The trap nights used for evaluation are excluded.

Full recipe, ablations and every number: the lepinet repository, starting at its README. The configuration is configs/20260804_B3rep5x_effnetv2s.yaml.

Limitations

• One insect per image. The model classifies; it does not detect. Run a detector first on scenes, trap screens or multi-insect photos (for example flat-bug) and classify the crops.
• Closed label set, uneven geography. It knows 12,041 species. Training images are 36 % North American, 33 % European and 15 % Asian, but only 6 % South American or African, as GBIF coverage is. A species outside the label set still receives one of these labels: use the thresholds, and expect the tropics to be served worse.
• Domain shift is large: compare the in-distribution and probe rows above. The thresholds come from Danish light traps and may not transfer to your camera or region. Re-fit them on a few hundred labelled images of your own if precision matters.
• Adults only, in practice. Larvae are about 1 % of the training images; caterpillars, eggs and pupae will be poorly identified.
• Look-alike species that need genitalia dissection or DNA cannot be separated from a photo. Treat the genus answer as the honest one there.
• Not for decisions about conservation status, pest control or toxicity without expert verification.

Licence

CC-BY-NC-4.0. About 77 % of the training images are licensed CC-BY-NC by their GBIF contributors, so the weights are released for non-commercial use. The lepinet training code is GPL-3.0.

Citation

No paper yet. Please cite the repository:

@software{mougeot_lepinet_2026,
  author = {Mougeot, Guillaume},
  title  = {lepinet: hierarchical Lepidoptera identification that knows what it does not know},
  year   = {2026},
  url    = {https://github.com/GuillaumeMougeot/lepinet}
}

Please also acknowledge the GBIF contributors whose images made this possible: GBIF.org (15 May 2025) GBIF Occurrence Download https://doi.org/10.15468/dl.hg37y9.