fotonet

fotonet is an open-source, compact NMS-free object detector for computer vision and machine learning workflows. It is built with PyTorch and provides a Python API, resumable training, COCO-style validation, small-object P2 variants, and ONNX or TorchScript export.

The repository is being published while the first official Nano checkpoint is still training. Model names currently construct untrained architectures. The trained weight, SHA256 checksum, canonical validation report, and automatic download hook will be published after training and release verification finish. There is no public AP claim yet.

Documentation: https://hazegreleases.github.io/fotonet/

Current source/package version: v0.8.0b2 (beta).

Install

python -m pip install fotonet

For a development checkout:

git clone https://github.com/hazegreleases/fotonet.git
cd fotonet
python -m pip install -e ".[dev]"

Inference

from fotonet import Fotonet

model = Fotonet("path/to/checkpoint.pt")
results = model.predict("image.jpg", conf=0.25, imgsz=640)

for detection in results[0].boxes:
    print(detection.cls, detection.conf, detection.xyxy)

Image and tensor calls return list[Results]. Video/webcam prediction returns an iterator only when stream=True.

Box transforms

from fotonet import AnchorPoint

crop = (
    results[0].boxes[0].transform
    .set_anchor(AnchorPoint.CENTER)
    .pixel_expand(40)
    .clamp()
    .crop(results[0].orig_img)
)

Training

Download the public launcher, then point it at a YOLO-format dataset:

curl -L https://hazegreleases.github.io/fotonet/examples/train.py -o train.py

python train.py \
  --model fotonetn \
  --data path/to/data.yaml \
  --epochs 300 \
  --batch 16 \
  --run-dir runs/fotonetn

Resume an interrupted run without starting a second training protocol:

python train.py \
  --model fotonetn \
  --data path/to/data.yaml \
  --epochs 300 \
  --batch 16 \
  --run-dir runs/fotonetn \
  --resume

No training is started by importing the package or by a launcher --dry-run.

Export

from fotonet import Fotonet

model = Fotonet("path/to/checkpoint.pt")
output = model.export(format="onnx", path="exports/fotonet.onnx", imgsz=640)
print(output["artifact"], output["metadata"])

Supported checkpoints and exports are self-identifying and tensor-only loaded. Missing or unknown schema versions fail closed; the runtime does not infer a model from filenames or tensor shapes.

Architecture and measurements

Every public model uses the same production Backbone, Neck, Head, and Detector implementation under fotonet.models.v1. P2 variants add a stride-4 prediction level for small-object experiments.

The current Nano graph has 1,042,936 training parameters (1,005,932 fused deployment parameters), 1.313 GMAC / 2.626 GFLOP at 640x640, and measured 218.33 images/s at batch 1 or 723.43 images/s at batch 8 on the declared RTX 4060 FP32 benchmark. These are graph/runtime measurements, not accuracy claims. See the model table for all ten variants and methodology.

S, M, L, and X will be resized in a later architecture revision. Planned parameter centers are 2.2M, 5.0M, 11.4M, and 33.8M respectively, with the explicit bands documented in the model table. Their future MAC/FLOP values will be measured after the graphs exist; they are not estimated here.

Documentation

• Installation
• Quick start
• Inference and results
• Training and resume
• Models and measured runtime
• Model configuration
• Export
• Transform API
• Security
• Contributing
• Production cleanup boundary

License

Apache License 2.0. See LICENSE.