NOTICE — Attribution and Provenance

This repository hosts an INT8-quantized ONNX of a YOLOX-Nano variant with ReLU activations and no depthwise convolutions, trained from randomly-initialised weights on COCO 2017 by Pablo Mendoza (@thefalley).

The trained weights, the ONNX export pipeline, the INT8 quantization, the decoder, and the inference scripts are released under the MIT License (see LICENSE).

Provenance chain

Training architecture: YOLOX-Nano-ti-lite        BSD-3 + Apache-2.0
   TexasInstruments/edgeai-yolox                 (used as build tool only)
   exps/default/yolox_nano_ti_lite.py
   (depth=0.33, width=0.25, ReLU activations, regular Conv2D — no depthwise)
        │
        │  Training loop (TI's train.py invoked locally)
        │  Dataset: COCO train2017 (CC BY 4.0)
        │  Hardware: own server NVIDIA GTX 1050 Ti, 4 GB VRAM
        │  Batch: 32, fp16 mixed precision, image size 416×416
        ▼
yolox_nano_relu.pth          (epoch N checkpoint — original work, MIT)
        │
        │  TI's tools/export_onnx.py (used as tool; --opset 13, --no-onnxsim)
        ▼
yolox_nano_relu_float.onnx   (this repo, MIT)
        │
        │  onnxruntime.quantize_static (MIT, used as tool)
        │  + COCO val2017 calibration (1000 images, CC BY 4.0, seed=42)
        ▼
yolox_nano_relu_int8_qop.onnx (this repo, MIT)

The TI source code is used as a build tool only and is not redistributed in this repository. The weights themselves are an original work output of the training process performed by Pablo Mendoza.

Component-level attribution

1. Training architecture & training code — Texas Instruments / edgeai-yolox

Project: TexasInstruments/edgeai-yolox
Source: https://github.com/TexasInstruments/edgeai-yolox
License: BSD-3-Clause + Apache-2.0
What we use: the experiment file exps/default/yolox_nano_ti_lite.py and the standard YOLOX training loop. The architecture replaces SiLU activations with ReLU and depthwise separable convolutions with regular Conv2D, originally proposed by TI for their Jacinto SoCs.
What we do NOT redistribute: any source code, configs, or compiled binaries from TI's repo. This HF repo contains only the trained weights and our own export/quantization/inference scripts.

2. Original YOLOX architecture — Megvii

Project: Megvii-BaseDetection/YOLOX
Source: https://github.com/Megvii-BaseDetection/YOLOX
License: Apache-2.0
Reference: Ge et al., "YOLOX: Exceeding YOLO Series in 2021", arXiv:2107.08430
What we use: the academic credit for the architecture. We do not use Megvii's pretrained weights — those use SiLU activations which our target hardware (an INT8 FPGA DPU) cannot run.

3. Training data — COCO train2017

Source: https://cocodataset.org
License: Creative Commons Attribution 4.0 (CC BY 4.0)

4. Calibration data — COCO val2017

Source: same dataset, validation split
1,000 images randomly sampled with random.Random(42).sample(...) (deterministic). No COCO image is embedded inside the published ONNX file.

5. Conversion and quantization tools — Microsoft ONNX Runtime, PyTorch

torch.onnx.export (BSD-3-Clause) — used as a tool, not redistributed
onnxruntime.quantization.quantize_static (MIT) — used as a tool, not redistributed

File-level integrity (SHA-256)

Hashes will be added once training reaches a milestone where the checkpoint is "frozen" for that version. The README always reflects the currently-published file hashes.

File	Size	SHA-256
`yolox_nano_relu_float.onnx`	***	***
`yolox_nano_relu_int8_qop.onnx`	***	***

Author of the trained weights and INT8 derivative

Pablo Mendoza — HuggingFace @thefalley

Companion repositories (same author, MIT-clean INT8 detectors):

Repository	INT8 size	mAP@0.5:0.95
`Thefalley/yolov4-leaky-416-int8-qop`	61.66 MiB	0.345
`Thefalley/yolov4-tiny-416-int8-qop`	5.83 MiB	0.163
`Thefalley/yolo-fastest-1.1-320-int8-qop`	0.47 MiB	(pending)

Contact

If you are a rights holder and believe this attribution is inaccurate or incomplete, please open an issue on this repository and it will be corrected promptly.