# NOTICE — Attribution and Provenance

This repository hosts an **INT8-quantized ONNX derivative** of
**YOLOv4-tiny-416**, prepared by **Pablo Mendoza** (`@thefalley`) for
deployment on a custom INT8 DPU (ZedBoard XC7Z020 FPGA).

The work this repo adds (ONNX export pipeline + INT8 quantization with
COCO calibration) is released under the MIT License (see `LICENSE`). All
upstream components keep their original licenses, listed below in
dependency order.

---

## Provenance chain

```
AlexeyAB / darknet                    (YOLO License v2 = public domain)
  yolov4-tiny.weights  (23.13 MiB)
  yolov4-tiny.cfg
        │
        │  parsed and loaded by:
        ▼
gwinndr / YOLOv4-Pytorch (MIT)         (used as conversion tool)
  utilities/configs.py::parse_config       — parses darknet .cfg dynamically
  utilities/weights.py::load_weights        — reads AlexeyAB binary .weights
  + small in-repo patch to handle [route] groups
  + DarknetRaw wrapper (this repo, MIT) that captures pre-YoloLayer outputs
        │
        │  torch.onnx.export(opset=13)
        ▼
yolov4-tiny-416_float.onnx            (this repo, MIT — derivative)
  2 raw outputs:
    out_stride16  — shape (1, 255, 26, 26)
    out_stride32  — shape (1, 255, 13, 13)
        │
        │  onnxruntime.quantize_static (MIT, used as tool)
        │  + COCO val2017 calibration (1000 images, CC BY 4.0)
        ▼
yolov4-tiny-416_int8_qop.onnx         (this repo, MIT — derivative)

External Python decoder (this repo, MIT) reproduces the standard YOLOv4
post-processing: sigmoid + scale_xy + grid offset + anchor multiplication
+ NMS. This decoder is not part of the ONNX graph (matches the standard
deployable design where the decoder runs on the host CPU after the DPU
executes the conv backbone+neck+head).
```

---

## Component-level attribution

### 1. Model weights — AlexeyAB / darknet

- **Project**: `AlexeyAB/darknet`
- **Source**: https://github.com/AlexeyAB/darknet
- **Weights URL**: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights
- **License**: YOLO License v2 — *"Darknet is public domain. Do whatever you want with it."*
- **What we use**: the trained `yolov4-tiny.weights` file (23.13 MiB) and the corresponding `yolov4-tiny.cfg`. No modifications.

### 2. Darknet → PyTorch conversion — gwinndr / YOLOv4-Pytorch

- **Project**: `gwinndr/YOLOv4-Pytorch`
- **Source**: https://github.com/gwinndr/YOLOv4-Pytorch
- **License**: MIT (Copyright (c) 2020 Damon Gwinn)
- **What we use**: the `parse_config` cfg parser and the `load_weights` AlexeyAB binary loader.
- **Modifications we made** (to handle yolov4-tiny's CSPDarknet route trick):
  - Patched `utilities/configs.py::parse_route_block` to read `groups` and `group_id` from `[route]` blocks.
  - Patched `model/layers/route.py::RouteLayer` to apply channel-wise split when `groups > 1`.
  - These patches are also released under MIT.
- We do NOT redistribute gwinndr's source code in this HF repo. We reference it as a build-time tool. To reproduce, clone gwinndr's repo and apply the patches in our companion firmware repository.

### 3. ONNX export

- **Tool**: `torch.onnx.export` (PyTorch core, BSD-3-Clause). Used as a tool, not redistributed.
- **Wrapper**: `DarknetRaw` (≈ 30 lines, this work, MIT) intercepts the pre-YoloLayer feature maps and exports them as 2 raw 4D tensors. The decoder lives in `inference.py` (Python, MIT).

### 4. INT8 quantization

- **Tool**: `onnxruntime.quantization.quantize_static` (Microsoft, MIT). Used as a tool, not redistributed.
- **Configuration**:
  - `quant_format = QuantFormat.QOperator`
  - `weight_type = QuantType.QInt8` (per-tensor, symmetric)
  - `activation_type = QuantType.QInt8` (per-tensor, asymmetric)
  - `per_channel = False`
  - `reduce_range = False`
- **Calibration data**: 1000 randomly-sampled images from MS COCO val2017 (CC BY 4.0). No COCO image is embedded inside the ONNX file; the dataset's role ends after calibration.

### 5. Calibration dataset — COCO val2017

- **Source**: https://cocodataset.org
- **License**: Creative Commons Attribution 4.0 (CC BY 4.0) for both images and annotations.

---

## File-level integrity (SHA-256)

| File | Size | SHA-256 |
|---|---:|---|
| `yolov4-tiny.weights` (AlexeyAB upstream)        | 24,251,276 B | `cf9fbfd0f6d4869b35762f56100f50ed05268084078805f0e7989efe5bb8ca87` |
| `yolov4-tiny-416_float.onnx`                     | 24,230,209 B | `eea691d460fd3eb5c1a250b4e5f822784cd44e11aaa77a24299b0952b9f4fc9f` |
| `yolov4-tiny-416_int8_qop.onnx`                  |  6,113,440 B | `c30c8f0a33b3a0edc13a2ca21726a288228e1448b3c38940f9da0c7d8cee4760` |

---

## Author of the INT8 derivative

**Pablo Mendoza** — HuggingFace [`@thefalley`](https://huggingface.co/thefalley)

Companion repositories:

| Repository | Purpose |
|---|---|
| [`Thefalley/yolov4-leaky-416-int8-qop`](https://huggingface.co/Thefalley/yolov4-leaky-416-int8-qop) | Larger sibling: full YOLOv4-Leaky-416 INT8 (61.66 MiB) for higher mAP |
| `Thefalley/dpu-firmware` (GitHub, ***) | Bare-metal C firmware + RTL for the custom DPU on ZedBoard XC7Z020 |

## 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.