Add NOTICE.md (clean MIT chain)

NOTICE.md

@@ -1,35 +1,46 @@
 # NOTICE — Attribution and Provenance
 
-This repository hosts an **INT8-quantized ONNX derivative** of YOLOv4-Leaky-416,
-prepared by **Pablo Mendoza** as part of a custom DPU FPGA toolchain for the
-ZedBoard XC7Z020.
+This repository hosts an **INT8-quantized ONNX derivative** of
+**YOLOv4-Leaky-416**, prepared by **Pablo Mendoza** (`@thefalley`) for
+deployment on a custom INT8 DPU (ZedBoard XC7Z020 FPGA).
 
-The work this repo adds (calibration + INT8 quantization in QOperator format)
-is released under the MIT License (see `LICENSE`). All upstream components
-keep their original licenses, listed below in dependency order.
+The work this repository adds (ONNX export pipeline + INT8 quantization
+with COCO calibration + decoder) is released under the MIT License (see
+`LICENSE`). All upstream components keep their original licenses, listed
+below in dependency order.
 
 ---
 
 ## Provenance chain
 
 ```
-AlexeyAB / darknet            (original .weights, YOLO License ≈ MIT)
+AlexeyAB / darknet                    (YOLO License v2 = public domain)
+  yolov4-leaky-416.weights  (245.78 MiB)
+  yolov4-leaky-416.cfg
         │
-        │  conversion via david8862's scripts
+        │  parsed and loaded by:
         ▼
-david8862 / keras-YOLOv3-model-set     (MIT)
+gwinndr / YOLOv4-Pytorch (MIT)         (used as conversion tool)
+  utilities/configs.py::parse_config
+  utilities/weights.py::load_weights
+  + DarknetRaw wrapper (this repo, MIT) that captures pre-YoloLayer outputs
         │
-        │  packaged in
+        │  torch.onnx.export(opset=13)
         ▼
-Xilinx / Vitis-AI Model Zoo (tf_yolov4-416_3.5)   (Apache 2.0)
+yolov4-leaky-416_float.onnx           (this repo, MIT)
+  3 raw outputs:
+    out_stride8   — shape (1, 255, 52, 52)
+    out_stride16  — shape (1, 255, 26, 26)
+    out_stride32  — shape (1, 255, 13, 13)
         │
-        │  source of the float .pb used as input
+        │  onnxruntime.quantize_static (MIT, used as tool)
+        │  + COCO val2017 calibration (1000 images, CC BY 4.0)
         ▼
-yolov4_float.onnx       (this repo, MIT — derivative)
-        │
-        │  INT8 static quantization with COCO val2017 calibration
-        ▼
-yolov4_int8_qop.onnx    (this repo, MIT — derivative)
+yolov4-leaky-416_int8_qop.onnx        (this repo, MIT)
+
+External Python decoder (this repo, MIT) reproduces standard YOLOv4
+post-processing: sigmoid + scale_xy + grid offset + anchor multiplication
++ NMS. The decoder lives outside the ONNX graph.
 ```
 
 ---
@@ -40,55 +51,38 @@ yolov4_int8_qop.onnx    (this repo, MIT — derivative)
 
 - **Project**: `AlexeyAB/darknet`
 - **Source**: https://github.com/AlexeyAB/darknet
-- **License**: YOLO License (effectively MIT-compatible; permits redistribution and modification)
-- **What we use**: the trained `yolov4-leaky-416.weights` file (indirectly, via the converted `.pb`).
+- **Wiki listing**: https://github.com/AlexeyAB/darknet/wiki/YOLOv4-model-zoo
+- **License**: YOLO License v2 — *"Darknet is public domain. Do whatever you want with it."*
+- **What we use**: the trained `yolov4-leaky-416.weights` file (245.78 MiB) and the corresponding `yolov4-leaky-416.cfg`. No modifications.
 
-### 2. Darknet → TensorFlow conversion — david8862
+### 2. Darknet → PyTorch conversion — gwinndr / YOLOv4-Pytorch
 
-- **Project**: `david8862/keras-YOLOv3-model-set`
-- **Source**: https://github.com/david8862/keras-YOLOv3-model-set
-- **License**: MIT
-- **What we use**: the `convert.py` and `keras_to_tensorflow.py` pipeline that produced the `.pb` we started from. We did not run these scripts ourselves; we used the `.pb` already generated by Xilinx's packaging.
+- **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.
+- We do NOT redistribute gwinndr's source code in this HF repo. It is a build-time tool.
 
-### 3. Float `.pb` packaging — Xilinx / Vitis-AI
+### 3. ONNX export
 
-- **Project**: `Xilinx/Vitis-AI` — Model Zoo entry `tf_yolov4-416_3.5`
-- **Source**: https://github.com/Xilinx/Vitis-AI/tree/master/model_zoo/model-list/tf_yolov4-416_3.5
-- **License**: Apache License 2.0
-- **What we use**: the file `float/yolov4-leaky-416.pb` (245 MB) packaged inside the model-zoo archive. We did not redistribute the archive itself — only the derivative ONNX models produced from that `.pb`.
-- **Note**: Per the README inside that model-zoo entry, the `.pb` was produced by running david8862's MIT-licensed conversion scripts on AlexeyAB's MIT-licensed weights. Xilinx did not train this model; they packaged it.
+- **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 3 raw 4D tensors.
 
-### 4. Calibration dataset — COCO val2017
+### 4. INT8 quantization
 
-- **Source**: https://cocodataset.org
-- **License**: Creative Commons Attribution 4.0 (CC BY 4.0) for the images; annotations are CC BY 4.0.
-- **What we use**: a small subset of `val2017` images was used **only as runtime input** to drive the activation calibrator inside `onnxruntime.quantization.quantize_static`. No COCO image is embedded inside the ONNX file; the dataset's role ends after calibration.
-
-### 5. Quantization tool — ONNX Runtime
-
-- **Project**: `microsoft/onnxruntime`
-- **Source**: https://github.com/microsoft/onnxruntime
-- **License**: MIT
-- **What we use**: the `onnxruntime.quantization.quantize_static` API with the following settings:
+- **Tool**: `onnxruntime.quantization.quantize_static` (Microsoft, MIT). Used as a tool, not redistributed.
+- **Configuration**:
   - `quant_format = QuantFormat.QOperator`
-  - `weight_type = QuantType.QInt8`
-  - `activation_type = QuantType.QInt8`
+  - `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 in the ONNX file.
 
----
+### 5. Calibration dataset — COCO val2017
 
-## Apache 2.0 compliance summary (Vitis-AI)
-
-Apache License 2.0 (Section 4) requires, when redistributing derivative
-works, that we:
-
-| Requirement | How we comply |
-|---|---|
-| Give recipients a copy of the License | Apache 2.0 is referenced in this NOTICE.md (full text at https://www.apache.org/licenses/LICENSE-2.0) |
-| State changes made | We changed: `.pb` → ONNX float (via `tf2onnx`), then ONNX float → INT8 ONNX (via `onnxruntime.quantize_static` with COCO calibration). No weights were modified by hand. |
-| Preserve copyright/attribution notices | This NOTICE.md preserves attribution to Xilinx, david8862 and AlexeyAB |
-| Carry forward the NOTICE file | Included as `NOTICE.md` in this repo |
+- **Source**: https://cocodataset.org
+- **License**: Creative Commons Attribution 4.0 (CC BY 4.0).
 
 ---
 
@@ -96,13 +90,9 @@ works, that we:
 
 | File | Size | SHA-256 |
 |---|---:|---|
-| `yolov4_float.onnx`           | 245.46 MB | `a168d4c5e07e83b393c3592822ea003aa8ff55d4606d2b29604d29741080aee9` |
-| `yolov4_int8_qop.onnx`        | 61.62 MB  | `0a4ccbea7081368b7e336b78d7dfefd83ec725a7bfa5cbdb825a917cb228e774` |
-| `quantize_float_to_int8.py`   | ~8 KB     | (see git commit hash on HF for the canonical version) |
-
-The float ONNX `yolov4-leaky-416.pb` (245.96 MB) is **not** rehosted here —
-it lives in the Xilinx Vitis-AI Model Zoo. Its SHA-256 at the time of this
-work was `e2428d0d890feb2e245b37899f7bcea22300b24f3b768aafc6fbb0c890342686`.
+| `yolov4-leaky-416.weights` (AlexeyAB upstream) | 257,717,640 B | `5b5b359940fd91e6d35bb8a957f6a8b27a05316889319173d39d25f8f33c0640` |
+| `yolov4-leaky-416_float.onnx`                  | 257,388,314 B | `d7277fc1c6522cb063999d2d72058fb15de6f15900c66d0093d535df0bcf200f` |
+| `yolov4-leaky-416_int8_qop.onnx`               |  64,655,943 B | `ca31b2c53227518f1e29cb50e59294e758b69de26f33e374f1e65c922d338da4` |
 
 ---
 
@@ -110,15 +100,15 @@ work was `e2428d0d890feb2e245b37899f7bcea22300b24f3b768aafc6fbb0c890342686`.
 
 **Pablo Mendoza** — HuggingFace [`@thefalley`](https://huggingface.co/thefalley)
 
-Source code that consumes this model:
+Companion repositories:
 
-| Repository | Status | Purpose |
-|---|---|---|
-| [`thefalley/yolov4-dpu-toolchain`](https://github.com/thefalley/yolov4-dpu-toolchain) | 🚧 **Pending upload** | Python toolchain that downloads this ONNX and produces firmware-ready `weights.bin` + `layer_configs.json`. |
-| [`thefalley/dpu-firmware`](https://github.com/thefalley/dpu-firmware) | 🚧 **Pending upload** | Bare-metal C firmware + RTL (SystemVerilog) for the custom DPU on ZedBoard XC7Z020. |
+| Repository | Purpose |
+|---|---|
+| [`Thefalley/yolov4-tiny-416-int8-qop`](https://huggingface.co/Thefalley/yolov4-tiny-416-int8-qop) | Smaller sibling: YOLOv4-tiny-416 INT8 (5.83 MiB) — fast embedded inference |
+| `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.
+incomplete, please open an issue on this repository and it will be
+corrected promptly.