Coral (Edge TPU) builds of yunet-2023mar and yolox-tiny-plate (CamStack D710 addendum)

faceDetection/yunet/README.md

@@ -17,6 +17,8 @@ Selectable on the `face-detection` step as `yunet-2023mar`.
 | ONNX | `onnx/camstack-yunet-2023mar.onnx` | fp32 |
 | OpenVINO IR | `openvino/camstack-yunet-2023mar.{xml,bin}` | fp32 |
 | CoreML | `coreml/camstack-yunet-2023mar.mlpackage` | fp16 |
+| TFLite (Coral Edge TPU) | `tflite/camstack-yunet-2023mar_edgetpu.tflite` | int8 |
+| TFLite (CPU, the Edge TPU file's source) | `tflite/camstack-yunet-2023mar_int8.tflite` | int8 |
 
 ## Source
 
@@ -38,6 +40,13 @@ named `cls_{8,16,32}`, `obj_{8,16,32}`, `bbox_{8,16,32}`, `kps_{8,16,32}`.
 The OpenVINO and CoreML files are converted from that patched ONNX. Built by
 `scripts/build-yunet-model.py` in the CamStack repository.
 
+The TFLite files are built by `scripts/build-yunet-edgetpu.py` from the patched
+ONNX: the two-node prefix is folded into the first convolution (the same
+function — the Edge TPU has no Gather), the graph is quantised to full integer
+(int8, per-channel weights; input uint8 with scale 1/255, so raw RGB bytes are
+the [0,1] tensor) on CamStack person crops, and the Edge TPU file is that graph
+compiled with `edgetpu_compiler` 14.1. Output names are unchanged.
+
 ## Training data
 
 YuNet was trained on **WIDER FACE** (<http://shuoyang1213.me/WIDERFACE/>),

faceDetection/yunet/tflite/camstack-yunet-2023mar_edgetpu.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7214dafaf18b8a5c55c7ea502c783e6a356043c2900ff9e02515a368285e982
+size 714256

faceDetection/yunet/tflite/camstack-yunet-2023mar_int8.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a369663e42539a7dc2b4152a53e7e5b6386215335d5cb979b81c44ab50b258e6
+size 128640

plateDetection/yolox-tiny-plate/README.md

@@ -18,9 +18,8 @@ The `plate-detection` step's only model, catalog id `yolox-tiny-plate`.
 | ONNX | `onnx/camstack-yolox-tiny-plate.onnx` | fp32 |
 | OpenVINO IR | `openvino/camstack-yolox-tiny-plate.{xml,bin}` | fp16 weights |
 | CoreML | `coreml/camstack-yolox-tiny-plate.mlpackage` | fp16 |
-
-No TFLite / Edge TPU build exists: YOLOX's Focus stem does not compile for the
-Edge TPU, and a Coral node runs this step on another device.
+| TFLite (Coral Edge TPU) | `tflite/camstack-yolox-tiny-plate_edgetpu.tflite` | int8 |
+| TFLite (CPU, the Edge TPU file's source) | `tflite/camstack-yolox-tiny-plate_int8.tflite` | int8 |
 
 ## Source
 
@@ -46,6 +45,15 @@ graph its BGR 0..255 tensor. The upstream output `[1, 8400, 6]`
 edited ONNX. Built by `scripts/build-yolox-plate-model.py` in the CamStack
 repository; see `SOURCE.md`.
 
+The TFLite files are built by `scripts/build-yolox-plate-edgetpu.py` from that
+edited ONNX: YOLOX's Focus stem (four strided slices and a concat) is rewritten
+as the mathematically equivalent 6x6 stride-2 convolution, the input prefix is
+folded into it, and the graph ends at the raw head — per stride `s`,
+`reg_s` (dx, dy, log w, log h) and `score_s` (obj*cls), which the CamStack
+`yolo` postprocessor decodes. It is quantised to full integer (int8,
+per-channel weights; input uint8 with scale 1/255) on CamStack vehicle crops,
+and the Edge TPU file is that graph compiled with `edgetpu_compiler` 14.1.
+
 ## Training data
 
 **Unknown terms.** Per the upstream NOTICE and manifest: an annotated

plateDetection/yolox-tiny-plate/tflite/camstack-yolox-tiny-plate_edgetpu.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e091263e9c93063f9da822387860a9cd21c0a554e714de276140921487b8e4d
+size 6865936

plateDetection/yolox-tiny-plate/tflite/camstack-yolox-tiny-plate_int8.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cca36a2f81d8788685fa9374a849783e9bc6a3aaea96fa85531a7e6bb3310f09
+size 5331536