Quantization support

Author: daniil-lyakhov

backends/openvino/README.md

@@ -46,7 +46,8 @@ Before you begin, ensure you have openvino installed and configured on your syst
 
 ```bash
 git clone https://github.com/daniil-lyakhov/openvino.git
-cd openvino && git checkout dl/executorch/yolo12
+# WIP
+cd openvino && git checkout dl/yolo12_complete_support
 git submodule update --init --recursive
 sudo ./install_build_dependencies.sh
 mkdir build && cd build

examples/models/yolo12/CMakeLists.txt

@@ -32,6 +32,7 @@ set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 find_package(executorch CONFIG REQUIRED PATHS ${EXECUTORCH_ROOT}/cmake-out)
 target_link_options_shared_lib(executorch)
 
+add_subdirectory(${EXECUTORCH_ROOT}/third-party/gflags gflags)
 set(link_libraries gflags)
 list(APPEND link_libraries portable_ops_lib portable_kernels)
 target_link_options_shared_lib(portable_ops_lib)

examples/models/yolo12/README.md

@@ -10,6 +10,17 @@ This example demonstrates how to perform inference of [Ultralytics YOLO12 family
 - [OpenVINO](../../../backends/openvino/README.md)
 - [XNNPACK](../../../backends/xnnpack/README.md)
 
+# Performance Evaluation
+
+| CPU                            | Model   | Backend  | Device | Precision | Average Latency, ms |
+|--------------------------------|---------|----------|--------|-----------|-----------------|
+| Intel(R) Core(TM) Ultra 7 155H | yolo12s | openvino | CPU    | FP32      | 88.3549         |
+| Intel(R) Core(TM) Ultra 7 155H | yolo12s | openvino | CPU    | INT8      | 53.066          |
+| Intel(R) Core(TM) Ultra 7 155H | yolo12l | openvino | CPU    | FP32      | 317.953         |
+| Intel(R) Core(TM) Ultra 7 155H | yolo12s | openvino | CPU    | INT8      | 150.846         |
+| Intel(R) Core(TM) Ultra 7 155H | yolo12s | xnnpack  | CPU    | FP32      | 169.36          |
+| Intel(R) Core(TM) Ultra 7 155H | yolo12l | xnnpack  | CPU    | FP32      | 436.876         |
+
 
 # Instructions
 
@@ -42,12 +53,16 @@ OpenVINO:
 python export_and_validate.py --model_name yolo12s --input_dims=[1920,1080]  --backend openvino --device CPU
 ```
 
+OpenVINO quantized model:
+```bash
+python export_and_validate.py --model_name yolo12s --input_dims=[1920,1080]  --backend openvino --quantize --video_input /path/to/calibration/video --device CPU
+```
+
 XNNPACK:
 ```bash
 python export_and_validate.py --model_name yolo12s --input_dims=[1920,1080] --backend xnnpack
 ```
 
-> **_NOTE:_**  Quantization is comming soon!
 
 Exported model could be validated using the `--validate` key:
 

examples/models/yolo12/export_and_validate.py

@@ -228,7 +228,6 @@ def main(
     model = YOLO(model_name)
 
     if quantize:
-        raise NotImplementedError("Quantization is comming soon!")
         if video_path is None:
             raise RuntimeError(
                 "Could not quantize model without the video for the calibration."