[rknn] update rkopt doc and code, now model is same for tk1/2

README.md

@@ -1,3 +1,9 @@
+#### Get model optimized for RKNN
+Exports model with optimization for RKNN, please refer here [README_rkopt_manual.md](./README_rkopt_manual.md)
+
+---
+<br>
+
 <div align="center">
 <p>
    <a align="left" href="https://ultralytics.com/yolov5" target="_blank">

README_rkopt_manual.md

@@ -1,16 +1,35 @@
-# YOLOv5 - rkopt 仓库
+# YOLOv5 - RKNN optimize
 
-- 基于 https://github.com/ultralytics/yolov5 代码修改，设配 rknpu 设备的部署优化
-- 切换分支 git checkout {分支名}
-- 目前支持分支:
-  - master
-    - maxpool/ focus 优化，输出改为个branch分支的输出。以上优化代码使用插入宏实现，不影响原来的训练逻辑，这个优化兼容修改前的权重，故支持官方给的预训练权重。
+## Source
 
-    - 修改激活函数 silu 为 relu
+  Base on https://github.com/ultralytics/yolov5 with commit id as d3ea0df8b9f923685ce5f2555c303b8eddbf83fd
 
-    - 训练的相关内容请参考 README.md 说明
 
-    - 导出模型时  python export.py --rknpu {rk_platform} 即可导出优化模型
 
-      (rk_platform支持 rk1808, rv1109, rv1126, rk3399pro, rk3566, rk3568, rk3588, rv1103, rv1106)
+## What different
+
+Inference result unchanged:
+
+- Optimize focus/SPPF block, getting better performance with same result
+- Change output node, remove post_process from the model. (post process is unfriendly in quantization)
+
+
+
+Inference result changed:
+
+- Using ReLU as activation layer instead of SiLU（Only valid when training new model）
+
+
+
+## How to use
+
+```
+python export.py --rknpu {rk_platform} --weight yolov5s.pt
+```
+
+- rk_platform support  rk1808, rv1109, rv1126, rk3399pro, rk3566, rk3562, rk3568, rk3588, rv1103, rv1106. (Actually the exported models are the same in spite of the exact platform )
+
+- the 'yolov5s.pt' could be replace with your model path
+- A file name "RK_anchors.txt" would be generated and it could be use during doing post_process in the outside. 
+- **NOTICE: Please call with --rknpu param, do not changing the default rknpu value in export.py.** 
 

export.py

@@ -526,7 +526,7 @@ def run(
             m.onnx_dynamic = dynamic
             m.export = True
 
-        if os.getenv('RKNN_model_hack', '0') == 'npu_1':
+        if os.getenv('RKNN_model_hack', '0') in ['npu_1','npu_2']:
             from models.common import Focus
             from models.common import Conv
             from models.common_rk_plug_in import surrogate_focus

models/common.py

@@ -232,7 +232,7 @@ def forward(self, x):
             return self.cv2(torch.cat(y, 1))
 
 
-if os.getenv('RKNN_model_hack', '0') in ['0','npu_2']:
+if os.getenv('RKNN_model_hack', '0') in ['0']:
     class SPPF(nn.Module):
         # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
         def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))
@@ -249,7 +249,7 @@ def forward(self, x):
                 y1 = self.m(x)
                 y2 = self.m(y1)
                 return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
-elif os.getenv('RKNN_model_hack', '0') == 'npu_1':
+elif os.getenv('RKNN_model_hack', '0') in ['npu_1','npu_2']:
     class SPPF(nn.Module):
         # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
         def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))