# Amba - darknet to amba
###### tags: `darknet`
## initial environment gcc

source SDK2.5/Package/cv2x_linux_sdk_2.5/ambarella//build/env/aarch64-linaro-gcc.env

> TOOLCHAIN PATH: /usr/local/linaro-aarch64-2018.08-gcc8.2
TOOLCHAIN NANE: aarch64-linux-gnu-gcc
TOOLCHAIN VERSION: 8.2.1

## initial environment cv25 cnn

source /usr/local/amba-cv-tools-2.1.8-20200117.ubuntu-18.04/env/cv25.env
source  SDK2.5/CV2x_Toolchain/cv2x_cnngen_samples_20200304/build/env/cv25.env

> =================================================
> AmbaCnnUtils Version: cv2.1.8.2.550
CnnUtils Version: cv2.1.8.2.550
Cavalry Version: version 2.1.7 ( HASH c5db5f1 )
Vas Version: cv22.1.8.2.550
=================================================
============================================================
CV TOOLCHAIN PATH: /usr/local/amba-cv-tools-2.1.8-20200117.ubuntu-18.04
CV TOOLCHAIN VERSION: 2.1.8-20200117

## darknet to amba

#### 有frozen graph 檔, 剪枝output layer
用 netorn 看input/output layer name

graph_surgery.py tf -p (convert_weights_pb_jl.py 產出的檔案).pb -o tiny_yolov3_frozen_exhaust_0811.pb -isrc "i:inputs|is:416,416,3" -on "detector/yolo-v3-tiny/Conv_10/Conv2D,detector/yolo-v3-tiny/Conv_11/Conv2D"

#### 檔案放置

cd SDK2.5/CV2x_Toolchain/cv2x_cnngen_samples_20200304/tensorflow/demo_networks/tiny_yolo_v3
# model 放置處
cd models
# 影像資料放置處(jpg檔) 大約200張
cd dra_img

#### 轉 amba

cd SDK2.5/CV2x_Toolchain/cv2x_cnngen_samples_20200304/
vim .config

.config file 內容直接貼上

BUILD_CNNGEN_TENSORFLOW_NETWORKS=y
BUILD_CNNGEN_TENSORFLOW_DEMO_NETWORKS=y
BUILD_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3=y
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_MODEL="tiny_yolov3_frozen_exhaust_0811.pb"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_PARSER_OPTION="-c act-force-fx16,coeff-force-fx16"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_IN_NODE="inputs"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_IN_H_W="416,416"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_IN_MEAN="0.0,0.0,0.0"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_IN_SCALE="1"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_IN_IS_BGR=0
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_OUT_NODE="detector/yolo-v3-tiny/Conv_10/Conv2D,detector/yolo-v3-tiny/Conv_11/Conv2D"
CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_OUT_FORMAT="fp32"

> 精度調整
> CONFIG_CNNGEN_TENSORFLOW_DEMO_TINY_YOLOV3_PARSER_OPTION
> int16: act-force-fx16,coeff-force-fx16
> int8: act-force-fx8,coeff-force-fx8
> 詳細自己看 spec SDK2.5/Document/Ambarella_CV2x_Flexible_Linux_SDK2.5.6_Document_VIVOTEK/Ambarella_CV2x_UG_Flexible_Linux_SDK2.5_CNNGen_Development_Flow.pdf

cavalry mode (轉換成 amba device 讀的模型檔)

make run_mode=cavalry

> 產出位置 out/tensorflow/demo_networks/tf_tiny_yolov3/cavalry_tf_tiny_yolov3/tf_tiny_yolov3_cavalry.bin