零基础导入自己YOLOv3或YOLOv3-Tiny模型

前言：通过 零基础入门darknet-YOLO3及YOLOv3-Tiny 文档操作，我们已经有了自己训练的YOLO3或YOLOv3-Tiny模型，接下来一步步演示如何转换自己的模型，并适配进我们khadas VIM3 开发板 android平台的khadas_android_npu_app。关于VIM3 ubuntu平台请参考此文档 1。

一，转换模型

模型转换流程图：

image1942×605 164 KB

1，导入模型

当前模型转换过程都是在 acuity-toolkit 目录下进行:

cd {workspace}/aml_npu_sdk/acuity-toolkit
cp {workspace}/yolov3-khadas_ai.cfg_train demo/model/
cp {workspace}/yolov3-khadas_ai_last.weights demo/model/
cp {workspace}/test.jpg demo/model/

如下修改：

hlm@Server:/users/hlm/npu/aml_npu_sdk/acuity-toolkit/demo$ git diff
diff --git a/acuity-toolkit/demo/0_import_model.sh b/acuity-toolkit/demo/0_import_model.sh
index 3198810..4671c9f 100755
--- a/acuity-toolkit/demo/0_import_model.sh
+++ b/acuity-toolkit/demo/0_import_model.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-NAME=mobilenet_tf
+NAME=yolov3
 ACUITY_PATH=../bin/
 
 convert_caffe=${ACUITY_PATH}convertcaffe
@@ -11,13 +11,13 @@ convert_onnx=${ACUITY_PATH}convertonnx
 convert_keras=${ACUITY_PATH}convertkeras
 convert_pytorch=${ACUITY_PATH}convertpytorch
 
-$convert_tf \
-    --tf-pb ./model/mobilenet_v1.pb \
-    --inputs input \
-    --input-size-list '224,224,3' \
-    --outputs MobilenetV1/Predictions/Softmax \
-    --net-output ${NAME}.json \
-    --data-output ${NAME}.data 
+#$convert_tf \
+#    --tf-pb ./model/mobilenet_v1.pb \
+#    --inputs input \
+#    --input-size-list '224,224,3' \
+#    --outputs MobilenetV1/Predictions/Softmax \
+#    --net-output ${NAME}.json \
+#    --data-output ${NAME}.data 
        
 #$convert_caffe \
 #    --caffe-model xx.prototxt   \
@@ -30,11 +30,11 @@ $convert_tf \
 #    --net-output ${NAME}.json \
 #    --data-output ${NAME}.data 
 
-#$convert_darknet \
-#    --net-input xxx.cfg \
-#      --weight-input xxx.weights \
-#    --net-output ${NAME}.json \
-#    --data-output ${NAME}.data 
+$convert_darknet \
+    --net-input ./model/yolov3-khadas_ai.cfg_train \
+    --weight-input ./model/yolov3-khadas_ai_last.weights \
+    --net-output ${NAME}.json \
+    --data-output ${NAME}.data 

--- a/acuity-toolkit/demo/data/validation_tf.txt
+++ b/acuity-toolkit/demo/data/validation_tf.txt
@@ -1 +1 @@
-./space_shuttle_224.jpg, 813
+./test.jpg

执行对应脚本：

bash 0_import_model.sh 

image1616×198 13.1 KB

2，对模型进行量化

hlm@Server:/users/hlm/npu/aml_npu_sdk/acuity-toolkit/demo$ git diff
diff --git a/acuity-toolkit/demo/1_quantize_model.sh b/acuity-toolkit/demo/1_quantize_model.sh
index 630ea7f..ee7bd00 100755
--- a/acuity-toolkit/demo/1_quantize_model.sh
+++ b/acuity-toolkit/demo/1_quantize_model.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-NAME=mobilenet_tf
+NAME=yolov3
 ACUITY_PATH=../bin/
 
 tensorzone=${ACUITY_PATH}tensorzonex
@@ -11,12 +11,12 @@ $tensorzone \
     --dtype float32 \
     --source text \
     --source-file data/validation_tf.txt \
-    --channel-mean-value '128 128 128 128' \
-    --reorder-channel '0 1 2' \
+    --channel-mean-value '0 0 0 256' \
+    --reorder-channel '2 1 0' \
     --model-input ${NAME}.json \
     --model-data ${NAME}.data \
     --model-quantize ${NAME}.quantize \
-    --quantized-dtype asymmetric_affine-u8 \
+    --quantized-dtype dynamic_fixed_point-i8 \

注意，这里 --quantized-dtype 跟ubuntu平台不一样。执行对应脚本：

bash 1_quantize_model.sh

image1131×240 11.6 KB

3，生成 case 代码

hlm@Server:/users/hlm/npu/aml_npu_sdk/acuity-toolkit/demo$ git diff
diff --git a/acuity-toolkit/demo/2_export_case_code.sh b/acuity-toolkit/demo/2_export_case_code.sh
index 85b101b..867c5b9 100755
--- a/acuity-toolkit/demo/2_export_case_code.sh
+++ b/acuity-toolkit/demo/2_export_case_code.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-NAME=mobilenet_tf
+NAME=yolov3
 ACUITY_PATH=../bin/
 
 export_ovxlib=${ACUITY_PATH}ovxgenerator
@@ -8,8 +8,8 @@ export_ovxlib=${ACUITY_PATH}ovxgenerator
 $export_ovxlib \
     --model-input ${NAME}.json \
     --data-input ${NAME}.data \
-    --reorder-channel '0 1 2' \
-    --channel-mean-value '128 128 128 128' \
+    --reorder-channel '2 1 0' \
+    --channel-mean-value '0 0 0 256' \

执行对应脚本：

bash 2_export_case_code.sh

image1279×1231 83.8 KB

当然只要运行脚本没报错，后续你可以按照上面修改完成后，一次性执行脚本：

bash 0_import_model.sh && bash 1_quantize_model.sh  && bas