yolov5 训练一个模型，转成fp16用NCNN部署到移动端，发现ncnn侧得分超低，并位置都不对

[ouxiand]

1.yolov5 训练一个车牌检测模型，转成fp16用NCNN部署到移动端 python3 export.py --weights /yolo/yolov5/weights/best.pt --include onnx --train python3 export.py --weights /yolo/yolov5/weights/best.pt --img 320 --include onnx --train python3 -m onnxsim /yolo/yolov5/weights/best.onnx /yolo/yolov5/weights/best-sim.onnx ./weights/onnx2ncnn /yolo/yolov5/weights/best-sim.onnx /yolo/yolov5/weights/best.param /yolo/yolov5/weights/best.bin ./weights/ncnnoptimize /yolo/yolov5/weights/best.param /yolo/yolov5/weights/best.bin /yolo/yolov5/weights/best-opt.param /yolo/yolov5/weights/best-opt.bin 65536

2.在安卓用NCNN部署 修改了：以下参数匹配模型输出参数 ex.input("images", in_pad); ex.extract("output", out); ex.extract("353", out); ex.extract("367", out); 参数匹配yolov5n的训练参数

发现同一张图片（车牌做了打码处理）在ncnn侧得分超低，并位置都不对！ 云端检测得分

NCNN检测得分

[ouxiand]

补充说明： 1.云端是用1080P的图片训练 imgsize是320 2.NCNN用320检测1080P的图片

3. param如下：

7767517 173 197 Input images 0 1 images Convolution Conv_0 1 1 images 122 0=16 1=6 3=2 4=2 5=1 6=1728 Swish Mul_2 1 1 122 124 Convolution Conv_3 1 1 124 125 0=32 1=3 3=2 4=1 5=1 6=4608 Swish Mul_5 1 1 125 127 Split splitncnn_0 1 2 127 127_splitncnn_0 127_splitncnn_1 Convolution Conv_6 1 1 127_splitncnn_1 128 0=16 1=1 5=1 6=512 Swish Mul_8 1 1 128 130 Split splitncnn_1 1 2 130 130_splitncnn_0 130_splitncnn_1 Convolution Conv_9 1 1 130_splitncnn_1 131 0=16 1=1 5=1 6=256 Swish Mul_11 1 1 131 133 Convolution Conv_12 1 1 133 134 0=16 1=3 4=1 5=1 6=2304 Swish Mul_14 1 1 134 136 BinaryOp Add_15 2 1 130_splitncnn_0 136 137 Convolution Conv_16 1 1 127_splitncnn_0 138 0=16 1=1 5=1 6=512 Swish Mul_18 1 1 138 140 Concat Concat_19 2 1 137 140 141 Convolution Conv_20 1 1 141 142 0=32 1=1 5=1 6=1024 Swish Mul_22 1 1 142 144 Convolution Conv_23 1 1 144 145 0=64 1=3 3=2 4=1 5=1 6=18432 Swish Mul_25 1 1 145 147 Split splitncnn_2 1 2 147 147_splitncnn_0 147_splitncnn_1 Convolution Conv_26 1 1 147_splitncnn_1 148 0=32 1=1 5=1 6=2048 Swish Mul_28 1 1 148 150 Split splitncnn_3 1 2 150 150_splitncnn_0 150_splitncnn_1 Convolution Conv_29 1 1 150_splitncnn_1 151 0=32 1=1 5=1 6=1024 Swish Mul_31 1 1 151 153 Convolution Conv_32 1 1 153 154 0=32 1=3 4=1 5=1 6=9216 Swish Mul_34 1 1 154 156 BinaryOp Add_35 2 1 150_splitncnn_0 156 157 Split splitncnn_4 1 2 157 157_splitncnn_0 157_splitncnn_1 Convolution Conv_36 1 1 157_splitncnn_1 158 0=32 1=1 5=1 6=1024 Swish Mul_38 1 1 158 160 Convolution Conv_39 1 1 160 161 0=32 1=3 4=1 5=1 6=9216 Swish Mul_41 1 1 161 163 BinaryOp Add_42 2 1 157_splitncnn_0 163 164 Convolution Conv_43 1 1 147_splitncnn_0 165 0=32 1=1 5=1 6=2048 Swish Mul_45 1 1 165 167 Concat Concat_46 2 1 164 167 168 Convolution Conv_47 1 1 168 169 0=64 1=1 5=1 6=4096 Swish Mul_49 1 1 169 171 Split splitncnn_5 1 2 171 171_splitncnn_0 171_splitncnn_1 Convolution Conv_50 1 1 171_splitncnn_1 172 0=128 1=3 3=2 4=1 5=1 6=73728 Swish Mul_52 1 1 172 174 Split splitncnn_6 1 2 174 174_splitncnn_0 174_splitncnn_1 Convolution Conv_53 1 1 174_splitncnn_1 175 0=64 1=1 5=1 6=8192 Swish Mul_55 1 1 175 177 Split splitncnn_7 1 2 177 177_splitncnn_0 177_splitncnn_1 Convolution Conv_56 1 1 177_splitncnn_1 178 0=64 1=1 5=1 6=4096 Swish Mul_58 1 1 178 180 Convolution Conv_59 1 1 180 181 0=64 1=3 4=1 5=1 6=36864 Swish Mul_61 1 1 181 183 BinaryOp Add_62 2 1 177_splitncnn_0 183 184 Split splitncnn_8 1 2 184 184_splitncnn_0 184_splitncnn_1 Convolution Conv_63 1 1 184_splitncnn_1 185 0=64 1=1 5=1 6=4096 Swish Mul_65 1 1 185 187 Convolution Conv_66 1 1 187 188 0=64 1=3 4=1 5=1 6=36864 Swish Mul_68 1 1 188 190 BinaryOp Add_69 2 1 184_splitncnn_0 190 191 Split splitncnn_9 1 2 191 191_splitncnn_0 191_splitncnn_1 Convolution Conv_70 1 1 191_splitncnn_1 192 0=64 1=1 5=1 6=4096 Swish Mul_72 1 1 192 194 Convolution Conv_73 1 1 194 195 0=64 1=3 4=1 5=1 6=36864 Swish Mul_75 1 1 195 197 BinaryOp Add_76 2 1 191_splitncnn_0 197 198 Convolution Conv_77 1 1 174_splitncnn_0 199 0=64 1=1 5=1 6=8192 Swish Mul_79 1 1 199 201 Concat Concat_80 2 1 198 201 202 Convolution Conv_81 1 1 202 203 0=128 1=1 5=1 6=16384 Swish Mul_83 1 1 203 205 Split splitncnn_10 1 2 205 205_splitncnn_0 205_splitncnn_1 Convolution Conv_84 1 1 205_splitncnn_1 206 0=256 1=3 3=2 4=1 5=1 6=294912 Swish Mul_86 1 1 206 208 Split splitncnn_11 1 2 208 208_splitncnn_0 208_splitncnn_1 Convolution Conv_87 1 1 208_splitncnn_1 209 0=128 1=1 5=1 6=32768 Swish Mul_89 1 1 209 211 Split splitncnn_12 1 2 211 211_splitncnn_0 211_splitncnn_1 Convolution Conv_90 1 1 211_splitncnn_1 212 0=128 1=1 5=1 6=16384 Swish Mul_92 1 1 212 214 Convolution Conv_93 1 1 214 215 0=128 1=3 4=1 5=1 6=147456 Swish Mul_95 1 1 215 217 BinaryOp Add_96 2 1 211_splitncnn_0 217 218 Convolution Conv_97 1 1 208_splitncnn_0 219 0=128 1=1 5=1 6=32768 Swish Mul_99 1 1 219 221 Concat Concat_100 2 1 218 221 222 Convolution Conv_101 1 1 222 223 0=256 1=1 5=1 6=65536 Swish Mul_103 1 1 223 225 Convolution Conv_104 1 1 225 226 0=128 1=1 5=1 6=32768 Swish Mul_106 1 1 226 228 Split splitncnn_13 1 2 228 228_splitncnn_0 228_splitncnn_1 Pooling MaxPool_107 1 1 228_splitncnn_1 229 1=5 3=2 5=1 Split splitncnn_14 1 2 229 229_splitncnn_0 229_splitncnn_1 Pooling MaxPool_108 1 1 229_splitncnn_1 230 1=5 3=2 5=1 Split splitncnn_15 1 2 230 230_splitncnn_0 230_splitncnn_1 Pooling MaxPool_109 1 1 230_splitncnn_1 231 1=5 3=2 5=1 Concat Concat_110 4 1 228_splitncnn_0 229_splitncnn_0 230_splitncnn_0 231 232 Convolution Conv_111 1 1 232 233 0=256 1=1 5=1 6=131072 Swish Mul_113 1 1 233 235 Convolution Conv_114 1 1 235 236 0=128 1=1 5=1 6=32768 Swish Mul_116 1 1 236 238 Split splitncnn_16 1 2 238 238_splitncnn_0 238_splitncnn_1 Interp Resize_121 1 1 238_splitncnn_1 243 0=1 1=2.000000e+00 2=2.000000e+00 Concat Concat_122 2 1 243 205_splitncnn_0 244 Split splitncnn_17 1 2 244 244_splitncnn_0 244_splitncnn_1 Convolution Conv_123 1 1 244_splitncnn_1 245 0=64 1=1 5=1 6=16384 Swish Mul_125 1 1 245 247 Convolution Conv_126 1 1 247 248 0=64 1=1 5=1 6=4096 Swish Mul_128 1 1 248 250 Convolution Conv_129 1 1 250 251 0=64 1=3 4=1 5=1 6=36864 Swish Mul_131 1 1 251 253 Convolution Conv_132 1 1 244_splitncnn_0 254 0=64 1=1 5=1 6=16384 Swish Mul_134 1 1 254 256 Concat Concat_135 2 1 253 256 257 Convolution Conv_136 1 1 257 258 0=128 1=1 5=1 6=16384 Swish Mul_138 1 1 258 260 Convolution Conv_139 1 1 260 261 0=64 1=1 5=1 6=8192 Swish Mul_141 1 1 261 263 Split splitncnn_18 1 2 263 263_splitncnn_0 263_splitncnn_1 Interp Resize_146 1 1 263_splitncnn_1 268 0=1 1=2.000000e+00 2=2.000000e+00 Concat Concat_147 2 1 268 171_splitncnn_0 269 Split splitncnn_19 1 2 269 269_splitncnn_0 269_splitncnn_1 Convolution Conv_148 1 1 269_splitncnn_1 270 0=32 1=1 5=1 6=4096 Swish Mul_150 1 1 270 272 Convolution Conv_151 1 1 272 273 0=32 1=1 5=1 6=1024 Swish Mul_153 1 1 273 275 Convolution Conv_154 1 1 275 276 0=32 1=3 4=1 5=1 6=9216 Swish Mul_156 1 1 276 278 Convolution Conv_157 1 1 269_splitncnn_0 279 0=32 1=1 5=1 6=4096 Swish Mul_159 1 1 279 281 Concat Concat_160 2 1 278 281 282 Convolution Conv_161 1 1 282 283 0=64 1=1 5=1 6=4096 Swish Mul_163 1 1 283 285 Split splitncnn_20 1 2 285 285_splitncnn_0 285_splitncnn_1 Convolution Conv_164 1 1 285_splitncnn_1 286 0=64 1=3 3=2 4=1 5=1 6=36864 Swish Mul_166 1 1 286 288 Concat Concat_167 2 1 288 263_splitncnn_0 289 Split splitncnn_21 1 2 289 289_splitncnn_0 289_splitncnn_1 Convolution Conv_168 1 1 289_splitncnn_1 290 0=64 1=1 5=1 6=8192 Swish Mul_170 1 1 290 292 Convolution Conv_171 1 1 292 293 0=64 1=1 5=1 6=4096 Swish Mul_173 1 1 293 295 Convolution Conv_174 1 1 295 296 0=64 1=3 4=1 5=1 6=36864 Swish Mul_176 1 1 296 298 Convolution Conv_177 1 1 289_splitncnn_0 299 0=64 1=1 5=1 6=8192 Swish Mul_179 1 1 299 301 Concat Concat_180 2 1 298 301 302 Convolution Conv_181 1 1 302 303 0=128 1=1 5=1 6=16384 Swish Mul_183 1 1 303 305 Split splitncnn_22 1 2 305 305_splitncnn_0 305_splitncnn_1 Convolution Conv_184 1 1 305_splitncnn_1 306 0=128 1=3 3=2 4=1 5=1 6=147456 Swish Mul_186 1 1 306 308 Concat Concat_187 2 1 308 238_splitncnn_0 309 Split splitncnn_23 1 2 309 309_splitncnn_0 309_splitncnn_1 Convolution Conv_188 1 1 309_splitncnn_1 310 0=128 1=1 5=1 6=32768 Swish Mul_190 1 1 310 312 Convolution Conv_191 1 1 312 313 0=128 1=1 5=1 6=16384 Swish Mul_193 1 1 313 315 Convolution Conv_194 1 1 315 316 0=128 1=3 4=1 5=1 6=147456 Swish Mul_196 1 1 316 318 Convolution Conv_197 1 1 309_splitncnn_0 319 0=128 1=1 5=1 6=32768 Swish Mul_199 1 1 319 321 Concat Concat_200 2 1 318 321 322 Convolution Conv_201 1 1 322 323 0=256 1=1 5=1 6=65536 Swish Mul_203 1 1 323 325 Convolution Conv_204 1 1 285_splitncnn_0 326 0=36 1=1 5=1 6=2304 Reshape Reshape_216 1 1 326 338 0=-1 1=12 2=3 Permute Transpose_217 1 1 338 output 0=1 Convolution Conv_218 1 1 305_splitncnn_0 340 0=36 1=1 5=1 6=4608 Reshape Reshape_230 1 1 340 352 0=-1 1=12 2=3 Permute Transpose_231 1 1 352 353 0=1 Convolution Conv_232 1 1 325 354 0=36 1=1 5=1 6=9216 Reshape Reshape_244 1 1 354 366 0=-1 1=12 2=3 Permute Transpose_245 1 1 366 367 0=1

[deepage]

现在强烈建议采用pnnx来转换pytorch模型，这里贴一下nihui新写的使用教程https://zhuanlan.zhihu.com/p/471357671

[nihui]

针对onnx模型转换的各种问题，推荐使用最新的pnnx工具转换到ncnn In view of various problems in onnx model conversion, it is recommended to use the latest pnnx tool to convert your model to ncnn

pip install pnnx
pnnx model.onnx inputshape=[1,3,224,224]

详细参考文档 Detailed reference documentation https://github.com/pnnx/pnnx https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx#how-to-use-pnnx