how to convert mnet.25.prototxt and mnet.25.caffemodel to mnet-deconv.prototxt and mnet-deconv.caffemodel

[Royzon]

Thanks for your sharing. I convert mnet.25 mxnet model to caffe model successfully by your tools, but I find this caffe model can't used in your project directly, so may you pleasure to help me solve it?

[clancylian]

what's your problems?

[Royzon]

the error message is here: [100%] Linking CXX executable retinaface [100%] Built target retinaface batchSize:1, channel:3, netHeight:416, netWidth:288. Caffe Parser: Invalid axis in crop layer - only spatial cropping is supported error parsing layer type Crop index 143 retinaface: retinaface/tensorrt/trtnetbase.cpp:279: void TrtNetBase::caffeToTRTModel(const string&, const string&, nvcaffeparser1::IPluginFactory*): Assertion `blobNameToTensor != nullptr' failed. Aborted (core dumped)

and I find there are many differences, why the converted model can not be used directly? $ diff mnet25.prototxt mnet-deconv-0517.prototxt 7c7 < shape: { dim: 1 dim: 3 dim: 416 dim: 288 }

shape: { dim: 1 dim: 3 dim: 320 dim: 320 }

1208a1209 bias_term: true 1247a1249 bias_term: true 1279a1282 bias_term: true 1318a1322 bias_term: true 1350a1355 bias_term: true 1389a1395 bias_term: true 1437a1444 bias_term: true 1448,1453c1455,1460 < dim: 2 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:2
  dim:-1
  dim:0
 }

} 1470,1475c1477,1482 < dim: 4 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:4
  dim:-1
  dim:0
 }

} 1487a1495 bias_term: true 1500a1509 bias_term: true 1513a1523 bias_term: true 1544,1545d1553 < bottom: "rf_c3_lateral_relu" < top: "rf_c3_upsampling" 1547a1556,1557 bottom: "rf_c3_lateral_relu" top: "rf_c3_upsampling" 1549,1556c1559,1566 < num_output: 64 < kernel_size: 4 < stride: 2 < pad: 1 < group: 64 < bias_term: false < weight_filler: { < type: "bilinear"

  kernel_size: 4
  stride: 2
  pad: 1
  num_output: 64
  group: 64
  bias_term: false
  weight_filler {
 type: "bilinear"

1558a1569 param { lr_mult: 0 decay_mult: 0 } 1568,1569c1579 < axis: 1 < offset: 0

  axis: 2

1593a1604 bias_term: true 1632a1644 bias_term: true 1664a1677 bias_term: true 1703a1717 bias_term: true 1735a1750 bias_term: true 1774a1790 bias_term: true 1822a1839 bias_term: true 1833,1838c1850,1855 < dim: 2 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:2
  dim:-1
  dim:0
 }

} 1855,1860c1872,1877 < dim: 4 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:4
  dim:-1
  dim:0
 }

} 1872a1890 bias_term: true 1885a1904 bias_term: true 1898a1918 bias_term: true 1929,1930d1948 < bottom: "rf_c2_aggr_relu" < top: "rf_c2_upsampling" 1932a1951,1952 bottom: "rf_c2_aggr_relu" top: "rf_c2_upsampling" 1934,1941c1954,1961 < num_output: 64 < kernel_size: 4 < stride: 2 < pad: 1 < group: 64 < bias_term: false < weight_filler: { < type: "bilinear"

  kernel_size: 4
  stride: 2
  pad: 1
  num_output: 64
  group: 64
  bias_term: false
  weight_filler {
 type: "bilinear"

1943a1964 param { lr_mult: 0 decay_mult: 0 } 1953,1954c1974 < axis: 1 < offset: 0

  axis: 2

1978a1999 bias_term: true 2017a2039 bias_term: true 2049a2072 bias_term: true 2088a2112 bias_term: true 2120a2145 bias_term: true 2159a2185 bias_term: true 2207a2234 bias_term: true 2218,2223c2245,2250 < dim: 2 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:2
  dim:-1
  dim:0
 }

} 2240,2245c2267,2272 < dim: 4 < dim: -1 < dim: 0 < } < axis: 1
< }

  dim:0
  dim:4
  dim:-1
  dim:0
 }

} 2257a2285 bias_term: true 2270a2299 bias_term: true

[clancylian]

can you use caffe model directly instead of use tensorRT model ? if caffe model is not occur error , you should set shape in .prototxt the same as the code.

[flazerain]

can you use caffe model directly instead of use tensorRT model ? if caffe model is not occur error , you should set shape in .prototxt the same as the code.

我自己训的模型怎么能变成mnet-deconv那样,转换的时候报错: Caffe Parser: Invalid axis in crop layer - only spatial cropping is supported

[nuanxinqing]

@Royzon Hi where did you download the mobilenet25 mxnet model?could you give me the website?thank u

[JaydonChion]

@clancylian , I encountered this error when I tried using your mnet25.caffemodel and ment25.prototxt to do inferencing on DeepStream. Deepstream auto convert the caffe model to tensorRT by generating engine plan file. The caffe model itself is working fine, the error occurs when it comes to tensorRT. Any thoughts on what might be the reason? Thanks

[qustions]

@JaydonChion have you solve this?