topic Re: Compiler errors with DW COnvolution? in Intel® Distribution of OpenVINO™ Toolkit

Compiler errors with DW COnvolution?

idata — Mon, 12 Feb 2018 02:10:13 GMT

Hi guys, i have a model in caffe for a separable convolution like the one in Mobile Net, but the compiler make this strange error:

`mvNCCompile v02.00, Copyright @ Movidius Ltd 2016

Fusing depthconv and conv in conv2/dw and conv2/sep

Traceback (most recent call last):

File "/usr/local/bin/mvNCCompile", line 118, in

create_graph(args.network, args.inputnode, args.outputnode, args.outfile, args.nshaves, args.inputsize, args.weights)

File "/usr/local/bin/mvNCCompile", line 101, in create_graph

net = parse_caffe(args, myriad_config)

File "/usr/local/bin/ncsdk/Controllers/CaffeParser.py", line 1394, in parse_caffe

network.attach(node)

File "/usr/local/bin/ncsdk/Models/Network.py", line 81, in attach

stage.attach_several(appropriate_nodes)

File "/usr/local/bin/ncsdk/Models/NetworkStage.py", line 689, in attach_several

parents.attach(self)

File "/usr/local/bin/ncsdk/Models/NetworkStage.py", line 412, in attach

taps[c,c*multiplier+i,y,x] = self.taps[y,x,c,i]

IndexError: index 10 is out of bounds for axis 2 with size 10

I say strange because my Network has this architecture: data- (conv/dw - conv/sep- BN- scal&bias) for three times, than a FC layer and a SoftMax activation, so i don't understand why the compiler locked on the second Convolution Separable Convolution, any suggestion?

Re: Compiler errors with DW COnvolution?

idata — Tue, 13 Feb 2018 04:40:28 GMT

@Ryose If you could provide your model files, I'd like to help you debug your network. Thanks.

Re: Compiler errors with DW COnvolution?

idata — Tue, 13 Feb 2018 17:33:09 GMT

@Tome_at_Intel sure, here is my .prototxt. I hope this helps!

name: "UNIPINET"
layer {
  name: "data"
  type: "Input"
  top: "data"
  input_param { shape: { dim: 1 dim: 1 dim: 63 dim: 13 } }
}
layer {
  name: "conv1/dw"
  type: "Convolution"
  bottom: "data"
  top: "conv1/dw"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 1
    bias_term: false
    pad: 0
    kernel_h: 15 
    kernel_w: 3
    group: 1
    #engine: CAFFE
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}

layer {
  name: "conv1/sep"
  type: "Convolution"
  bottom: "conv1/dw"
  top: "conv1/sep"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 16
    bias_term: true
    pad: 0
    kernel_size: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv1/sep/bn"
  type: "BatchNorm"
  bottom: "conv1/sep"
  top: "conv1/sep"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv1/sep/bn/scale"
  type: "Scale"
  bottom: "conv1/sep"
  top: "conv1/sep"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "relu1/sep"
  type: "ReLU"
  bottom: "conv1/sep"
  top: "conv1/sep"
}
layer {
  name: "conv2/dw"
  type: "Convolution"
  bottom: "conv1/sep"
  top: "conv2/dw"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 16
    bias_term: false
    pad: 0
    kernel_h: 10
    kernel_w: 3
    group: 16
    #engine: CAFFE
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2/sep"
  type: "Convolution"
  bottom: "conv2/dw"
  top: "conv2/sep"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: true
    pad: 0
    kernel_size: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2/sep/bn"
  type: "BatchNorm"
  bottom: "conv2/sep"
  top: "conv2/sep"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv2/sep/bn/scale"
  type: "Scale"
  bottom: "conv2/sep"
  top: "conv2/sep"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}

layer {
  name: "relu2/sep"
  type: "ReLU"
  bottom: "conv2/sep"
  top: "conv2/sep"
}
layer {
  name: "conv3/dw"
  type: "Convolution"
  bottom: "conv2/sep"
  top: "conv3/dw"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 0
    kernel_h: 5
    kernel_w: 3
    group: 64
    #engine: CAFFE
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3/sep"
  type: "Convolution"
  bottom: "conv3/dw"
  top: "conv3/sep"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 128
    bias_term: true
    pad: 0
    kernel_size: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3/sep/bn"
  type: "BatchNorm"
  bottom: "conv3/sep"
  top: "conv3/sep"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv3/sep/bn/scale"
  type: "Scale"
  bottom: "conv3/sep"
  top: "conv3/sep"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}

layer {
  name: "relu3/sep"
  type: "ReLU"
  bottom: "conv3/sep"
  top: "conv3/sep"
}
layer {
  name: "avg_pool"
  type: "Pooling"
  bottom: "conv3/sep"
  top: "avg_pool"
  pooling_param {
    pool: AVE
    global_pooling: true
  }
}
layer {
  name: "fc"
  type: "InnerProduct"
  bottom: "avg_pool"
  top: "fc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 12
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "fc/bn"
  type: "BatchNorm"
  bottom: "fc"
  top: "fc"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "fc/bn/scale"
  type: "Scale"
  bottom: "fc"
  top: "fc"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}

layer {
  name: "output"
  type: "Softmax"
  bottom: "fc"
  top: "output"
}

Re: Compiler errors with DW COnvolution?

idata — Wed, 14 Feb 2018 01:57:02 GMT

@Ryose I see that your network is using non-square convolutions (i.e. in layer conv1/dw: 15 x 3). At the moment, we only have support for square convolutions on the NCS. Please visit https://movidius.github.io/ncsdk/Caffe.html for more details about limitations and known issues.

Re: Compiler errors with DW COnvolution?

idata — Fri, 16 Feb 2018 22:00:48 GMT

thanks for the reference @Tome_at_Intel

Re: Compiler errors with DW COnvolution?

idata — Sat, 17 Feb 2018 07:24:54 GMT

@Ryose I apologize, the limitations from the Caffe site is not up to date. We do support non-square convolutions, but for depth-wise convolutions we only support 3x3 convolutions at the moment.

Re: Compiler errors with DW COnvolution?

idata — Sun, 18 Mar 2018 21:17:44 GMT

what about simple 1*1 dwise?

I think thats not too hard to implement.

Re: Compiler errors with DW COnvolution?

idata — Tue, 20 Mar 2018 06:43:42 GMT

@jokilokis A 1x1 depthwise convolution seems strange. You are taking a single input pixel from a single input channel and multiplying it by a number, effectively scaling the image. It does not seem provide any filtering like a 3x3 depthwise convolution. I'm curious to know what kind of network you are using.