Re: Completely wrong values from NCS

idata · ‎12-20-2017

I have been working on implementing Caffe MTCNN on the ncs.

After compiling a PNet prototxt with models and running a sample application, I figured that the values are vastly incorrect from what I should be getting from a Caffe application running the same model.

Given the PNet weight definition (det1.caffemodel acquired from https://github.com/DuinoDu/mtcnn/tree/master/model ), all the values should be in the range of [-1, 1], however the output acquired via the NCS ranges from [-65344, 62912].

There is no correlation to be found between the outputs from NCS and Caffe app.

Is there a problem with the Caffe softmax layer on NCS?

Please help. Thanks.

input: "data"

input_shape {

dim: 1

dim: 3

dim: 328

dim: 148

}

layer {

type: "Convolution"

bottom: "data"

top: "conv1"

param {

lr_mult: 1

decay_mult: 1

}

param {

lr_mult: 2

decay_mult: 0

}

convolution_param {

num_output: 10

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0

}

layer {

type: "PReLU"

bottom: "conv1"

top: "conv1"

}

layer {

type: "Pooling"

bottom: "conv1"

top: "pool1"

pooling_param {

pool: MAX

kernel_size: 2

stride: 2

}

layer {

type: "Convolution"

bottom: "pool1"

top: "conv2"

param {

lr_mult: 1

decay_mult: 1

}

param {

lr_mult: 2

decay_mult: 0

}

convolution_param {

num_output: 16

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0

}

layer {

type: "PReLU"

bottom: "conv2"

top: "conv2"

}

layer {

type: "Convolution"

bottom: "conv2"

top: "conv3"

param {

lr_mult: 1

decay_mult: 1

}

param {

lr_mult: 2

decay_mult: 0

}

convolution_param {

num_output: 32

kernel_size: 3

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0

}

layer {

type: "PReLU"

bottom: "conv3"

top: "conv3"

}

layer {

type: "Convolution"

bottom: "conv3"

top: "conv4-1"

param {

lr_mult: 1

decay_mult: 1

}

param {

lr_mult: 2

decay_mult: 0

}

convolution_param {

num_output: 2

kernel_size: 1

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0

}

layer {

type: "Convolution"

bottom: "conv3"

top: "conv4-2"

param {

lr_mult: 1

decay_mult: 1

}

param {

lr_mult: 2

decay_mult: 0

}

convolution_param {

num_output: 4

kernel_size: 1

stride: 1

weight_filler {

type: "xavier"

}

bias_filler {

type: "constant"

value: 0

}

layer {

type: "Softmax"

bottom: "conv4-1"

top: "prob1"

}

layer {

type: "Concat"

bottom: "conv4-2"

bottom: "prob1"

top: "concat"

}

layer {

type: "Flatten"

bottom: "concat"

top: "output"

}

idata · ‎12-20-2017

@sejunkim Caffe on NCS supports basically configuration with 3 channels and square images. You can try to change the input_shape to {1,3,X,X} and check if the issue persists.

idata · ‎01-10-2018

@sejunkim I think the shape should be 12_12,24_24 and 48*48, do you have a demo run based on NCS?

idata · ‎01-12-2018

I found that there is something wrong in transferring prototxt file to graph file when "top" and "bottom" are the same name. You can change the top name of PRelu layers and try again

idata · ‎01-12-2018

@Tome_at_Intel Did you find the problem I mentioned above?

Regarding to PNet:

name: "PNet"
input: "data"
input_shape {
  dim: 1
  dim: 3
  dim: 328
  dim: 148
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 10
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "prelu1"
  type: "PReLU"
  bottom: "conv1"
  top: "prelu1"
}

layer {
  name: "pool1"
  type: "Pooling"
  bottom: "prelu1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

layer {
  name: "conv2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "prelu2"
  type: "PReLU"
  bottom: "conv2"
  top: "prelu2"
}

layer {
  name: "conv3"
  type: "Convolution"
  bottom: "prelu2"
  top: "conv3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "prelu3"
  type: "PReLU"
  bottom: "conv3"
  top: "prelu3"
}

layer {
  name: "conv4-1"
  type: "Convolution"
  bottom: "prelu3"
  top: "conv4-1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 2
    kernel_size: 1
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "conv4-2"
  type: "Convolution"
  bottom: "prelu3"
  top: "conv4-2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4
    kernel_size: 1
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "prob1"
  type: "Softmax"
  bottom: "conv4-1"
  top: "prob1"
}

layer {
  name: "concat"
  type: "Concat"
  bottom: "conv4-2"
  bottom: "prob1"
  top: "concat"
}

layer {
  name: "output"
  type: "Flatten"
  bottom: "concat"
  top: "output"
}

I have modified several places which are related to PRelu. These two prototxt can get same results in Caffe while different results in Movidius

idata · ‎11-21-2018

@sejunkim

Have you made any progress with MTCNN? Could you share some experience?