Error while running

[xxd137494375]

[DNNDK] Invalid mean value for DPU kernel. kernel name : anet_0

There is a error at run time，How to solve this problem?

Thanks!

[gguasti]

hello

The invalid mean value might be due to a wrong setup in the prototxt file describing the CNN architecture. For the final step (compilation) the deployed.prototxt should have commented all references to the mean_file. Only the mean values should be present.

[xxd137494375]

How can I modify this network?

layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
mirror: true
crop_size: 126
scale: 0.017
}
image_data_param {
source: "/home/aa/demo/Classification/train.txt"
root_folder: "/home/aa/demo/Classification/all-images/"
batch_size: 50
new_width: 126
new_height: 126
is_color: false
shuffle: true
}
}
layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
mirror: false
crop_size: 126
scale: 0.017
}
image_data_param {
source: "/home/aa/demo/Classification/test.txt"
root_folder: "/home/aa/demo/Classification/all-images/"
batch_size: 25
new_width: 126
new_height: 126
is_color: false
}
}

[xxd137494375]

Do I need to retrain?

[aluo]

@xxd137494375How do you set the mean value in float prototxt? Here below is an example -

name: "ResNet-50"
layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: false
    mean_value: 104
    mean_value: 107
    mean_value: 123
  }
  image_data_param {
#    source: "./data/imagenet_256/calibration.txt"
#    root_folder: "./data/imagenet_256/calibration_images/"
    source: "/down/dnndk_2.08/common/dataset/imagenet/val.txt"
    root_folder: "/down/dnndk_2.08/common/dataset/imagenet/val/"
    batch_size: 10
    shuffle: false
    new_height: 224
    new_width: 224
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 64
    pad: 3
    kernel_size: 7
    stride: 2
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"

[xxd137494375]

My float.prototxt are below，i did not set the mean_value



layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
mirror: true
crop_size: 126
scale: 0.017
}
image_data_param {
source: "/home/aa/demo/Classification/train.txt"
root_folder: "/home/aa/demo/Classification/all-images/"
batch_size: 50
new_width: 126
new_height: 126
is_color: false
shuffle: true
}
}
layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
mirror: false
crop_size: 126
scale: 0.017
}
image_data_param {
source: "/home/aa/demo/Classification/test.txt"
root_folder: "/home/aa/demo/Classification/all-images/"
batch_size: 25
new_width: 126
new_height: 126
is_color: false
}
}

[xxd137494375]

Besides，I use black and white images.

[marron]

Not certain if this was resolved, sharing my experience just in case.

I ran into the same issue as you, with the following prototxt definition:

   transform_param {
      scale: 0.00390625
      mirror: false
   }

Note the 1/256 scale is for the binary input image, similar to what you mentionned.

I was able to resolve the issue by adding the missing "mean_value" value (just one, since this is a grayscale image (not rgb).

   transform_param {
      scale: 0.00390625
      mirror: false
      mean_value: 0
   }

Hope this helps !

AlbertaBeef ;)