ActivityNet: mvNCCompile error: image inference not supported

idata · ‎08-10-2017

I am trying to compile a caffe model I have taken from an example I found at https://github.com/willxie/activity

When I run:

python3 ./mvNCCompile.pyc /home/niall/Desktop/activity-master/activitynet_deploy.prototxt

I get the folowing output

mvNCCompile v02.00, Copyright @ Movidius Ltd 2016

__ WARNING: using empty weights __

[Error 24] Toolkit Error: >1 image inference not supported.

Here is my network:

name: "ActivityNet"

input: "data_x"
input_shape {
    dim:50
    dim:1
    dim:1
    dim:64
}
input: "data_y"
input_shape {
    dim:50
    dim:1
    dim:1
    dim:64
}
input: "data_z"
input_shape {
    dim:50
    dim:1
    dim:1
    dim:64
}

layer {
  name: "conv1_x"
  type: "Convolution"
  bottom: "data_x"
  top: "conv1_x"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 3
    kernel_h: 1
    kernel_w: 20
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1_x"
  type: "Pooling"
  bottom: "conv1_x"
  top: "pool1_x"
  pooling_param {
    pool: MAX
    kernel_h: 1
    kernel_w: 3
    stride: 3
  }
}

layer {
  name: "conv1_y"
  type: "Convolution"
  bottom: "data_y"
  top: "conv1_y"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 3
    kernel_h: 1
    kernel_w: 20
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1_y"
  type: "Pooling"
  bottom: "conv1_y"
  top: "pool1_y"
  pooling_param {
    pool: MAX
    kernel_h: 1
    kernel_w: 3
    stride: 3
  }
}

layer {
  name: "conv1_z"
  type: "Convolution"
  bottom: "data_z"
  top: "conv1_z"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 3
    kernel_h: 1
    kernel_w: 20
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1_z"
  type: "Pooling"
  bottom: "conv1_z"
  top: "pool1_z"
  pooling_param {
    pool: MAX
    kernel_h: 1
    kernel_w: 3
    stride: 3
  }
}

layer {
  name: "concat"
  bottom: "pool1_x"
  bottom: "pool1_y"
  bottom: "pool1_z"
  top: "concat"
  type: "Concat"
  concat_param {
    axis: 0
  }
}


layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "concat"
  top: "ip1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 1024
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "ip1"
  top: "ip1"
}
layer {
  name: "drop1"
  type: "Dropout"
  bottom: "ip1"
  top: "ip1"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "ip2"
  type: "InnerProduct"
  bottom: "ip1"
  top: "ip2"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 30
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu2"
  type: "ReLU"
  bottom: "ip2"
  top: "ip2"
}
layer {
  name: "drop2"
  type: "Dropout"
  bottom: "ip2"
  top: "ip2"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "ip3"
  type: "InnerProduct"
  bottom: "ip2"
  top: "ip3"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 6
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "prob"
  type: "Softmax"
  bottom: "ip3"
  top: "prob"
}

I do not know why this error is happening as I do not pass any image when I call mvNCComplile.pyc.

I am using toolkit version 1.07.03 with the older version of the ncs, the fathom ncs.

idata · ‎08-14-2017

I am interested in applying the NCS in an industrial application that uses uni-variate sensors (such as temperature sensors) rather than camera images. I have found some example models that use LSTM networks for the most part on accelerometer data (using the HAR dataset) and even one that uses convolutional networks in the post above.

However I have had no luck compiling them for the ncs. What are the requirements of the caffe model that is fed to the mvNCCompile.pyc script? Can the NCS only support convolutional networks at this stage?

idata · ‎08-14-2017

@John Thank you for bringing this to our attention. ActivityNet is a network utilizes multiple inputs and at the moment, the NCS supports networks using at most one input. Although there isn’t a timeline for this feature yet, we are planning to support networks that use multiple inputs in the future.

idata · ‎08-15-2017

Hi Tome, thanks for your reply. I have tried removing the y, z and concat layers in my model, leaving me with just one input (x). However I still get the error

[Error 24] Toolkit Error: >1 image inference not supported.

Any ideas on why I am getting this error?

Note: I have included the path to the caffemodel file when I call mvNCCompile.pyc so I no longer have the empty weights warning detailed in my original post.

My hope is that if I can get this to work I can then process the data as one input with 3 channels. At this stage I just need some confirmation that the ncs will be usable with time-series data.

idata · ‎08-15-2017

Is it possible that the dimensions of my input layer are an invalid size?

idata · ‎08-29-2017

Hi,

I have manipulated my data to have all accelerometer values merged into 1 hdf5 dataset with dimensions: no. of samples (6959) x 1 channels x height 3 x width 64.

I have also changed my convolution kernal from 1_3 to 3_3 in size. The model trains ok, however again when I got to compile it with mvNCCompile I get image inference not supported. When I change the batchsize in my deploy network to 1, I get :

ValueError: negative dimensions are not allowed

Can someone tell me what part of my data/model the ncs does not like? all the layers in my deploy/validation network are supported. Perhaps there s some mismatch between my training model and my deploy/validation model that I am missing? I would greatly appreciate any assistance

idata · ‎08-13-2018

I am also getting the image error, the topic below has link to the repo I have made, cannot seem to find other info on the net about this @John did you ever resolve this mate ?

https://ncsforum.movidius.com/discussion/1068/error-24-toolkit-error-1-image-inference-not-supported#latest