Understanding and Designing 'Neural' Networks

I trained my first customised network, as designed in the previous log, on an Amazon AWS GPU and successfully got convergence, which means the network does work:

Training took about 3.5 hours and it got an mAP (val) (mean actual precision) of about 10, which is not particularly good, but it's a good start. It's very possible that if I had kept on training that this might have continued to slowly increase, but my budget for this session was $10, so it had to stop!

After messing about with some networks for a couple of months in an attemt to design computer vision for machines, I started to notice a few re-occurring topics and buzz words such as 'prototext' and 'weights'.

Curiosity has now got the better of me and I worked out that the various prototext files associated with a network describe it's structure in reasonably simple and human readable terms eg:

layer {
  name: "pool2/3x3_s2"
  type: "Pooling"
  bottom: "conv2/norm2"
  top: "pool2/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}

I presume that this is a python layer, but I'm not sure, but it does not matter for now. 

The fantastic Nvidia Digits software enables print out of a fancy graphical representation of the whole network and, starting with the renowned bvlc_googlenet.caffemodel, I thought I'd try and hack it and learn something through experimentation.

One of the first thing I looked for was symmetry and repetition, with the desire to simplify what initially look very complicated. I noticed that the above layer describes a 'link' between other blocks of layers that seem to repeat themselves about 6 times:

...... in the massive bvlc_googlenet network:

..... and in this way I managed to simplify it by removing what looked like about 6 large blocks of repeating layers to this:

...... And looking at this diagram very carefully, there's still one big block that repeats and should also be able to be removed. I tried removing it, but unfortunately gave this error:

Creating layer coverage/sig
Creating Layer coverage/sig
coverage/sig <- cvg/classifier
coverage/sig -> coverage
Setting up coverage/sig
Top shape: 2 1 80 80 (12800)
Memory required for data: 851282688
Creating layer bbox/regressor
Creating Layer bbox/regressor
bbox/regressor <- pool5/drop_s1_pool5/drop_s1_0_split_1
bbox/regressor -> bboxes
Setting up bbox/regressor
Top shape: 2 4 80 80 (51200)
Memory required for data: 851487488
Creating layer bbox_mask
Creating Layer bbox_mask
bbox_mask <- bboxes
bbox_mask <- coverage-block
bbox_mask -> bboxes-masked
Check failed: bottom[i]->shape() == bottom[0]->shape()      

...... So that's it for now ..... and here's my 'simplified' network prototext for object detection:

# DetectNet network

# DetectNet network

# Data/Input layers
name: "DetectNet"
layer {
  name: "train_data"
  type: "Data"
  top: "data"
  data_param {
    backend: LMDB
    source: "examples/kitti/kitti_train_images.lmdb"
    batch_size: 10
  }
  include: { phase: TRAIN }
}
layer {
  name: "train_label"
  type: "Data"
  top: "label"
  data_param {
    backend: LMDB
    source: "examples/kitti/kitti_train_labels.lmdb"
    batch_size: 10
  }
  include: { phase: TRAIN }
}
layer {
  name: "val_data"
  type: "Data"
  top: "data"
  data_param {
    backend: LMDB
    source: "examples/kitti/kitti_test_images.lmdb"
    batch_size: 6
  }
  include: { phase: TEST stage: "val" }
}
layer {
  name: "val_label"
  type: "Data"
  top: "label"
  data_param {
    backend: LMDB
    source: "examples/kitti/kitti_test_labels.lmdb"
    batch_size: 6
  }
  include: { phase: TEST stage: "val" }
}
layer {
  name: "deploy_data"
  type: "Input"
  top: "data"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 640
      dim: 640
    }
  }
  include: { phase: TEST not_stage: "val" }
}

# Data transformation layers
layer {
  name: "train_transform"
  type: "DetectNetTransformation"
  bottom: "data"
  bottom: "label"
  top: "transformed_data"
  top: "transformed_label"
  detectnet_groundtruth_param: {
    stride: 16
    scale_cvg: 0.4
    gridbox_type: GRIDBOX_MIN
    coverage_type: RECTANGULAR
    min_cvg_len: 20
    obj_norm: true
    image_size_x: 640
    image_size_y: 640
    crop_bboxes: false
    object_class: { src: 1 dst: 0} # obj class 1 -> cvg index 0
  }
   detectnet_augmentation_param: {
    crop_prob: 1
    shift_x: 32
    shift_y: 32
    flip_prob: 0.5
    rotation_prob: 0
    max_rotate_degree: 5
    scale_prob: 0.4
    scale_min: 0.8
    scale_max: 1.2
    hue_rotation_prob: 0.8
    hue_rotation: 30
    desaturation_prob: 0.8
    desaturation_max: 0.8
  }
  transform_param: {
    mean_value: 127
  }
  include: { phase: TRAIN }
}
layer {
  name: "val_transform"
  type: "DetectNetTransformation"
  bottom: "data"
  bottom: "label"
  top: "transformed_data"
  top: "transformed_label"
  detectnet_groundtruth_param: {
    stride: 16
    scale_cvg: 0.4
    gridbox_type: GRIDBOX_MIN
    coverage_type: RECTANGULAR
    min_cvg_len: 20
    obj_norm: true
    image_size_x: 640
    image_size_y: 640
    crop_bboxes: false
    object_class: { src: 1 dst: 0} # obj class 1 -> cvg index 0
  }
  transform_param: {
    mean_value: 127
  }
  include: { phase: TEST stage: "val" }
}
layer {
  name: "deploy_transform"
  type: "Power"
  bottom: "data"
  top: "transformed_data"
  power_param {
    shift: -127
  }
  include: { phase: TEST not_stage: "val" }
}

# Label conversion layers
layer {
  name: "slice-label"
  type: "Slice"
  bottom: "transformed_label"
  top: "foreground-label"
  top: "bbox-label"
  top: "size-label"
  top: "obj-label"
  top: "coverage-label"
  slice_param {
    slice_dim: 1
    slice_point: 1
    slice_point: 5
    slice_point: 7
    slice_point: 8
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "coverage-block"
  type: "Concat"
  bottom: "foreground-label"
  bottom: "foreground-label"
  bottom: "foreground-label"
  bottom: "foreground-label"
  top: "coverage-block"
  concat_param {
    concat_dim: 1
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "size-block"
  type: "Concat"
  bottom: "size-label"
  bottom: "size-label"
  top: "size-block"
  concat_param {
    concat_dim: 1
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "obj-block"
  type: "Concat"
  bottom: "obj-label"
  bottom: "obj-label"
  bottom: "obj-label"
  bottom: "obj-label"
  top: "obj-block"
  concat_param {
    concat_dim: 1
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "bb-label-norm"
  type: "Eltwise"
  bottom: "bbox-label"
  bottom: "size-block"
  top: "bbox-label-norm"
  eltwise_param {
    operation: PROD
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "bb-obj-norm"
  type: "Eltwise"
  bottom: "bbox-label-norm"
  bottom: "obj-block"
  top: "bbox-obj-label-norm"
  eltwise_param {
    operation: PROD
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}

######################################################################
# Start of convolutional network
######################################################################

layer {
  name: "conv1/7x7_s2"
  type: "Convolution"
  bottom: "transformed_data"
  top: "conv1/7x7_s2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 3
    kernel_size: 7
    stride: 2
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "conv1/relu_7x7"
  type: "ReLU"
  bottom: "conv1/7x7_s2"
  top: "conv1/7x7_s2"
}

layer {
  name: "pool1/3x3_s2"
  type: "Pooling"
  bottom: "conv1/7x7_s2"
  top: "pool1/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}

layer {
  name: "pool1/norm1"
  type: "LRN"
  bottom: "pool1/3x3_s2"
  top: "pool1/norm1"
  lrn_param {
    local_size: 5
    alpha: 0.0001
    beta: 0.75
  }
}

layer {
  name: "conv2/3x3_reduce"
  type: "Convolution"
  bottom: "pool1/norm1"
  top: "conv2/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "conv2/relu_3x3_reduce"
  type: "ReLU"
  bottom: "conv2/3x3_reduce"
  top: "conv2/3x3_reduce"
}

layer {
  name: "conv2/3x3"
  type: "Convolution"
  bottom: "conv2/3x3_reduce"
  top: "conv2/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 192
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "conv2/relu_3x3"
  type: "ReLU"
  bottom: "conv2/3x3"
  top: "conv2/3x3"
}

layer {
  name: "conv2/norm2"
  type: "LRN"
  bottom: "conv2/3x3"
  top: "conv2/norm2"
  lrn_param {
    local_size: 5
    alpha: 0.0001
    beta: 0.75
  }
}

layer {
  name: "pool2/3x3_s2"
  type: "Pooling"
  bottom: "conv2/norm2"
  top: "pool2/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
########################################################### 3a starts
layer {
  name: "inception_3a/1x1"
  type: "Convolution"
  bottom: "pool2/3x3_s2"
  top: "inception_3a/1x1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3a/relu_1x1"
  type: "ReLU"
  bottom: "inception_3a/1x1"
  top: "inception_3a/1x1"
}

layer {
  name: "inception_3a/3x3_reduce"
  type: "Convolution"
  bottom: "pool2/3x3_s2"
  top: "inception_3a/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_3a/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_3a/3x3_reduce"
  top: "inception_3a/3x3_reduce"
}

layer {
  name: "inception_3a/3x3"
  type: "Convolution"
  bottom: "inception_3a/3x3_reduce"
  top: "inception_3a/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_3a/relu_3x3"
  type: "ReLU"
  bottom: "inception_3a/3x3"
  top: "inception_3a/3x3"
}

layer {
  name: "inception_3a/5x5_reduce"
  type: "Convolution"
  bottom: "pool2/3x3_s2"
  top: "inception_3a/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.2
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3a/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_3a/5x5_reduce"
  top: "inception_3a/5x5_reduce"
}
layer {
  name: "inception_3a/5x5"
  type: "Convolution"
  bottom: "inception_3a/5x5_reduce"
  top: "inception_3a/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3a/relu_5x5"
  type: "ReLU"
  bottom: "inception_3a/5x5"
  top: "inception_3a/5x5"
}

layer {
  name: "inception_3a/pool"
  type: "Pooling"
  bottom: "pool2/3x3_s2"
  top: "inception_3a/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}

layer {
  name: "inception_3a/pool_proj"
  type: "Convolution"
  bottom: "inception_3a/pool"
  top: "inception_3a/pool_proj"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3a/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_3a/pool_proj"
  top: "inception_3a/pool_proj"
}
################################################################### start of connecting block
layer {
  name: "inception_3a/output"
  type: "Concat"
  bottom: "inception_3a/1x1"
  bottom: "inception_3a/3x3"
  bottom: "inception_3a/5x5"
  bottom: "inception_3a/pool_proj"
  top: "inception_3a/output"
}
################################################################### 3a finish
################################################################### start of connecting block
layer {
  name: "pool3/3x3_s2"
  type: "Pooling"
  bottom: "inception_3a/output"
  top: "pool3/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
################################################################### 4a starts
layer {
  name: "inception_4a/1x1"
  type: "Convolution"
  bottom: "pool3/3x3_s2"
  top: "inception_4a/1x1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 192
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4a/relu_1x1"
  type: "ReLU"
  bottom: "inception_4a/1x1"
  top: "inception_4a/1x1"
}

layer {
  name: "inception_4a/3x3_reduce"
  type: "Convolution"
  bottom: "pool3/3x3_s2"
  top: "inception_4a/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4a/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_4a/3x3_reduce"
  top: "inception_4a/3x3_reduce"
}

layer {
  name: "inception_4a/3x3"
  type: "Convolution"
  bottom: "inception_4a/3x3_reduce"
  top: "inception_4a/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 208
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4a/relu_3x3"
  type: "ReLU"
  bottom: "inception_4a/3x3"
  top: "inception_4a/3x3"
}

layer {
  name: "inception_4a/5x5_reduce"
  type: "Convolution"
  bottom: "pool3/3x3_s2"
  top: "inception_4a/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.2
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4a/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_4a/5x5_reduce"
  top: "inception_4a/5x5_reduce"
}
layer {
  name: "inception_4a/5x5"
  type: "Convolution"
  bottom: "inception_4a/5x5_reduce"
  top: "inception_4a/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 48
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4a/relu_5x5"
  type: "ReLU"
  bottom: "inception_4a/5x5"
  top: "inception_4a/5x5"
}
layer {
  name: "inception_4a/pool"
  type: "Pooling"
  bottom: "pool3/3x3_s2"
  top: "inception_4a/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_4a/pool_proj"
  type: "Convolution"
  bottom: "inception_4a/pool"
  top: "inception_4a/pool_proj"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4a/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_4a/pool_proj"
  top: "inception_4a/pool_proj"
}
################################################################### 4a finish
layer {
  name: "inception_4a/output"
  type: "Concat"
  bottom: "inception_4a/1x1"
  bottom: "inception_4a/3x3"
  bottom: "inception_4a/5x5"
  bottom: "inception_4a/pool_proj"
  top: "inception_4a/output"
}
################################################################### start of connecting block

layer {
  name: "inception_5a/1x1"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_5a/1x1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_1x1"
  type: "ReLU"
  bottom: "inception_5a/1x1"
  top: "inception_5a/1x1"
}

layer {
  name: "inception_5a/3x3_reduce"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_5a/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 160
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_5a/3x3_reduce"
  top: "inception_5a/3x3_reduce"
}

layer {
  name: "inception_5a/3x3"
  type: "Convolution"
  bottom: "inception_5a/3x3_reduce"
  top: "inception_5a/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 320
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_3x3"
  type: "ReLU"
  bottom: "inception_5a/3x3"
  top: "inception_5a/3x3"
}
layer {
  name: "inception_5a/5x5_reduce"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_5a/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.2
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_5a/5x5_reduce"
  top: "inception_5a/5x5_reduce"
}
layer {
  name: "inception_5a/5x5"
  type: "Convolution"
  bottom: "inception_5a/5x5_reduce"
  top: "inception_5a/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_5x5"
  type: "ReLU"
  bottom: "inception_5a/5x5"
  top: "inception_5a/5x5"
}
layer {
  name: "inception_5a/pool"
  type: "Pooling"
  bottom: "inception_4a/output"
  top: "inception_5a/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_5a/pool_proj"
  type: "Convolution"
  bottom: "inception_5a/pool"
  top: "inception_5a/pool_proj"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5a/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_5a/pool_proj"
  top: "inception_5a/pool_proj"
}
################################################################### start of connecting block
layer {
  name: "inception_5a/output"
  type: "Concat"
  bottom: "inception_5a/1x1"
  bottom: "inception_5a/3x3"
  bottom: "inception_5a/5x5"
  bottom: "inception_5a/pool_proj"
  top: "inception_5a/output"
}
################################################################### start of connecting block
layer {
  name: "pool5/drop_s1"
  type: "Dropout"
  bottom: "inception_5a/output"
  top: "pool5/drop_s1"
  dropout_param {
    dropout_ratio: 0.4
  }
}
layer {
  name: "cvg/classifier"
  type: "Convolution"
  bottom: "pool5/drop_s1"
  top: "cvg/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 1
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.
    }
  }
}
layer {
  name: "coverage/sig"
  type: "Sigmoid"
  bottom: "cvg/classifier"
  top: "coverage"
}
layer {
  name: "bbox/regressor"
  type: "Convolution"
  bottom: "pool5/drop_s1"
  top: "bboxes"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.
    }
  }
}

######################################################################
# End of convolutional network
######################################################################

# Convert bboxes
layer {
  name: "bbox_mask"
  type: "Eltwise"
  bottom: "bboxes"
  bottom: "coverage-block"
  top: "bboxes-masked"
  eltwise_param {
    operation: PROD
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "bbox-norm"
  type: "Eltwise"
  bottom: "bboxes-masked"
  bottom: "size-block"
  top: "bboxes-masked-norm"
  eltwise_param {
    operation: PROD
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "bbox-obj-norm"
  type: "Eltwise"
  bottom: "bboxes-masked-norm"
  bottom: "obj-block"
  top: "bboxes-obj-masked-norm"
  eltwise_param {
    operation: PROD
  }
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}

# Loss layers
layer {
  name: "bbox_loss"
  type: "L1Loss"
  bottom: "bboxes-obj-masked-norm"
  bottom: "bbox-obj-label-norm"
  top: "loss_bbox"
  loss_weight: 2
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}
layer {
  name: "coverage_loss"
  type: "EuclideanLoss"
  bottom: "coverage"
  bottom: "coverage-label"
  top: "loss_coverage"
  include { phase: TRAIN }
  include { phase: TEST stage: "val" }
}

# Cluster bboxes
layer {
    type: 'Python'
    name: 'cluster'
    bottom: 'coverage'
    bottom: 'bboxes'
    top: 'bbox-list'
    python_param {
        module: 'caffe.layers.detectnet.clustering'
        layer: 'ClusterDetections'
        param_str : '640, 640, 16, 0.6, 2, 0.02, 22, 1'
    }
    include: { phase: TEST }
}

# Calculate mean average precision
layer {
  type: 'Python'
  name: 'cluster_gt'
  bottom: 'coverage-label'
  bottom: 'bbox-label'
  top: 'bbox-list-label'
  python_param {
      module: 'caffe.layers.detectnet.clustering'
      layer: 'ClusterGroundtruth'
      param_str : '640, 640, 16, 1'
  }
  include: { phase: TEST stage: "val" }
}
layer {
    type: 'Python'
    name: 'score'
    bottom: 'bbox-list-label'
    bottom: 'bbox-list'
    top: 'bbox-list-scored'
    python_param {
        module: 'caffe.layers.detectnet.mean_ap'
        layer: 'ScoreDetections'
    }
    include: { phase: TEST stage: "val" }
}
layer {
    type: 'Python'
    name: 'mAP'
    bottom: 'bbox-list-scored'
    top: 'mAP'
    top: 'precision'
    top: 'recall'
    python_param {
        module: 'caffe.layers.detectnet.mean_ap'
        layer: 'mAP'
        param_str : '640, 640, 16'
    }
    include: { phase: TEST stage: "val" }
}