Changing the initial stride from 16 to 8

[ehol...@gmail.com]

Hello,
We were trying to refine DetectNet so that it would look at more of our input data (images) by changing the initial stride from 16 to 8. We initially got an error that said "bottom[i]->shape == bottom[0]->shape failed". We finally figured out how to make this work. If you look at the pool3/3x3_s2 layer the output shape is (480,63,84) which the last two numbers are our image width and height divided by 16. This is not correct as this "new" stride of 16 propagates to the lower layers. We resolved this issue by making the stride=1 and the kernel 1x1 for this layer only. The rest propagated just fine. We realize that having a pooling layer that uses a stride of 1 and a kernel size of 1x1 actually does nothing. You can surgically remove this layer and things should just work. However we left this layer alone and we are successfully training with a stride of 8. We are posting this to see if other people have had this same problem and how they resolved it
Thanks.

[Alper ALT]

Hello and thanks for sharing this info. May I assume a lower stride helps with smaller objects, or objects standing so close to each other ? Or, whats your expectations from a smaller stride? Thanks again.

[ehol...@gmail.com]

Yeah. We have fairly small objects in our data set and wanted to use a smaller stride to detect them hoping the mAP would improve. We are still running the smaller stride runs and the mAP is improving. I also forgot to mention that we had to drop down from 4 to 2 titans as the small stride and kernel required more memory.

[Alper ALT]

Perfect. I will exactly follow your changes to see if any improvements. It will be a "true" improvement if it can individually detect the objects standing nearby, (very close or barely touching to each other). If your mod can solve this issue, my test accuracy will possibly beat human eye :)

[Alper ALT]

Ok, first test is not quite encouraging. mAP went straight from the beginning and never exceeded 3. Normal version provides 65.x mAP..

[ehol...@gmail.com]

Hi

I would give it some time as it is re-learning things. Quick questions: (1) Did you use a pre-trained model (2) Did you have to drop graphics video cards (reduce them to fit the data) (3) Was the bottle neck in the error at the same layer as the pool3x3/s2?

It is encouraging that you didn't get the mis-shape error. I would give it some time. Maybe reduce your learning rate but otherwise keep the same network parameters.

[Alper ALT]

Hi, I appreciate for the reply.  First, I want to be sure that modifications are at right points;

# Data transformation layers
layer {
  name: "train_transform"
  type: "DetectNetTransformation"
  bottom: "data"
  bottom: "label"
  top: "transformed_data"
  top: "transformed_label"
  detectnet_groundtruth_param: {
    stride: 8

-----------------------

 layer {
  name: "val_transform"
  type: "DetectNetTransformation"
  bottom: "data"
  bottom: "label"
  top: "transformed_data"
  top: "transformed_label"
  detectnet_groundtruth_param: {
    stride: 8

-------------------------

 layer {
  name: "pool3/3x3_s2"
  type: "Pooling"
  bottom: "inception_3b/output"
  top: "pool3/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 1
    stride: 1
  }
}

-------------------------
1) I have never used a pretrained model file up to now, in any of my trainings. I can still achieve 65mAP this way, in 30 to 50 epochs (my image complexity is not high, like detecting rice grains over solid colored table, this helps to mAP I think).
 
2) Not sure what you mean here, graphics video card? I am using a Maxwell TitanX with 12GB RAM..
 
3) A complex question for the beginner, I am afraid I dont know how to determine if a bottleneck appears somewhere in the model.  But one observation, no change in mAP observed at learning rate change-points, which normally occurs with standard model, therefore I am not sure it will likely to increase with more number of epochs, and training completed in 3 hours which normally takes 6-7 hr with standard model, same dataset, until epoch 30 . Another thing, with default res, (1248x384),default batch size (32?) works fine at 12GB ram, after mod, at batch=3, 11.1GB of 12GB was in use.
 
Unit now, I could run it two times, one with batch@3 and accu@4 (picture in above message); And second run was with batch&accu @ 1. I am really excited for ability to test a lower stride, but I think I will need some more help..

[ehol...@gmail.com]

Hi

Everything looks correct. I would have to see the entire DetectNet.prototxt file (if you can post it I can CTRL-A (select all) CTRL-C (copy) and post it here I can copy it and take a look at it. It looks like your learning rate is way too low. If I use a learning rate of 0.001 things tend to go wild. This is a good thing when starting out training an image set. Anyway if you post you entire prototxt I will run it through our data set and see what happens. Be aware that your images have to be padded so that (width mod stride) = (height mod stride) = 0 but I think you already do that. I have also changed other default parameters but I can take a look. Also be careful using another pre-trained model. It really shouldn't make that much of a difference since they are usually mostly just weights but I would try not using one at all first. This may take many epochs to see a non-zero mAP. It is odd that your mAP starts non-zero.

Thanks.

[Alper ALT]

Hi, here is a copy of whats available in "custom network" area. I set image padding to 1248 and 384 and no entry at resize. All 2478 images are at 1248x384. Also checked mAP graph it starts at 1st epoch as 1.4, and ends at 30th epoch at 1.6x.. All other settings are copy of detectnet walkthrough except "Minimum box size (in pixels) for validation set" is set to zero, at lmdb dataset prep. screen.

# 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: 384
      dim: 1248
    }
  }
  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: 8

    scale_cvg: 0.4
    gridbox_type: GRIDBOX_MIN
    coverage_type: RECTANGULAR
    min_cvg_len: 20
    obj_norm: true
    image_size_x: 1248
    image_size_y: 384
    crop_bboxes: true
    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: 8

    scale_cvg: 0.4
    gridbox_type: GRIDBOX_MIN
    coverage_type: RECTANGULAR
    min_cvg_len: 20
    obj_norm: true
    image_size_x: 1248
    image_size_y: 384
    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
  }
}

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"
}

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"
}

layer {
  name: "inception_3b/1x1"
  type: "Convolution"
  bottom: "inception_3a/output"
  top: "inception_3b/1x1"
  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: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_3b/relu_1x1"
  type: "ReLU"
  bottom: "inception_3b/1x1"
  top: "inception_3b/1x1"
}

layer {
  name: "inception_3b/3x3_reduce"
  type: "Convolution"
  bottom: "inception_3a/output"
  top: "inception_3b/3x3_reduce"
  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: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3b/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_3b/3x3_reduce"
  top: "inception_3b/3x3_reduce"
}
layer {
  name: "inception_3b/3x3"
  type: "Convolution"
  bottom: "inception_3b/3x3_reduce"
  top: "inception_3b/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: "inception_3b/relu_3x3"
  type: "ReLU"
  bottom: "inception_3b/3x3"
  top: "inception_3b/3x3"
}

layer {
  name: "inception_3b/5x5_reduce"
  type: "Convolution"
  bottom: "inception_3a/output"
  top: "inception_3b/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_3b/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_3b/5x5_reduce"
  top: "inception_3b/5x5_reduce"
}
layer {
  name: "inception_3b/5x5"
  type: "Convolution"
  bottom: "inception_3b/5x5_reduce"
  top: "inception_3b/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_3b/relu_5x5"
  type: "ReLU"
  bottom: "inception_3b/5x5"
  top: "inception_3b/5x5"
}

layer {
  name: "inception_3b/pool"
  type: "Pooling"
  bottom: "inception_3a/output"
  top: "inception_3b/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_3b/pool_proj"
  type: "Convolution"
  bottom: "inception_3b/pool"
  top: "inception_3b/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_3b/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_3b/pool_proj"
  top: "inception_3b/pool_proj"
}
layer {
  name: "inception_3b/output"
  type: "Concat"
  bottom: "inception_3b/1x1"
  bottom: "inception_3b/3x3"
  bottom: "inception_3b/5x5"
  bottom: "inception_3b/pool_proj"
  top: "inception_3b/output"

}

layer {
  name: "pool3/3x3_s2"
  type: "Pooling"
  bottom: "inception_3b/output"
  top: "pool3/3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 1
    stride: 1
  }
}

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"
}
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"
}

layer {
  name: "inception_4b/1x1"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_4b/1x1"
  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: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4b/relu_1x1"
  type: "ReLU"
  bottom: "inception_4b/1x1"
  top: "inception_4b/1x1"
}
layer {
  name: "inception_4b/3x3_reduce"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_4b/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 112
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4b/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_4b/3x3_reduce"
  top: "inception_4b/3x3_reduce"
}
layer {
  name: "inception_4b/3x3"
  type: "Convolution"
  bottom: "inception_4b/3x3_reduce"
  top: "inception_4b/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 224
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4b/relu_3x3"
  type: "ReLU"
  bottom: "inception_4b/3x3"
  top: "inception_4b/3x3"
}
layer {
  name: "inception_4b/5x5_reduce"
  type: "Convolution"
  bottom: "inception_4a/output"
  top: "inception_4b/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.2
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4b/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_4b/5x5_reduce"
  top: "inception_4b/5x5_reduce"
}
layer {
  name: "inception_4b/5x5"
  type: "Convolution"
  bottom: "inception_4b/5x5_reduce"
  top: "inception_4b/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4b/relu_5x5"
  type: "ReLU"
  bottom: "inception_4b/5x5"
  top: "inception_4b/5x5"
}
layer {
  name: "inception_4b/pool"
  type: "Pooling"
  bottom: "inception_4a/output"
  top: "inception_4b/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_4b/pool_proj"
  type: "Convolution"
  bottom: "inception_4b/pool"
  top: "inception_4b/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_4b/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_4b/pool_proj"
  top: "inception_4b/pool_proj"
}
layer {
  name: "inception_4b/output"
  type: "Concat"
  bottom: "inception_4b/1x1"
  bottom: "inception_4b/3x3"
  bottom: "inception_4b/5x5"
  bottom: "inception_4b/pool_proj"
  top: "inception_4b/output"
}

layer {
  name: "inception_4c/1x1"
  type: "Convolution"
  bottom: "inception_4b/output"
  top: "inception_4c/1x1"
  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: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4c/relu_1x1"
  type: "ReLU"
  bottom: "inception_4c/1x1"
  top: "inception_4c/1x1"
}

layer {
  name: "inception_4c/3x3_reduce"
  type: "Convolution"
  bottom: "inception_4b/output"
  top: "inception_4c/3x3_reduce"
  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: 0.09
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

layer {
  name: "inception_4c/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_4c/3x3_reduce"
  top: "inception_4c/3x3_reduce"
}
layer {
  name: "inception_4c/3x3"
  type: "Convolution"
  bottom: "inception_4c/3x3_reduce"
  top: "inception_4c/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4c/relu_3x3"
  type: "ReLU"
  bottom: "inception_4c/3x3"
  top: "inception_4c/3x3"
}
layer {
  name: "inception_4c/5x5_reduce"
  type: "Convolution"
  bottom: "inception_4b/output"
  top: "inception_4c/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.2
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4c/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_4c/5x5_reduce"
  top: "inception_4c/5x5_reduce"
}
layer {
  name: "inception_4c/5x5"
  type: "Convolution"
  bottom: "inception_4c/5x5_reduce"
  top: "inception_4c/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4c/relu_5x5"
  type: "ReLU"
  bottom: "inception_4c/5x5"
  top: "inception_4c/5x5"
}
layer {
  name: "inception_4c/pool"
  type: "Pooling"
  bottom: "inception_4b/output"
  top: "inception_4c/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_4c/pool_proj"
  type: "Convolution"
  bottom: "inception_4c/pool"
  top: "inception_4c/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_4c/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_4c/pool_proj"
  top: "inception_4c/pool_proj"
}
layer {
  name: "inception_4c/output"
  type: "Concat"
  bottom: "inception_4c/1x1"
  bottom: "inception_4c/3x3"
  bottom: "inception_4c/5x5"
  bottom: "inception_4c/pool_proj"
  top: "inception_4c/output"
}

layer {
  name: "inception_4d/1x1"
  type: "Convolution"
  bottom: "inception_4c/output"
  top: "inception_4d/1x1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 112
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4d/relu_1x1"
  type: "ReLU"
  bottom: "inception_4d/1x1"
  top: "inception_4d/1x1"
}
layer {
  name: "inception_4d/3x3_reduce"
  type: "Convolution"
  bottom: "inception_4c/output"
  top: "inception_4d/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 144
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4d/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_4d/3x3_reduce"
  top: "inception_4d/3x3_reduce"
}
layer {
  name: "inception_4d/3x3"
  type: "Convolution"
  bottom: "inception_4d/3x3_reduce"
  top: "inception_4d/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 288
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4d/relu_3x3"
  type: "ReLU"
  bottom: "inception_4d/3x3"
  top: "inception_4d/3x3"
}
layer {
  name: "inception_4d/5x5_reduce"
  type: "Convolution"
  bottom: "inception_4c/output"
  top: "inception_4d/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: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4d/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_4d/5x5_reduce"
  top: "inception_4d/5x5_reduce"
}
layer {
  name: "inception_4d/5x5"
  type: "Convolution"
  bottom: "inception_4d/5x5_reduce"
  top: "inception_4d/5x5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_4d/relu_5x5"
  type: "ReLU"
  bottom: "inception_4d/5x5"
  top: "inception_4d/5x5"
}
layer {
  name: "inception_4d/pool"
  type: "Pooling"
  bottom: "inception_4c/output"
  top: "inception_4d/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_4d/pool_proj"
  type: "Convolution"
  bottom: "inception_4d/pool"
  top: "inception_4d/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_4d/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_4d/pool_proj"
  top: "inception_4d/pool_proj"
}
layer {
  name: "inception_4d/output"
  type: "Concat"
  bottom: "inception_4d/1x1"
  bottom: "inception_4d/3x3"
  bottom: "inception_4d/5x5"
  bottom: "inception_4d/pool_proj"
  top: "inception_4d/output"
}

layer {
  name: "inception_4e/1x1"
  type: "Convolution"
  bottom: "inception_4d/output"
  top: "inception_4e/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_4e/relu_1x1"
  type: "ReLU"
  bottom: "inception_4e/1x1"
  top: "inception_4e/1x1"
}
layer {
  name: "inception_4e/3x3_reduce"
  type: "Convolution"
  bottom: "inception_4d/output"
  top: "inception_4e/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_4e/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_4e/3x3_reduce"
  top: "inception_4e/3x3_reduce"
}
layer {
  name: "inception_4e/3x3"
  type: "Convolution"
  bottom: "inception_4e/3x3_reduce"
  top: "inception_4e/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_4e/relu_3x3"
  type: "ReLU"
  bottom: "inception_4e/3x3"
  top: "inception_4e/3x3"
}
layer {
  name: "inception_4e/5x5_reduce"
  type: "Convolution"
  bottom: "inception_4d/output"
  top: "inception_4e/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_4e/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_4e/5x5_reduce"
  top: "inception_4e/5x5_reduce"
}
layer {
  name: "inception_4e/5x5"
  type: "Convolution"
  bottom: "inception_4e/5x5_reduce"
  top: "inception_4e/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_4e/relu_5x5"
  type: "ReLU"
  bottom: "inception_4e/5x5"
  top: "inception_4e/5x5"
}
layer {
  name: "inception_4e/pool"
  type: "Pooling"
  bottom: "inception_4d/output"
  top: "inception_4e/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_4e/pool_proj"
  type: "Convolution"
  bottom: "inception_4e/pool"
  top: "inception_4e/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_4e/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_4e/pool_proj"
  top: "inception_4e/pool_proj"
}
layer {
  name: "inception_4e/output"
  type: "Concat"
  bottom: "inception_4e/1x1"
  bottom: "inception_4e/3x3"
  bottom: "inception_4e/5x5"
  bottom: "inception_4e/pool_proj"
  top: "inception_4e/output"
}



layer {
  name: "inception_5a/1x1"
  type: "Convolution"
  bottom: "inception_4e/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_4e/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_4e/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_4e/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"
}
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"
}

layer {
  name: "inception_5b/1x1"
  type: "Convolution"
  bottom: "inception_5a/output"
  top: "inception_5b/1x1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 384
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5b/relu_1x1"
  type: "ReLU"
  bottom: "inception_5b/1x1"
  top: "inception_5b/1x1"
}
layer {
  name: "inception_5b/3x3_reduce"
  type: "Convolution"
  bottom: "inception_5a/output"
  top: "inception_5b/3x3_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  convolution_param {
    num_output: 192
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5b/relu_3x3_reduce"
  type: "ReLU"
  bottom: "inception_5b/3x3_reduce"
  top: "inception_5b/3x3_reduce"
}
layer {
  name: "inception_5b/3x3"
  type: "Convolution"
  bottom: "inception_5b/3x3_reduce"
  top: "inception_5b/3x3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 384
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5b/relu_3x3"
  type: "ReLU"
  bottom: "inception_5b/3x3"
  top: "inception_5b/3x3"
}
layer {
  name: "inception_5b/5x5_reduce"
  type: "Convolution"
  bottom: "inception_5a/output"
  top: "inception_5b/5x5_reduce"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 48
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5b/relu_5x5_reduce"
  type: "ReLU"
  bottom: "inception_5b/5x5_reduce"
  top: "inception_5b/5x5_reduce"
}
layer {
  name: "inception_5b/5x5"
  type: "Convolution"
  bottom: "inception_5b/5x5_reduce"
  top: "inception_5b/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: 0.1
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}
layer {
  name: "inception_5b/relu_5x5"
  type: "ReLU"
  bottom: "inception_5b/5x5"
  top: "inception_5b/5x5"
}
layer {
  name: "inception_5b/pool"
  type: "Pooling"
  bottom: "inception_5a/output"
  top: "inception_5b/pool"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 1
    pad: 1
  }
}
layer {
  name: "inception_5b/pool_proj"
  type: "Convolution"
  bottom: "inception_5b/pool"
  top: "inception_5b/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_5b/relu_pool_proj"
  type: "ReLU"
  bottom: "inception_5b/pool_proj"
  top: "inception_5b/pool_proj"
}
layer {
  name: "inception_5b/output"
  type: "Concat"
  bottom: "inception_5b/1x1"
  bottom: "inception_5b/3x3"
  bottom: "inception_5b/5x5"
  bottom: "inception_5b/pool_proj"
  top: "inception_5b/output"
}
layer {
  name: "pool5/drop_s1"
  type: "Dropout"
  bottom: "inception_5b/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 : '1248, 352, 16, 0.6, 3, 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 : '1248, 352, 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 : '1248, 352, 16'
    }
    include: { phase: TEST stage: "val" }

[ehol...@gmail.com]

I'll check and get back to you

[Alper ALT]

Ok I have lowered LR to 0.001, running to 30 epochs, no mAP yet... Meanwhile, may I ask whats your dataset image resolution used in stride-8 trainings?

[Alper ALT]

And If this helps for anything, a working (stride-16) model vs stride-8 model.

[ehol...@gmail.com]

Hi,

I ran your network (The end didn't post but I just added the standard 8 stride network. Only a couple of lines). It is imperative that you do not use a pre-trained model as you must establish a baseline. The mAP was zero until epoch 32, then spiked to 17.11 then back to zero until epoch 47 where it rose considerably and a standard learning curve began. At around epoch 63 the mAP was 63%. I think it should rise more. The learning rate was 0.001 which is quite aggressive. I will take the maximum mAP and use that as a starting point for the next run. So with my data your network worked fine. My image size is (3,1344,1008) where (3=RGB, 1344 is the width and 1008 is the height). You should have your images divisible by 8. If they are not you can pad them to be so. Your image preparation is also very important. Let me know how you are making out. Everything seems OK on my side but then again I don't have your image set. 

Thanks.

[Alper ALT]

Ok I just have solved the issue and below image is from ongoing training.

The missing thing was at those last lines, there are 3 more "16" sitting there. I have changed them to 8 also, then feed a pretrained model along (same dataset, stride-16). Please see image. I wonder how it will react to small/close objects now. In fact I have never waited until epoch 30, After completion of this, I can start a new one. And yes, my image sizes are compatible with 8, 16 and 32. Thank you very much again. One last question, how stride-8 yielded for you? I mean, could you get the higher accuracy you aimed?

[ehol...@gmail.com]

We haven't really benefited by the reduced stride. Still max mAP = 70%. I may go back to a stride of 16. Yours seems to really be doing well! Nice mAP! Congratulations. Let me know if I can be of more help.

[Alper ALT]

Thank you, we will see how it will perform with problem images/objects. Btw, I tend to think stride occurs at X axis but it actually takes place on Y axis too, right? And do you see any benefit to push it down to stride-4 ?

[Alper ALT]

A comparison.

[sulthan...@triassicsolutions.com]

I have changed my code as mentioned but getting an error.Pls help me find out the reason

ERROR: Layer 'inception_4a/1x1' references bottom 'pool3/3x3_s2' at the TRAIN stage however this blob is not included at that stage. Please consider using an include directive to limit the scope of this layer.

Traceback (most recent call last):
  File "/usr/share/digits/digits/scheduler.py", line 507, in run_task
    task.run(resources)
  File "/usr/share/digits/digits/task.py", line 184, in run
    self.before_run()
  File "/usr/share/digits/digits/model/tasks/caffe_train.py", line 138, in before_run
    self.save_files_generic()
  File "/usr/share/digits/digits/model/tasks/caffe_train.py", line 598, in save_files_generic
    CaffeTrainTask.net_sanity_check(train_val_network, caffe_pb2.TRAIN)
  File "/usr/share/digits/digits/model/tasks/caffe_train.py", line 1472, in net_sanity_check
    layer.name, bottom, "TRAIN" if phase == caffe_pb2.TRAIN else "TEST"))
CaffeTrainSanityCheckError: Layer 'inception_4a/1x1' references bottom 'pool3/3x3_s2' at the TRAIN stage however this blob is not included at that stage. Please consider using an include directive to limit the scope of this layer.

[sulthan...@triassicsolutions.com]

# Data transformation layers layer {

name: "train_transform"
type: "DetectNetTransformation" bottom: "data"
bottom: "label"
top: "transformed_data"
top: "transformed_label" detectnet_groundtruth_param: {

stride: 8
scale_cvg: 0.4
gridbox_type: GRIDBOX_MIN
coverage_type: RECTANGULAR
min_cvg_len: 20
obj_norm: true

image_size_x: 1000
image_size_y: 1000

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: 8
scale_cvg: 0.4
gridbox_type: GRIDBOX_MIN
coverage_type: RECTANGULAR
min_cvg_len: 20
obj_norm: true

image_size_x: 1000
image_size_y: 1000

layer {

top: 'bbox-list-class0' python_param {

module: 'caffe.layers.detectnet.clustering'

layer: 'ClusterDetections'

param_str : '1000, 1000, 8, 0.6, 3, 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-class0' python_param {

module: 'caffe.layers.detectnet.clustering' layer: 'ClusterGroundtruth'
param_str : '1000, 1000, 8, 1'

}

include: { phase: TEST stage: "val" } }

layer {
type: 'Python'
name: 'score-class0'
bottom: 'bbox-list-label-class0' bottom: 'bbox-list-class0'
top: 'bbox-list-scored-class0' python_param {

module: 'caffe.layers.detectnet.mean_ap'

layer: 'ScoreDetections' }

include: { phase: TEST stage: "val" } }

layer {
type: 'Python'
name: 'mAP-class0'
bottom: 'bbox-list-scored-class0' top: 'mAP-class0'
top: 'precision-class0'
top: 'recall-class0' python_param {

module: 'caffe.layers.detectnet.mean_ap' layer: 'mAP'
param_str : '1000, 1000, 8'

[ehol...@gmail.com]

Hi Sulthan,
I cant seem to see your entire network including the offending layer. It sounds like you cjanged something more than what would be expected like the layer name (bottom name) or in which stage it is used. If you c an post the entire network I would be happy to take a look.

[Sulthana Shafi]

Thank you ehol.Now iam using the custonnet as attached.But in contrary, to the net ,iam using actual dimension of 1000x1000;source LMDB file location is not as in the net used.Though i have used without changing kitti location and dimension except the stride to 8 for now to correct the error mentioned earlier. Also attaching the screenshots of graph obtained with the above net on my dataset.Even now MAP=0always.Please suggest the changes for detecting coverage as well as boundingbox.

[Sulthana Shafi]

Iam also attaching the old custom net and graph.pls have a look at it too.

On Friday, March 31, 2017 at 1:46:44 AM UTC+5:30, ehol...@gmail.com wrote:

[Bhargav Kanakiya]

Hello,

I have been trying to train a detectnet model to detect very small images. Do you have any suggestions apart from this post? How small were the objects of interest in your dataset? Please let me know.

Any help is much appreciated.

Thank you,

Bhargav

[Jon Watte]

The smart idea behind DetectNet is to detect:

- is this "block" part of a thing I want to detect?

- if so, where might the edge of that thing be within the block?

Then there's a separate pass that aggregates all of these detections, and calculates bounding boxes of actual detected things.

If your things are very small, you may instead want to use another approach.

For example, if your things are 10-15 pixels in size each way, you could sweep a 20x20 convolution kernel across the image by a stride of 4 or so, and you could have the entire convolution pipeline mirror a single-image detector like LeNet.

The output would then be a heat map of detected object locations.

This is a fairly traditional object detection architecture (before DetectNet) AFAICT, and you should be able to find literature on it with some searching.

Or just code it up, throw it at your training data set, and see what you get :-)

Sincerely,

jw

Sincerely,

Jon Watte


--
"I find that the harder I work, the more luck I seem to have." -- Thomas Jefferson

--
You received this message because you are subscribed to the Google Groups "DIGITS Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to digits-users+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/digits-users/c9781ee6-a4a8-4cdb-87e7-b6f58cbc9b3c%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Bhargav Kanakiya]

Thanks, Jon, for some pointers. I think I have got a fair idea of the DetectNet and have been playing around with the clustering parameters because, during inference, I am getting multiple bounding boxes for one object. I believe that the Non-Maximum Suppression could help me here rather than changing the clustering parameters.

To unsubscribe from this group and stop receiving emails from it, send an email to digits-users...@googlegroups.com.

[Alfe Sani]

Hello

So did you get any error like out of memory ? i am trying to use stride 8 with 1920 x 1080 resolution too . But when the training starts It shows me out of memory error as following 

Failed to allocate 133693440 bytes on device 0. Total memory: 12787122176, Free: 107347968, dev_info[0]: total=12787122176 free=107347968

I Have 12 gb memory in GPU and which should be enough for 1080 resolution images . though smaller stride can take more memory , but digits trying to allocate the memory which is much less than the total memory i have and showing that i have only very limited memory remaining . But in fact i haven't used my gpu for anything else during that time . Only digits and caffe . If you possibly know any way to help me out ?

Regards

Sani