• Results of Training the Customised Network

    Capt. Flatus O'Flaherty ☠ • 12/12/2018 at 13:57 • 0 comments

    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!

  • Baby steps

    Capt. Flatus O'Flaherty ☠ • 12/11/2018 at 16:15 • 0 comments

    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:...
    Read more »