CODEWITHSUNDEEP
caffetraining-data
Gio
Gio

Reputation: 1

Feed trainind data python layer (FCN CAFFE)

I'm new with caffe. I'm trying to train nyud-fcn32s-color-d from https://github.com/shelhamer/fcn.berkeleyvision.org. I don't understand how to manage the training data before the definition of solver and the training phase. I've downloaded nyud-v2 .mat dataset from http://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html, but I don't know how to feed it with the nyud_layer (python layer). Sorry for the question, I'm a newbie, it's important to me make this training so any help will be very appreciated.

Upvotes: 0

Views: 226

Answers (1)

rkellerm
rkellerm

Reputation: 5512

You should add tou your train/val prototxt the python layer, defined as 

layer {
  name: "color"
  type: "Python"
  top: "color"
  top: "hha"
  top: "label"
  python_param {
    module: "nyud_layers"
    layer: "NYUDSegDataLayer"
    param_str: "{\'tops\': [\'color\', \'hha\', \'label\'], \'seed\': 1337, \'nyud_dir\': \'../data/nyud\', \'split\': \'test\'}"
  }
}

And fill the correct params, as described in the implementation:

- nyud_dir: path to NYUDv2 dir
- split: train / val / test
- tops: list of tops to output from {color, depth, hha, label}
- randomize: load in random order (default: True)
- seed: seed for randomization (default: None / current time)

Upvotes: 1

Related Questions