Config.json reformatting

[snowsignal]

I think that we should change some aspects of config.json parsing. First of all, we should add a new value, 'predict', in the main JSON object, which will contain "weight", and "images" sub-properties. We also should add a 'production' value, which contains "timeout" and "tmplocation" sub-properties. Here's a potential example of what it could look like:

{
  "model" : {
    "architecture": "Tiny Yolo",
    "input_size": 416,
    "anchors": [0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828],
    "max_box_per_image": 3,
    "labels": ["bucket"]
  },

  "train": {
    "train_image_folder": "/scratch/coe_andrew/bucket/bucket-dataset/images/",
    "train_annot_folder": "/scratch/coe_andrew/bucket/bucket-dataset/annotations/",

    "train_times": 15,
    "pretrained_weights": "",
    "batch_size": 24,
    "learning_rate": 1e-5,
    "nb_epoch": 50,
    "warmup_batches": 250,

    "object_scale": 5.0 ,
    "no_object_scale": 1.0,
    "coord_scale": 1.0,
    "class_scale": 1.0,

    "saved_weights_name": "save_tiny.h5",
    "debug": true
  },

  "valid": {
    "valid_image_folder": "",
    "valid_annot_folder": "",

    "valid_times": 1
  },
  "predict": {
    "weight": "data/dataset.h5",
    "images": [
      "images/test.jpg",
      "images/test2.jpg"
    ]
  },
  "production" : {
    "timeout" : 5,
    "tmplocation" : "tmp/",
    "weight" : "data/dataset.h5"
  }
}

The "weight" property determines what weight file to use when testing, if one is not specified in the command lines. "images" contains a test batch of images to run the predict function on. In the "production" object, "timeout" is the maximum processing time for an image in the processing queue (in seconds), "tmplocation" is the location to write out video camera input to (and to read from), and "weight" specifies the weight file to use.

I think that this should probably be done before we actually start working on the camera reading - or "production" script.

This would involve doing some rewriting on the predict.py script, and I think that we should work on seperating the CLI parsing from predict.py before we start working on it.

[snowsignal]

I'm going to go ahead and start working on this.