hyperparameter & pre-train issues

[muye5]

• Does this detection model train totally from scratch with COCO-Dataset or VOC?
  Should we do pre-train with ImageNet?

• GRID_W = 18 & GRID_H = 10
  How to set these two hyper-parameters ?
  (e.g. input image size: 32 * (13, 13) = (416, 416) as described in paper ? ?)

• The anchor file missed

Thanks!

[wojciechmo]

1. If you would like to train with COCO or VOC you should download and prepare data by yourself. But in such case it is easier to use trained model - as point 2 in description states: Conversion from Darknet to Tensorflow framework done with darkflow project. Point of this project is to train with your own data which are summarized in csv file.

2. GRID_W and GRID_H are spatial dimensions of final 3D tensor, where each cell of GRID_WxGRID_H makes a vector with predicted rois and classes. For GRID_W = 18 and GRID_H = 10 input training images should be of size 572x300.

3. As it's written in description you should prepare two files before training - one is anchor file. You can specify anchors dimensions as you wish, or you can use some clustering algorithm as it is done in paper.

Best regards

[colwind]

could i use the tfrecord file that generated in other repository?

[Tveek]

@WojciechMormul ,how to calculate 572x300 by GRID_WxGRID_H ? but,in make_tfrecord.py file,in_w, in_h = grid_w*net_scale, grid_h*net_scale ,so strange？