/yolo2_tensorflow

Primary LanguagePython

yolo2_tensorflow

this project is the implements of the yolo2 paper from scrath, i will record the troble and complete detail here

Troble:

  1. NAN , LOSS INCrease ERROR

    when i use :

    priors = tf.reshape(Anchors, [1, 1, 1, anchor_num, 2])
pred_wh = tf.exp(pred[..., 3:5]) * priors

    it always occurred NAN error, because pred[..., 3:5] always over than hundred after 1~2 train epoch,

    And train loss is always been biger in training, i need to figure it out

Implement Detail:

  1. auto_kemeans: use kemeans to generate the anchor_box automaticlly