search

georgesung / ssd_tensorflow_traffic_sign_detection

Implementation of Single Shot MultiBox Detector in TensorFlow, to detect and classify traffic signs
MIT License
530 stars 222 forks source link

Train loss and Validation loss are very high!! #34

Open Jasonsun1993 opened 6 years ago

Jasonsun1993 commented 6 years ago

I trained the whole LISA dataset, but when the train code was running. I found that the Train loss and Validation loss were up to 2600! Does anybody know what the problem is ??

xinxueying commented 6 years ago

I have the same problem. The trainning loss at beginning is over 800, and so as the validation loss. And after about 20000 round training, it drop to about 200.

xinxueying commented 6 years ago

I did some edit to some api, so it can train under tensorflow1.4.1 . Besides this, nothing changed.

Jasonsun1993 commented 6 years ago

@xinxueying After training, did you test the images in samples_images folder? Did it have good results?

xinxueying commented 6 years ago

@Jasonsun1993 Problem solved， I changed the learning rate of training to 0.5 (original code use the default 0.01,which is very small considering training from scratch): OPT = tf.train.AdadeltaOptimizer() change to OPT = tf.train.AdadeltaOptimizer(learning_rate=0.5)

Jasonsun1993 commented 6 years ago

@xinxueying Really? Did you train the whole data set?? How about the results?? Besides, can I have your Wechat number? I want to discuss about this code further.

zengxyu commented 6 years ago

Have you got the dataset? I can't open the link LISA Traffic Sign Dataset(http://cvrr.ucsd.edu/LISA/lisa-traffic-sign-dataset.html). If you got the dataset, could you please send it to me?

pageedward commented 5 years ago

@xinxueying what is the val loss when u ended training?