search

szaza / dataset-generator

It is a very simple program to generate Pascal VOC style learning dataset from images. It generates images and XML style annotations with bounding box coordinates.
Do What The F*ck You Want To Public License
11 stars 4 forks source link

What should be possible loss to train YOLO model? #1

Open msarfrazcss opened 5 years ago

msarfrazcss commented 5 years ago

Hey guys i am tarining a model on my custom dataset, Can you tell me please What should be possible loss to train YOLO model, below is the training console:

step 343 - loss 17.77151107788086 - moving ave loss 18.94973713932652 step 344 - loss 17.025156021118164 - moving ave loss 18.757279027505685 step 345 - loss 17.319908142089844 - moving ave loss 18.6135419389641 step 346 - loss 17.478084564208984 - moving ave loss 18.49999620148859 step 347 - loss 17.533782958984375 - moving ave loss 18.403374877238168

At what loss i should stop the training?

szaza commented 5 years ago

Hi, There is no such a value what should be the loss. It is only a value, which shows you the difference (distance) between the expected and the detected object. During the training phase it will converge to a value. Machine learning lessons from Coursera helped me a lot to understand how the training phase works, you can also check them here: https://www.coursera.org/learn/machine-learning. Probably, this part contains the full answer for your question: https://www.coursera.org/lecture/machine-learning/gradient-descent-intuition-GFFPB;