search

bourdakos1 / Custom-Object-Detection

Custom Object Detection with TensorFlow
https://medium.freecodecamp.org/tracking-the-millenium-falcon-with-tensorflow-c8c86419225e
MIT License
347 stars 181 forks source link

scores drop to zero #9

Closed yossitsadok closed 6 years ago

yossitsadok commented 6 years ago

Trying to reproduce your demo, I run your training example with your images and "faster_rcnn_resnet101.config" . Although the loss drops gradually the prediction scores drop fast to zero after few hundred of steps. Changing the learning rate didn't solved it.

Do you have any suggestion?

bourdakos1 commented 6 years ago

How many images are you using to train and how many classes of objects? Also, how does your dataset compare to the coco dataset?

yossitsadok commented 6 years ago

I was using the dataset you provided (Star Wars images) and faster_rcnn_resnet101_coco as the base model. I'm using the default parameters and definitions that in your config file faster_rcnn_resnet101.config

bourdakos1 commented 6 years ago

Oh, okay, maybe try running it longer? It should work to train the provided dataset out of the box, I recommend at least running it for 1,000 steps to see if it goes up at all. If you have any screenshots of the output that might help me help you a bit more