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

How to initialize base model #8

Closed aeon0 closed 7 years ago

aeon0 commented 7 years ago

Did you initialize the base classifier layers with pre-trained values when training the full ssd model? I have a custom dataset, but my model will not converge to anything. I was wondering if I have to train my base classifier first with my custom dataset before training the ssd model.

georgesung commented 7 years ago

I did not use any pre-trained weights, but it would have been a good idea. You can try to use pre-trained ImageNet weights if possible (look for AlexNet pre-trained ImageNet weights). I didn't have time to implement loading the pre-trained weights, but that would have been a good addition if time permitted.