search

xggIoU / centernet_tensorflow_wilderface_voc

This is the unofficial implementation of the "CenterNet:Objects as Points".Just a simple try with self-modified shufflenetv2 and yolov3.If you want better results, you need more experiments.
MIT License
111 stars 25 forks source link

heatmap loss convergence #11

Open rafikg opened 5 years ago

rafikg commented 5 years ago

@xggIoU Why the heatmap loss does not converge and what did you change? I am using DLA version and I had the same problem of convergence. I tried with mean_square_loss and focal_loss and still having convergence problem ? Some hints ?

xggIoU commented 5 years ago

Whether the loss converges mainly depends on the heat map produced. The main change is to ensure that the weight value outside the bbox is zero.

rafikg commented 5 years ago

Thanks @xggIoU, I think if you use the adaptive gaussian raduis from cornernet which mainly calculate the radius based on the size of the bbox, it will never had this problem. However, I have trained my DLA model around 80 epochs and the loss is still around ~4.0!!! I see that you trained few epochs and you got something on the training dataset!