Closed livst closed 6 years ago
If you post VGG+C3D network definition (at gist or here), I can try.
Try adding BN layers between convolutions, they work quite well with the VGG type definition. Why? Because it makes the optimisation less sensitive to the learning rate. Hope it helps. Cheers.
Closing this issue.
Hi,
I tried to use the 3D conv method with the VGG network, but the network doesn't converge, the loss doesn't go down. Since the prototxt is very large, i've posted just the first layers, all of the other layers are set just as these ones.