Closed jediofgever closed 4 years ago
@WangYueFt , Thank you for your reply, Also , if my point cloud contains more than say 2048 points, what will be expected behaviour of network to it , do I need to have strictly 2048 points ?
@WangYueFt , Thank you for your reply, Also , if my point cloud contains more than say 2048 points, what will be expected behaviour of network to it , do I need to have strictly 2048 points ?
You don't need to. But in each batch, you have to pad them into the same number of points.
@WangYueFt, I did create data files same as .h5 format and trained network but even in epoch 0 I got 0.99 train accuracy.
Train 0, loss: 1.376507, train acc: 0.993896, train avg acc: 0.993896
Test 0, loss: 1.233287, test acc: 1.000000, test avg acc: 1.000000
Train 1, loss: 1.233263, train acc: 1.000000, train avg acc: 1.000000
Test 1, loss: 1.233124, test acc: 1.000000, test avg acc: 1.000000
Train 2, loss: 1.233195, train acc: 1.000000, train avg acc: 1.000000
Test 2, loss: 1.233121, test acc: 1.000000, test avg acc: 1.000000
Train 3, loss: 1.233166, train acc: 1.000000, train avg acc: 1.000000
Test 3, loss: 1.233120, test acc: 1.000000, test avg acc: 1.000000
Train 4, loss: 1.233150, train acc: 1.000000, train avg acc: 1.000000
Test 4, loss: 1.233116, test acc: 1.000000, test avg acc: 1.000000
Does this mean model if overfitting ?
Hey there @jediofgever
I'm attempting to train (the equivariant version (VNN) of) this model on my own dataset as well. I think I may be overlooking something, as I'm running into an error.
If you or @WangYueFt could take a look at it and give me some pointers, or if I need to add some more info, that'd be much appreciated.
Link to my issue on the VNN github: https://github.com/FlyingGiraffe/vnn/issues/13
You probably need to do two things: