-
|
I'm kind of confused about the best way to implement a Convolution block. I'm working with 3D medical images (so spatial_dims=3) and have specified dropout_dim=3, but not set norm_dim. Additionally, I set adn_ordering="AND" (activation → normalization → dropout). However, in my training, I'm observing the performance is worse / training is unstable. Thank you for any suggestions! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hi, setting Unstable training can have many reasons. Make sure to start with a small simple network and gradually make it more complex. A Recipe for Training Neural Networks by Andrej Karpathy is probably the best starting point I know. Good look! :) |
Beta Was this translation helpful? Give feedback.
Hi,
setting
spatial_dims=3is correct in this case.dropout_dimandnorm_dimdepend on your architecture. I would advice starting with copying prominent architectures like theResNetfor classification/regression and theUNetfor segmentation tasks.Unstable training can have many reasons. Make sure to start with a small simple network and gradually make it more complex. A Recipe for Training Neural Networks by Andrej Karpathy is probably the best starting point I know. Good look! :)