I'm using a Unet for PixtoPix conversion of seismic data. It works great with one exception. It doesn't handle extreme values of the input correctly. My input is a 2d tensor which I normalize with mean and std. After the normalization, the data distribution is Gaussian-like (with mean=0 and std=1), but with a long tail of extreme values. Extreme values are important for my problem as they contain signals I'm trying to extract from overall noisy input, so I cannot crop these values. The dynamic range of the input can vary over orders of magnitude and it is dataset-specific. Overall my global Unet model works great, but it fails for those extremes. Any ideas of how to handle outliers in the input correctly? Is there any amplitude clipping within the network layers (Conv2d, ConvTranspose2d, BatchNorm2d, LeakyReLU, MaxPool2d, Dropout) that may affect the overall result? I'm using PyTorch to build CNN architecture.

New contributor
Julia is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.

Your Answer

Julia is a new contributor. Be nice, and check out our Code of Conduct.

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.