maxkazmsft

Extend GenericShiftedDataset with stride and max_samples_per_ts functionalities

4

**Is your feature request related to a current problem? Please describe.** Right now `GenericShiftedDataset` only strides the data by 1 time step - in some applications, I might want to...

# Features - generation of 3D output in test.py (3D segmentation maps) with *geometric* and *arithmetic* feature map averages in inline and crossline directions - added generation of segy files...

Fix seeds in both train and test python drive scripts to make sure results are reproducible on re-run on a single type of GPU. Switching GPUs will most likely change...

When using patch-based and section-based approaches, generate whole volume (score the entire test set) and write that volume into segy format at the end of scoring; for approaches where crosslines...

add unit test to check that each input slice has uniform distribution - facilitates uniform sampling of training data for better model accuracy

1

Related to #259 - this is essentially a test for it can test that for checkerboard dataset train and val patches have a uniform 50/50 class distribution across all patches

add Azure ML Training Pipeline Jupyter notebook which showcases the use of recently added pipeline as detailed here: https://github.com/microsoft/seismic-deeplearning/blob/contrib/interpretation/deepseismic_interpretation/azureml_pipelines/README.md Probably a good idea to re-write this README into notebook format

create HRNet Azure ML notebook based on existing HRNet notebook - showcases the use of Azure ML

1

blocked by Azure ML pipelines

Notebook which describes how to load patch based and section based data

Loader should work with generic 3D segy files; these should have optional metadata but not required.