olgaliak / seismic-deeplearning Goto Github PK

Currently the code assumes hardcoded "test" folders and splits.
Support for custom paths will make it possible to easily test on separate volumes of seismic data.

Currently the prepare_dutchf3.py selects the edge of the cube as test and the core as train.

It is important to train with all parts of the cube so when we run the tests/validation we have more diverse data

Add empty init to cv_lib/cv_lib/segmentation/dutchf3 so that our azureml_requirements.txt for AML can reference the microsoft repo and not the forked repo

Section slices could have quite a bit of duplicated data.
To shorten training time you could split every other (or even more) sections.

This functionality enables train\test on separate input seismic volumes

1. In the staging branch patch is not used to create the crosslines or inlines but it is in the master branch. When using this script the index generated in the txt files seem to be past our data dimension boundaries. We want to use patch to create the crosslines and inlines

2. The first patch highlighted is being used, but the second is not actually being used in the script as seen above. We should rename the second patch to patch_size to avoid confusion.

3. I created splits for two different volumes and compared them. If you have dimensions for example (400, 500, 350) with a stride of 50 and patch of 100. I would expect the master branch for crosslines to go from horz_locations = range(0, 400-100, 50) so it would go from 0 to 300 at 50 stride intervals so we would see a print out up to 250 there but I would like to see it go out to 300 since we have data until 400. Although if we add 300+1 this would fix this. So I am recommending to consider including a plus one at that point incase the dimensions are exactly on a mark like this.

Segy management unit test need to work with publicly available tiny segy file

Modify existing code samples and cofigs to include the option of resuming training.
For example: given a checkpoint from epoch 100 run the training for 50 more epochs.

Restart where training left off.
Update: Dutch_F3 notebook
Re-use min_epoch param in the config
param for weighs-path to resume -- double check what param we reuse

Add Segy data prep utils and needed unit tests (see also #1 ) to deepseimic repo

Add data normalization utils and needed unit tests (see also #1 ) to deepseimic repo

Work with Deepseimic team and pick a dataset that we can use to run AML pipeline on

Add config and supporting scripts. Validate on a sample dataset (see #2 )

The build failed when updating patch to patch_size in ablation.sh. Recreate this failed build so it can be investigated further.