Skip to content
This repository has been archived by the owner on Nov 16, 2023. It is now read-only.

Release Notes v0.1.2
Compare
Choose a tag to compare
@maxkazmsft maxkazmsft released this 20 May 17:29
· 6 commits to master since this release
v0.1.2
e136533
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

April 2020 release, our goal was to harden the current capabilities in the repository and ensure correctness in the results. We made specific investments this month to make sure users of the repository can produce the stated results and reproduce them. Developers can now leverage Docker and Jupyter Lab to train and develop on any platform. We have added a checkerboard dataset which will helps users validate and test the code. Users of the repository can now train and validate data in either the inline or crossline directions, or both simultaneously. We have also improved the Tensorboard visualizations and added the capability to export intermediate data in debug mode.

New features:

  • Docker image now ships with Jupyter Lab to allow users to train and develop code on any platform. #237 #255 #242
  • Added training and validation set logging to Tensorboard with better visualizations: also refactored the train.py script for Dutchf3 to enhance readability, maintainability. #236
  • Added the ability to train and validate the data in the inline or crossline directions only. #253
  • Added a new checkerboard dataset to allow users to assess the correctness of the codebase. #249

Enhancements:

  • Multi-GPU parallel builds: speeds up build execution for Pull Requests. #246
  • Completely removed OpenCV as a dependency from the DeepSeismic Repo. #240
  • Removes the HorovodLRScheduler class since its no longer used. #240

Bugfixes:

  • Fixed an issue where the generated training and validation splits can overlap (in different directions). #233 #143
  • Hardcoded number of classes in Dutch F3 dataloader: can now specify a variable number of classes for Dutch F3 dataset, e.g. map some classes as a single class; also allows for checkerboard dataset or any other dataset to be loaded through Dutch F3 format. #144 #285 #145
  • Fixed a visualization bug where the seismic images were not normalized correctly. #217
  • Fixed a visualization bug where the predictions were not masked where the input image was padded. This improves the ability to visually inspect and evaluate the results. This solves Issue #230.
  • Fixes a potential issue where Tensorboard can crash when a large training batchsize is used. Now the number of images visualized in Tensorboard from every batch has an upper limit.
  • Fixes Issue where the epoch number for the images in Tensorboard was always logged as 1 (therefore, not allowing the user to see the epoch number of the different results in Tensorboard. #218
  • Fixes an issue where the seeds used for the notebooks are not the same as those used for the training scripts #244
  • Fixed an issue where code in prepare_dutchf3.py wasn't using the stride to create patch locations #287
  • Fixed an issue where the borders of images shown in Tensorboard were not displayed correctly #230
Assets 2
Loading