v1.1.0

This release does not encompass major changes to the repository, but it is associated with snapshotted weights available on Globus.

The Network saved here achieves about 0.8 mIoU on the validation set. Weights available here.

(As a note, the data is available here )

To run inference with this model, do something like this:

python /home/cadams/ThetaGPU/CosmicTagger/bin/exec.py --config-name SCC_21 mode=inference \
mode.weights_location=output/tensorflow/scc21/untether_sota_8nodes/checkpoints/ \
run.id=test-live \
data.data_directory=/lus/eagle/projects/datascience/CosmicTaggerSCC_21/data_files/ \
data.file=cosmic_tagging_val.h5 \
framework=tensorflow run.iterations=100

Inference results look like this (8xA100, mixed precision with XLA enabled):

2022-04-20 14:04:33,103 - INFO - Inference report: 218 batches processed for 6976 entries.
2022-04-20 14:04:33,103 - INFO -   plane0/Total_Accuracy: 0.9997
2022-04-20 14:04:33,104 - INFO -   plane0/Non_Bkg_Accuracy: 0.9404
2022-04-20 14:04:33,104 - INFO -   plane0/Neutrino_IoU: 0.6720
2022-04-20 14:04:33,104 - INFO -   plane0/Cosmic_IoU: 0.9220
2022-04-20 14:04:33,104 - INFO -   plane1/Total_Accuracy: 0.9997
2022-04-20 14:04:33,104 - INFO -   plane1/Non_Bkg_Accuracy: 0.9491
2022-04-20 14:04:33,105 - INFO -   plane1/Neutrino_IoU: 0.6779
2022-04-20 14:04:33,105 - INFO -   plane1/Cosmic_IoU: 0.9259
2022-04-20 14:04:33,105 - INFO -   plane2/Total_Accuracy: 0.9997
2022-04-20 14:04:33,105 - INFO -   plane2/Non_Bkg_Accuracy: 0.9467
2022-04-20 14:04:33,105 - INFO -   plane2/Neutrino_IoU: 0.6534
2022-04-20 14:04:33,105 - INFO -   plane2/Cosmic_IoU: 0.9234
2022-04-20 14:04:33,105 - INFO -   Average/Total_Accuracy: 0.9997
2022-04-20 14:04:33,106 - INFO -   Average/Non_Bkg_Accuracy: 0.9454
2022-04-20 14:04:33,106 - INFO -   Average/Neutrino_IoU: 0.6678
2022-04-20 14:04:33,106 - INFO -   Average/Cosmic_IoU: 0.9238
2022-04-20 14:04:33,106 - INFO -   Average/mIoU: 0.7958
2022-04-20 14:04:33,106 - INFO - Total time to batch_process: 115.9773
2022-04-20 14:04:33,106 - INFO - Total time to batch process except first iteration: 115.1700, throughput: 60.2935
2022-04-20 14:04:33,106 - INFO - Total time to batch process except first two iterations: 114.6399, throughput: 60.2932
2022-04-20 14:04:33,106 - INFO - Total time to batch process last 40 iterations: 21.1240, throughput: 60.5947

v1.0.2

This release is the last release to use 'argparse` instead of hydra-config and omegaconfig. Freezing it in case roll back is needed.

v1.0.1

The purpose of this release is to begin generating citeable DOIs on Zenodo.

v1.0.0

First full release of CosmicTagger. Build tests are passing. Since the latest update, many improvements:

• Streamlined IO interface
• Unified Learning rate scheduler
• Networks are identical in torch and tensor flow, and this is checked in the tests.
• Downsampling is dynamic, so there is only one file to use for each piece of the dataset at any resolution.

Release target after merging tf1.X

This release incorporates all the latest updates from tf and pytorch. It does not yet support TF2.0.

Version 1.0 Release candidate.

The release is meant to be a jumping off point for a stable version of this repository. There are many significant changes that have come and so it is good to snapshot and make sure there is a tag where everything works. Changes include:

• synthetic data preloading available (create synthetic data before training loop)
• Dropped support for previous input files at half resolution (they had a bug anyways), now only supporting full resolution input files. (If you want them, contact me)
• Added support for on-the-fly downsampling of input using the frameworks. Can run at half and even quarter resolution. Only works properly with new files.
• Added Continuous integration for network creation, training with real data at all resolutions, training with synthetic data at all resolutions, inference with a restored model. There is a "full" test suite that will test every major network architecture change but it is very expensive to run because of tensor flows overhead in graph creation. There is an IO-only test for real and synthetic data.
• Added a small test file at full resolution for CI tests as well as local tests. Only 5 events included.
• Added a requirements file for TravisCI, but also may be used to install requirements for this package. Note that pytorch and tensorflow are both in the requirements but are not necessarily both needed. The network is implemented in both frameworks.

Tests are passing with this freeze, which is good. I will also verify training converges in real situations. When this is all done, I will update to a "real" release from this pre-release.