Ctlearn

ctapipe-stage1 migration

Major Features
* Support of the official CTA stage 1 files (the dl1dh data format is still supported but will be deprecated in the future)
* Upgrade to dl1dh v0.10.0 and ctapipe v0.10.5
* Added ResNet-RNN model
* New feature: Freeze the backbone in deep stereo models like the ResNet-RNN
* pypi installation for CTLearn

Minor Improvements
* Add CTA and MAGIC example files
* Improve input file handling in predict mode to process real data in standard convention

Bug Fixes and Other Changes

DL full event reconstruction

Major Features
* Enable energy and arrival direction regression
* Add residual blocks to construct ResNet architectures
* Modify singleCNN model to construct VGG models LucaRomanato
* Enable attention mechanism
* Add drawio diagrams nietootein
* Count examples by class and regression values aribrill

Minor Improvements
* Add ability to overwrite the random seed from the command line
* Add new ctlearn mode to recursively train and predict
* Predict on several file lists at once
* cttearn as a command line tool
* Update from plotting scripts aribrill
* Add notebook to visualize IACT regression metrics using ctaplot
* Upgrade TensorFlow version to 1.15.3

Bug Fixes and Other Changes

Remarks
* Stereo models haven't been benchmarked with this version.

v0.4.0-legacy

Major Features
- Replaced `DataLoader`, `DataProcessor`, and `ImageMapper` with `DL1DataReader` from the `DL1-Data-Handler` package (115, 82, 46, 73).
- Greatly revised configuration format to use DL1DH parameter names and inputs and simplify all sections.

Minor Improvements
- Simplified `input_fn` in `run_model` to remove unnecessary required parameters (41).
- Added option to list data files directly in configuration file (40).
- Added explicit dictionary of label names to list of class names to configuration file.
- Added `load_only` mode in `run_model` to load the data and print info without running a model.
- Simplified `predict` mode in `run_model` to iterate through the data only once.

Bug Fixes and Other Changes
- Added kludge when loading data to manually convert unsigned dtypes to the next-higher signed dtype, as TensorFlow cannot automatically perform this conversion.
- Removed scripts specific to processing and image mapping (`plot_camera_image` and `visualize_bounding_boxes`).
- Removed scripts made obsolete by `load_only` mode (`print_dataset_metadata` and `print_run_metadata`).
- Moved `test_image_mapper` notebook to DL1DH.
- Replaced direct dependencies on Astropy, OpenCV, Pillow, PyTables, and SciPy with a dependency on DL1-Data-Handler installed using pip.

Major Features

Minor Improvements
- Upgraded Python version to 3.7.3.
- Upgraded TensorFlow version to 1.13.1.

Bug Fixes and Other Changes
- Fixed typo in image mapper.

Major Features
- Added FACT, H.E.S.S.-I, H.E.S.S.-II, and MAGIC cameras to `ImageMapper`.
- Added bilinear interpolation, bicubic interpolation, nearest neighbor interpolation, rebinning, image shifting, and axial addressing image mapping methods in `ImageMapper`.
- Added support for running models using data of multiple telescope types.
- Added `use_peak_times` data loading option to load peak arrival times from data files.

Minor Improvements
- Added `auto_configuration.py` script to automatically change the paths in benchmark configuration files.
- Added argument in `run_multiple_configurations.py` to resume from a particular run.
- Added `summarize_results.py` to summarize the results of a set of runs.
- Rationalized the metadata variables returned by `DataLoader`.
- Added `test_image_mapper.ipynb` for testing the image mapping methods of `ImageMapper`.
- Changed telescope names for compatibility with [ctapipe](https://github.com/cta-observatory/ctapipe) camera names.
- Refactored `ImageMapper` to implement all image mapping methods as matrix operations, so that more expensive calculations are performed only during initialization.

Bug Fixes and Other Changes
- Fixed `DivisionByZeroError` in `apply_cuts` during `HDF5DataLoader` initialization.
- Renamed internal variables in and generally cleaned up `DataLoader`.
- Refactored `DataLoader._load_metadata()` into smaller functions for clarity and efficiency.
- Fixed incorrect logging of array examples by class.
- Changed model loading to use included CTLearn models by default.
- Added contributing guidelines.
- Made `run_model.py` append the CTLearn version number to config files.
- Updated TensorFlow version to v1.12.
- Added benchmark configuration files for CTLearn v0.3.0.
- Removed deprecated models.