Training framework & tools for PyTorch-based machine learning projects.
This package provides a training framework and CLI for PyTorch-based machine learning projects. This is free software distributed under the Apache Software License version 2.0 built by researchers and developers from the Centre de Recherche Informatique de Montréal / Computer Research Institute of Montreal (CRIM).
To get a general idea of what this framework can be used for, visit the FAQ page. For installation instructions, refer to the installation guide. For usage instructions, refer to the user guide. The auto-generated documentation is available via readthedocs.io.
Development is still on-going — the API and internal classes may change in the future.
- Added SelectChannels transform operator for inplace sample modifications that might need it.
- Fixed conda package builds for tagged deployments
- Refactored and cleaned up HDF5 data extraction/parsing classes
- Added dataset interfaces for BigEarthNet, Agri-Vis challenge
- Update classification task to allow multi-label classification
- Added activation layer customization for in-framework ResNet archs
- Updated default move_tensor behavior to be non-blocking
- Added trainer implementation for auto-encoder-type models
- Added Orion reporting support for hyperparameter explorations
- Added SLURM cluster utilities (tmpdir getter, launch scripts)
- Skip image save call during metric rendering if the provided value is None as employed by basic logger/reporter.
- Add JSON implementation for thelper.train.utils.ClassifLogger.
- Fix concepts to handle any variation of upper/lower concept name.
- Employ requirements.txt within conda-env.yml to kept dependencies in sync.
- Fixes built docker image not using appropriate dependencies enforced through requirements.txt.
- Fix version comparison check when validating configuration and/or checkpoint against package version. Version can now have a release part which was not considered.
- Fix incorrect calculation of sample coordinates in thelper.data.geo.parsers.SlidingWindowDataset.
- Remove not_skip = __init__.py config option for isort since __init__.py is included since 4.3.5. Also force isort<5 since many import checks break suddenly (e.g.: direct import with as alias break).
- Update this changelog to use rst links (renders on github and readthedocs)
- Add infer mode for classification of geo-referenced rasters
- Add Dockerfile-geo to build thelper with pre-installed geo packages
- Add geo-related build instructions to travis-ci build steps
- Add auto-documentation of makefile targets and docker related targets
- Removed optional dependencies from conda build env
- Travis deploy test w/ split conda/docker stages
- Split travis deploy stage into two phases
- Fixed draw_segment threshold usage & params lookup
- Fixed FCResNet embedding getter wrt latest pooling update
- Update all matplotlib plots to use 160 dpi by default
- Refactor trainer data/metric writer to save all viz data
- Added viz pkg w/ t-SNE & UMAP support for in-trainer usage
- Fixed geo pkg documentation build issue related to mocking
- Fixed type and output format checks in numerous metrics
- Updated all callback readers to rely on new utility function
- Cleaned and optimize coordconv implementation
- Added U-Net architecture implementation to nn package
- Added IoU metric implementation
- Added support for SRM kernels and SRM convolutions
- Updated documentation (install, faq, maintenance)
- Added fixed weight sampler to data package
- Added lots of extra unit tests
- Added efficientnet 3rd-party module wrapper
- Fixed potential conflicts in task class names ordering
- Fixed pytest-mock scope usage in metrics utests
- Updated common resnet impl to support segmentation heads
- Fixed samples usage for auto-weighting of loss functions
- Cleaned up samples usage in loader factory data splitter
- Add GDL compatibility module to geo package
- Fix segmentation task dontcare default color mapping
- Cleaned up and simplified coordconv implementation
- Update segmentation trainer to use long-typed label maps
- Cleaned up augmentor/albumentations demo configurations
- Removed travis check in deploy stage for master branch
- Added geo subpackage
- Added geo vector/raster parsing classes
- Added ogc module for testbed15-specific utilities
- Added testbed15 train/viz configuration files
- Cleaned up makefile targets & coverage usage
- Replaced tox build system with makefile completely
- Merged 3rdparty configs into setup.cfg
- Updated travis to rely on makefile directly
- Added extra logging calls in trainer and framework utils
- Cleaned up data configuration parsing logger calls
- Bypassed full device check when specific one is requested
- Moved drawing utilities to new module
- Cleaned up output root/save directory parsing
- Cleaned up potential circular imports
- Moved optional dependency imports inside relevant functions
- Added support for root directory specification via config
- Updated config load/save to make naming optional
- Fixed potential issue when reinstantiating custom ResNet
- Fixed ClassifLogger prediction logger w/o groundtruth
- Add cli/config override for task compatibility mode setting
- Cleaned up dependency lists, docstrings
- Fixed bbox iou computation with mixed int/float
- Fixed dontcare label deletion in segmentation task
- Cleaned up training session output directory localization
- Fixed object detection trainer empty bbox lists
- Fixed exponential parsing with pyyaml
- Fixed bbox display when using integer coords values
- Fixed collate issues for pytorch >= 1.2
- Fixed null-size batch issues
- Cleaned up params#kwargs parsing in trainer
- Added pickled hashed param support utils
- Added support for yaml-based session configuration
- Added concept decorators for metrics/consumer classes
- Cleaned up shared interfaces to fix circular dependencies
- Added detection (bbox) logger class
- Fixed nn modules constructor args forwarding
- Updated class importer to allow parsing of non-package dirs
- Fixed file-based logging from submodules (e.g. for all data)
- Cleaned and API-fied the CLI entrypoints for external use
- Fixed travis timeouts on long deploy operations
- Added output path to trainer callback impls
- Added new draw-and-save display callback
- Added togray/tocolor transformation operations
- Cleaned up matplotlib use and show/block across draw functions
- Fixed various dependency and logging issues
- Fixed torch version checks in custom default collate impl
- Fixed bbox predictions forwarding and evaluation in objdetect
- Refactored metrics/callbacks to clean up trainer impls
- Added pretrained opt to default resnet impl
- Fixed objdetect trainer display and prediction callbacks
- Refactored metrics/consumers into separate interfaces
- Added unit tests for all metrics/prediction consumers
- Updated trainer callback signatures to include more data
- Updated install doc with links to anaconda/docker hubs
- Cleaned drawing functions args wrt callback refactoring
- Added eval module to optim w/ pascalvoc evaluation funcs
- Fixed issues when reloading objdet model checkpoints
- Fixed issues when trying to use missing color maps
- Fixed backward compat issues when reloading old tasks
- Cleaned up object detection drawing utilities
- Fixed travis conda build dependencies & channels
- Update documentation use cases (model export) & faq
- Cleanup module base class config backup
- Fixed docker build and automated it via travis
- Fix metrics RawPredictions not returning predictions during eval
- Fix parsing of checkpoint base path
- Added dockerfile for containerized builds
- Added object detection task & trainer implementations
- Added CLI model/checkpoint export support
- Added CLI dataset splitting/HDF5 support
- Added baseline superresolution implementations
- Added lots of new unit tests & docstrings
- Cleaned up transform & display operations
- Cleaned up build tools & docstrings throughout api
- Added user guide in documentation build
- Update tasks to allow dataset interface override
- Cleaned up trainer output logs
- Added fully convolutional resnet implementation
- Fixup various issues related to fine-tuning via ‘resume’
- Updated conda build recipe for python variants w/ auto upload
- Added framework checkpoint/configuration migration utilities
- Fixed minor config parsing backward compatibility issues
- Fixed minor bugs related to query & drawing utilities
- Fix travis-ci conda build/env path
- Fix travis-ci conda channel setup
- Fix openssl dependency
- Fixed travis-ci matrix configuration
- Added travis-ci deployment step for pypi
- Fixed readthedocs documentation building
- Updated readme shields & front page look
- Cleaned up cli module entrypoint
- Fixed openssl dependency issues for travis tox check jobs
- Updated travis post-deploy to try to fix conda packaging (wip)
- Added typedef module & cleaned up parameter inspections
- Cleaned up all drawing utils & added callback support to trainers
- Added support for albumentation pipelines via wrapper
- Updated all trainers/schedulers to rely on 0-based indexing
- Updated travis/rtd configs for auto-deploy & 3.6 support
- Added regression/segmentation tasks and trainers
- Added interface for pascalvoc dataset
- Refactored data loaders/parsers and cleaned up data package
- Added lots of new utilities in base trainer implementation
- Added new unit tests for transformations
- Refactored transformations to use wrappers for augments/lists
- Added new samplers with dataset scaling support
- Added baseline implementation for FCN32s
- Added mae/mse metrics implementations
- Added trainer support for loss computation via external members
- Added utils to download/verify/extract files
- Minor fixups and updates for CCFB02 compatibility
- Added RawPredictions metric to fetch data from trainers
- Fixed readthedocs sphinx auto-build w/ mocking.
- Refactored package structure to avoid env issues.
- Rewrote seeding to allow 100% reproducible sessions.
- Cleaned up config file parameter lists.
- Cleaned up session output vars/logs/images.
- Add support for eval-time augmentation.
- Update transform wrappers for multi-channels & lists.
- Add gui module w/ basic segmentation annotation tool.
- Refactored task interfaces to allow merging.
- Simplified model fine-tuning via checkpoints.
- Completed first documentation pass.
- Fixed travis/rtfd builds.
- Fixed device mapping/loading issues.
- Initial release (work in progress).
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size thelper-0.6.2.tar.gz (471.8 kB)||File type Source||Python version None||Upload date||Hashes View|