Version history (through 0.0.33)#

Frozen history; releases up to and including 0.0.33. Newer entries live in the CHANGELOG.md at the repository root (rendered on the Changelog page).

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. 0.0.33 - 2026-10-11 —————————————————————————-

Added#

  • Add torch_ecg.utils.seed_everything (alias of DEFAULTS.set_seed) and torch_ecg.utils.worker_init_fn, a worker_init_fn for torch.utils.data.DataLoader which reseeds DEFAULTS.RNG per worker: PyTorch only reseeds the global random / torch / numpy.random states in workers, so with the default fork start method every worker inherited a copy of the parent’s DEFAULTS.RNG and all workers produced identical random streams for the augmentations.

  • Add codecov.yml with flag carry-forward, matching the tiered CI (see below).

  • test_download.py: regression tests for path traversal in tar extraction (sibling directory, .., absolute member names, symlink/hardlink escapes).

Changed#

  • CI workflow fixes: the docs workflow’s concurrency group is now scoped per ref (the previous global pages group with cancel-in-progress let any PR’s docs build cancel the docs builds of master/dev and of every other open PR); the test workflow also triggers on master pushes (validates the true merge commit and refreshes the codecov baseline); an ubuntu-24.04 canary job paves the way for the ubuntu-22.04 runner deprecation (full removal 2027-04); various workflow cleanups (obsolete setup.py install_egg_info workaround removed, dead importlib-metadata pin dropped, unused Hatch install removed, archived actions-ecosystem action replaced with the gh CLI).

  • Drop support for Python 3.9 (end-of-life since 2025-10; the latest torch already requires >=3.10) and add Python 3.14: requires-python >= 3.10, classifiers 3.10-3.14, and the CI matrix becomes 3.10 / 3.11 / 3.12 / 3.13 / 3.14 (3.11 was previously missing from the matrix; 3.13/3.14 were tested but not listed in the classifiers).

  • Replace the pyEDFlib dependency with the actively maintained pure-Python edfio (pyEDFlib’s last PyPI release was 2025-06, with no wheels for newer Python versions). The EDF reading used by the NSRR readers (e.g. SHHS) now goes through a small pyedflib.EdfReader-compatible adapter on top of edfio (torch_ecg.databases._edf), so no reader code changes and no behavior changes; verified by a round-trip test in test_base.py. This also unblocks Python 3.14 (pyEDFlib was the only dependency without compatible wheels). Cython, which was only needed to build pyEDFlib from source, is dropped from the dependencies.

  • Unify all randomness of the augmenters on DEFAULTS.RNG (mixup, cutmix, random_masking, stretch_compress, baseline_wander and the AugmenterManager used to draw from the Python random module). NOTE this changes the random streams: old seeds no longer reproduce old results.

  • DEFAULTS.set_seed now also seeds the global numpy.random state (see the in-code comment for the rationale); it used to seed random and torch but not numpy.random. The legacy numpy.random state is seeded with seed % 2**32 since np.random.seed rejects values beyond 32-bit.

  • Tiered CI (run-pytest.yml): pull requests and dev pushes run the fast lane (pytest -m "not db"), while the full suite including all database-downloading tests runs nightly and on manual dispatch (workflow_dispatch input suite). All test modules that downloaded databases at import (collection) time now download in session-scoped fixtures marked db which skip gracefully on download failure. Jobs get an explicit timeout-minutes so that a job killed by GitHub’s 6-hour limit fails visibly instead of showing up as “cancelled”.

  • Replace the Dropbox-hosted remote-model test fixtures of test_utils_nn.py with GitHub release assets (test-assets-v1), since the Dropbox links are traffic-suspended.

Fixed#

  • Fix an import-time race in the one-time database_citation.csv migration in CitationMixin: the unlink() after merging is now missing_ok=True. On Python 3.14+ the Linux default multiprocessing start method is forkserver (not fork), so DataLoader workers re-import the package concurrently; two workers could both pass the exists() check and both delete the file, and the loser crashed with FileNotFoundError at import (observed as DataLoader worker exited unexpectedly failing CI on 3.14).

  • Harden the safe tar extraction in download.py, completing the CVE-2007-4559 mitigation introduced in 2022 (commit 8f22086a), which correctly blocked the primary ../absolute-path vectors: the membership check now uses os.path.realpath + os.path.commonpath instead of string-prefix commonprefix (which a sibling directory whose name merely starts with the destination’s name could bypass), additionally refuses to extract members through symlink/hardlink members (which extractall would create first and then follow outside the destination), and validates that link members point inside the destination.

  • Remove all eval calls from the package (15 sites): config-driven code execution in make_attention_layer (unknown names now always raise ValueError) and setup_criterion (the "nn."-prefixed path now uses getattr); replaced dynamic name resolution with getattr / globals() / locals() / explicit dicts in cfg.py, misc.py, utils_signal.py, utils_metrics.py (including the wave-delineation metric collection in compute_metrics_waveform and its copy in ludb.compute_metrics), utils_nn.py, databases/base.py, ecg_arrhythmia_knowledge.py, and the plot methods of cinc2020/cinc2021/cpsc2021/sph/ludb. Behavior is unchanged except that invalid names now raise the documented errors instead of executing arbitrary expressions.

  • misc.timeout now uses a synchronous SIGALRM-based timer on POSIX main threads. The previous threading.Timer + PyThreadState_SetAsyncExc injection delivers the exception only at bytecode boundaries and could deadlock with the interpreter’s exception machinery (first observed hanging CI on Python 3.11 for hours); the thread-injection fallback is kept for Windows and non-main threads. test_read_event_scalars (downloads a real TensorBoard event file at test time) is now marked db so the core CI lane stays network-free.

  • Harden CFG in place (public API unchanged): - dict-typed fields are converted with the base CFG class instead

    of self.__class__, fixing the crash of CFG subclasses with stricter constructors (documented as _KNOWN_ISSUES in components/outputs.py: ClassificationOutput(..., d={"d": 1}) used to raise AssertionError); the issue notes are removed

    • keys that would shadow the dict API (keys/items/get/…) are now rejected with a ValueError instead of silently breaking the container

    • update deep-copies mapping values that replace non-mapping ones (e.g. None), so the updated config no longer shares mutable values with the config it was updated from

    • __repr__/__str__ keep callable values as <callable: ...> placeholders instead of silently hiding them

    • explicit __deepcopy__ that bypasses subclass __init__ and rebuilds both the item and the attribute views; the pickle format is untouched, so previously saved checkpoints still load

  • BaselineWanderAugmenter is fully vectorized and no longer spawns two multiprocessing.Pool instances per forward call (each forward now takes milliseconds instead of process-pool startup costs, and the fork-related duplication of random streams and Windows spawn incompatibilities are gone). The baseline wander amplitude and the Gaussian noise are now scaled by the ECG amplitude as documented (ampl_ratio are ratios of the signal amplitude); they used to be used as absolute values, with the per-lead amplitude computed but never used.

0.0.32 - 2026-10-08#

Added#

  • Implement a generic Registry class and establish registries for backbones, models, attention layers, preprocessors, augmenters, optimizers, schedulers, and losses.

  • Add forward_features and compute_features_output_shape methods to all CNN backbones to provide a standardized API for SSL and feature extraction.

Changed#

  • Refactor ECG_CRNN, PreprocManager, AugmenterManager, and BaseTrainer to utilize the new registry system for dynamic component construction and decoupling.

  • Enhance SizeMixin to support static shape inference for feature maps.

  • Make the function remove_spikes_naive in torch_ecg.utils.utils_signal support 2D and 3D input signals.

  • Use save_file and load_file from the safetensors package for saving and loading files in place of torch.save and torch.load in the CkptMixin class in torch_ecg.utils.utils_nn.

  • Add retry mechanism to the http_get function in torch_ecg.utils.download module.

  • Add length verification in the http_get function in torch_ecg.utils.download module.

  • Harden the http_get function in torch_ecg.utils.download module: retry and raise clear errors on persistent 202 (Accepted) queued downloads (e.g. figshare portal domains), refuse to save HTML error pages as archive files, and raise on empty (0-byte) downloads, instead of silently writing broken files.

  • Replace dead or unreliable database download URLs: the mirror of the CPSC2019 training set and the CACHET-CADB-Mini / SPH-Mini test subsets now point to Zenodo records; SPH and CACHET-CADB now use the central ndownloader.figshare.com file URLs (portal domains put downloads behind a queue that replies 202).

  • Move data downloads in database tests out of the module level (the pytest collection phase) into session-scoped fixtures which skip the test module on download failure, and mark those tests with the new db marker.

  • Fix the best-model checkpointing in BaseTrainer: the snapshot of the best weights is now a detached CPU copy (it used to be a reference to the parameter tensors and drifted along with the training), it is written to disk immediately when the monitored metric improves (so that the best weights survive a crash of the training process), and the saved file and the return value of train() now contain the best-epoch weights instead of the final-epoch weights. CkptMixin.save gains a state_dict parameter, and BaseTrainer falls back to torch.save for models without a save method.

  • Make BaseTrainer.resume_from_checkpoint raise an explicit NotImplementedError explaining why resuming is not supported by the current checkpoint format (it used to be a silent, untested half-implementation), and log the error detail when removing outdated checkpoints fails.

  • Fix the thr argument of ClassificationMetrics, which was silently ignored end-to-end (__call__ dropped it, and compute did not forward it to metrics_from_confusion_matrix), so passing e.g. thr=0.3 always computed metrics at the default 0.5.

  • Fix json.dump misuse in MITDBDataset and CINC2021Dataset (missing file-object argument), which crashed _slice_data with force_recompute=True after the full recomputation.

  • Fix the weight_mask branch in MITDBDataset for the af_event task: the condition never held (task == [list] instead of membership testing), and once fixed, the categorical rhythm_mask crashed generate_weight_mask (binary-only). The af_event labels are now binarized (AF vs non-AF, as for the rr_lstm task) before the weight mask is generated; the multiclass rhythm_segmentation target gets no weight mask.

  • Fix PhysioNetDataBase.download silently doing nothing when the AWS CLI is unavailable and the database has no compressed archive URL: it logged “Downloading the uncompressed version…” but returned without downloading; it now falls through to wfdb.dl_database.

  • Fix the .T transpose in preprocess_multi_lead_signal, which reversed all axes and corrupted outputs for 3-D (batched) channel_last inputs; only the last two axes are swapped now.

  • Validate at construction time that the torch Resample preprocessor has dst_fs or siglen set (with fs alone it used to crash at forward time with an obscure message).

  • Enhance the pytorch preprocessors module: all preprocessors now have pure PyTorch implementations, and NumPy fallbacks for NumPy array inputs.

Deprecated#

Removed#

Fixed#

  • Robustly handle dimension inference and initialization in ECG_CRNN models, especially for cases with None or Identity modules.

  • Address several compatibility issues for Python 3.13, including docstring indentation and NaN comparisons in dataclasses.

  • Improve error handling and encoding robustness in CitationMixin when reading cache files.

  • Resolve CodeQL warnings regarding incomplete URL substring sanitization in tests.

  • Correctly update the _df_metadata attribute of the PTBXL database reader classes after filtering records.

  • Enhance the save method of the torch_ecg.utils.utils_nn.CkptMixin class: non-safe items in the configs are removed before saving the model.

  • Fix errors when monitor is specified while the validation dataloader is not provided in the torch_ecg.components.BaseTrainer class.

  • Enhance the save_checkpoint method of the torch_ecg.components.BaseTrainer class: non-safe items in the configs are removed before saving the model.

  • Fix potential errors when deepcopying a torch_ecg.cfg.CFG object: previously, deepcopying such an object like CFG({“a”: {1: 0.1, 2: 0.2}}) would result in an error.

  • Fix potential bugs in contextmanager torch_ecg.utils.timeout: restore the previously installed SIGALRM handler after use, cancel any pending alarm reliably in a finally block, avoid installing a handler when duration <= 0 (preventing unintended global side-effects), and thereby eliminate spurious TimeoutError exceptions that could be triggered later by unrelated signal.alarm calls due to the old implementation not reinstating the original handler.

  • Fix bugs in utility function torch_ecg.utils.make_serializable: the previous implementation does not drop some types of unserializable items correctly. Two additional parameters drop_unserializable and drop_paths are added.

  • Fix CkptMixin.save() silently truncating checkpoint filenames that contain decimal values (e.g. …metric_0.91): pathlib.Path.with_suffix(".safetensors") treated .91 as the existing suffix and replaced it, producing …metric_0.safetensors instead of the correct …metric_0.91.safetensors. The method now appends .safetensors for paths without a recognised extension, and returns the final Path used so callers can track it.

  • Fix BaseTrainer checkpoint cleanup permanently failing: saved_models stored the raw stem path while the actual file on disk had a .safetensors suffix, causing every os.remove() call to raise FileNotFoundError. The trainer now stores the path returned by save_checkpoint(), and handles both single-file (os.remove) and directory (shutil.rmtree) checkpoints.

Security#

0.0.31 - 2025-01-28#

Added#

  • Add functions for downloading PhysioNet data from AWS S3. It is now made the default way to download data from PhysioNet.

  • Add easydict as a dependency for backward compatibility (loading old models using safe-mode torch.load with weights_only=True. Extra dependencies are added with torch.serialization.add_safe_globals).

Changed#

  • Test files (in the sample-data directory) are updated.

  • Add keyword argument weights_only to from_checkpoint and from_remote methods of the models (indeed the CkptMixin class). The default value is "auto", which means the behavior is the same as before. It checks if torch.serialization has add_safe_globals attribute. If it does, it will use safe-mode torch.load with weights_only=True. Otherwise, it will use torch.load with weights_only=False.

Deprecated#

  • Support for Python 3.7, 3.8 is deprecated. The minimum supported Python version is now 3.9. In the pyproject.toml file, the field requires-python is updated from >=3.7 to >=3.9.

Removed#

  • Restrictions on the version of wfdb and numpy packages are removed.

Fixed#

  • Fix IO issues with several PhysioNet databases.

Security#

  • Models are now loaded using safe-mode torch.load with weights_only=True by default.

0.0.30 - 2024-10-10#

Added#

  • Add support for AWS S3 in the download utility function http_get in the torch_ecg.utils.download module. PhysioNet now provides data download links from AWS S3, and the download utility function can now handle these links if AWS CLI is installed. This feature is implemented but not put into use yet.

Changed#

  • Change the default value of the method argument of the torch_ecg.utils.utils_signal.resample_irregular_timeseries function from "spline" to "interp1d". The former is probably not correctly implemented.

  • Update the logger classes: add checking of the write access of the log_dir. If the directory is not writable, the default log dir ~/.cache/torch_ecg/logs is used (ref. torch_ecg.cfg.DEFAULTS.log_dir).

  • Update the selection mechanism of the final model for the trainer classes: if no monitor is specified, the last model is selected by default (previously, no model was selected and saved).

  • The main part of the _setup_criterion method of the BaseTrainer class is moved to the function setup_criterion in the torch_ecg.models.loss module. The method is simplified and enhanced.

Deprecated#

  • Script setup.py is deprecated. The package building system is switched to hatch.

Removed#

  • Remove redundancy in base trainer classes: identical if blocks are removed from the _setup_criterion method of the BaseTrainer class.

Fixed#

  • Fix potential error in getting model name in the trainer classes.

  • Fix bugs in the CINC2020 and CINC2021 database reader classes for parsing the header files.

0.0.29 - 2024-07-21#

Added#

  • Add keyword argument with_suffix to function torch_ecg.utils.misc.get_record_list_recursive3.

  • Add function _download_from_google_drive to the torch_ecg.utils.download module for downloading files from Google Drive.

  • Add gdown as a dependency in the requirements.txt file.

  • Add database reader class PTBXLPlus for the PTB-XL+ database in torch_ecg.databases.physionet_databases.

  • Add github-release job to the publish action for creating a release on GitHub automatically.

Changed#

  • Improve the main training loop method of the base trainer class torch_ecg.components.trainers.BaseTrainer.

  • Allow passing additional keyword arguments to pass to requests.head in the url_is_reachable function of the torch_ecg.utils.download module (via adding the **kwargs argument).

  • Restrict version of numpy to be <=2.0.0 in the requirements.txt file. numpy version 2.0.0 is a breaking update, and a large proportion of the dependencies of this project are not compatible with it yet.

  • Enhance the cls_to_bin function and rename it to one_hot_encode in the torch_ecg.utils.utils_data module.

Fixed#

  • Enhance compatibility for different pandas versions.

  • Fix errors for taking length of an empty database reader class.

Security#

  • Fix code scanning alert - Incomplete regular expression for hostnames #21.

  • Fix code scanning alert - Incomplete URL substring sanitization #23.

0.0.28 - 2024-04-02#

Added#

  • Add CD workflow for the publish action with GitHub Action.

  • Add an optional argument return_fs for the load_data method for the database reader classes. If True, the sampling frequency of the record is returned along with the data as a tuple. To keep the behavior consistent, the default value is False.

  • Add an optional parameter fs for the function compute_receptive_field in the torch_ecg.utils.utils_nn module. If fs is provided, the receptive field is computed based on the sampling frequency.

  • Add method compute_receptive_field for several convolutional neural network models (layers) in the torch_ecg.models._nets module.

  • Add helper function make_serializable in the torch_ecg.utils.misc module for making an object serializable (with the json package). It will convert all numpy arrays to list in an object, and also convert numpy data types to python data types in the object recursively.

  • Add helper function url_is_reachable in the torch_ecg.utils.download module for checking if a URL is reachable.

  • Add database reader class PTBXL for the PTB-XL database in torch_ecg.databases.physionet_databases.

  • Add class method from_remote for CkptMixin classes. It is used to load a model from a remote location (e.g., a URL) directly.

  • Add sphinx-emoji-favicon as a dependency for generating the favicon for the documentation.

  • Add utility function ecg_plot from ecg-image-kit.

  • Add pyarrow as a dependency in the requirements.txt file.

  • Add benchmark study train_crnn_cinc2023 for the CinC2023 challenge.

Changed#

  • Change the default value reset_index of the utility function torch_ecg.utils.utils_data.stratified_train_test_split from True to False.

  • Enhance the decorator torch_ecg.utils.misc.add_kwargs so that the signature of the decorated function is also updated.

  • Update the documentation: use sphinx_toolbox.collapse and sphinxcontrib.bibtex; add citation info in the index page.

  • Make Dataset classes accept slice index for the __getitem__ method.

Deprecated#

  • Support for Python 3.6 is deprecated. The minimum supported Python version is updated to 3.7.

Removed#

  • Remove broken links in the docstrings of the database reader classes.

  • Remove unused scripts formatting.sh and push2pypi.sh.

Fixed#

  • Fix errors in the decorator torch_ecg.utils.misc.add_kwargs when a bound method is decorated.

  • Fix bugs related to data overflow for preprocessor classes that work with numpy arrays as reported in issue #12.

  • Fix bugs in augmentor class StretchCompress in the torch_ecg.augmenters module.

  • Fix dtype error when calling compute_class_weight from sklearn.utils.

  • Fix the issue when handling nan values in in computing metrics.

  • Fix errors for the ApneaECG database reader class when passing a path that does not exist or a path that contains no records at initialization.

0.0.27 - 2023-03-14#

Added#

  • Add default configs for blocks of the ResNet model in the torch_ecg.models.cnn module.

  • Add RegNet model in the torch_ecg.models.cnn module.

  • Add CutMix augmentor in the torch_ecg.augmenters module.

  • Add support for torch.nn.Dropout1d in the models.

  • Add .readthedocs.yml to the project. The documentation is now hosted on Read the Docs besides GitHub Pages.

Changed#

  • Move torch_ecg.utils.preproc to torch_ecg.utils._preproc.

  • Allow embed_dim of SelfAttention layer not divisible by num_heads via adding a linear projection layer before the multi-head attention layer.

  • Documentation is largely improved.

Deprecated#

  • Drop compability for older versions of torch (1.5 and below).

Removed#

  • Remove protobuf from the requirements.txt file.

  • Clear unused methods in the CINC2020 and CINC2021 database reader classes.

  • Clear unused layers in the torch_ecg.models._nets module.

  • Remove the torch_ecg.utils._pantompkins module. It contains the implementation of the Pan-Tompkins algorithm for QRS detection, modified from old versions of the wfdb package. It is moved to the legacy folder of the project.

  • Remove WandbLogger class from the torch_ecg.components.loggers module.

Fixed#

  • Fix bugs when passing units=None for the load_data method of the PhysioNet database reader classes.

0.0.26 - 2022-12-25#

Added#

  • Add a default load_data method for physionet databases reader classes in the base class PhysioNetDataBase. In most cases, in the inherited classes, one does not need to implement the load_data method, as the default method is sufficient. This method is a slight improvement over wfdb.rdrecord.

  • Add decorator add_kwargs in the torch_ecg.utils.misc module for adding keyword arguments to a function or method.

  • Add functions list_databases, list_datasets in the torch_ecg.datasets module for listing available databases reader classes and Dataset classes.

  • Add save method for the CkptMixin class. It is used to save the model to a file.

  • Add _normalize_leads a method of the base _DataBase class in the torch_ecg.databases.base module. It is used to normalize the names of the ECG leads.

  • Add subsampling functionality for database reader classes.

  • Add benchmark study train_mtl_cinc2022 for the CinC2022 challenge.

  • Add CITATIONS.bib file for storing BibTeX entries of the papers related to the project.

  • Add 10 sample data from the CPSC2018 database for testing in the sample-data directory.

Changed#

  • Use CitationMixin from the bib-lookup package as the base class for the DataBaseInfo class in torch_ecg.databases.base.

  • Use CitationMixin as one of the base classes for the models in torch_ecg.models.

  • Allow dummy (empty) preprocessor managers, a warning instead of an error is raised in such cases.

  • Enhance error message for the computation of metrics.

  • Add keyword argument requires_grad and include_buffers to the torch_ecg.utils.utils_nn.compute_module_size function. The dtype argument is removed as the data type of the model is now inferred from the model itself.

  • Improve several database reader classes: CPSC2018, CPSC2021, CINC2017, ApneaECG, MITDB, SPH.

  • Add asymmetric zero pad for convolution layers, so that when stride = 1 and kernel_size is even, strict "same" padding is conducted.

  • Use loggers instead of print in database reader classes.

  • Integrate code coverage into the CI workflow. The coverage report is generated and uploaded to Codecov.

  • More unit tests are added, and the existing ones are updated. Code coverage is largely improved.

Deprecated#

  • Drop compatibility for tqdm < 4.29.1

Removed#

  • Remove unused rpeaks detection methods in the torch_ecg.utils.rpeaks module.

  • Remove _normalize_leads method in LUDB database reader class.

  • Remove unused functions in the file of the CPSC2020 database reader class.

Fixed#

  • Fix bugs in the config class torch_ecg.cfg.CFG.

  • Fix errors in the plot method of CINC2020 and CINC2021 database reader classes.

Security#

  • CVE-2007-4559 patch: Fix a potential security vulnerability in the torch_ecg.utils.download.http_get function.

0.0.25 - 2022-10-08#

Added#

  • Add docstring utility function remove_parameters_returns_from_docstring in torch_ecg.utils.misc.

  • Add abstract property database_info to the base class _DataBase in torch_ecg.databases.base so that when implementing a new database reader class that inherits from the base class, its DataBaseInfo must be implemented and assigned to the property.

  • Add method get_citation to the base abstract class _DataBase in torch_ecg.databases.base which enhances the process for getting citations for the databases.

  • Add database reader class CACHET_CADB for the CACHET-CADB database in torch_ecg.databases.other_databases.

  • Add download method for the base abstract class CPSCDataBase in torch_ecg.databases.base.

Changed#

  • Improve the warning message for passing an non-existing path when initializing a database reader class.

  • Change the default behavior of the download method for PhysioNetDataBase class: default to download the compressed version of the database.

  • Update the README file in the torch_ecg/databases directory.

Fixed#

  • Use register_buffer in custom loss classes for constant tensors to avoid potential device mismatch issues.

  • Rename and update the data file physionet_dbs.csv.tar.gz to physionet_dbs.csv.gz to comply with the changement of the pandas.read_csv function from version 1.4.x to 1.5.x.

  • Fix the incorrect usage of NoReturn type hints. It is replaced with None to indicate that the function/method does not return anything.

0.0.24 - 2022-08-13 [YANKED]#

This release was yanked.

0.0.23 - 2022-08-09#

Added#

  • Add collate_fn as an optional argument for BaseTrainer class in torch_ecg.components.trainers.

Changed#

  • Let db_dir attribute of the database reader classes be absolute when instantiated, to avoid potential pathlib errors.

  • Update utility function torch_ecg.utils.utils_nn.adjust_cnn_filter_lengths`: avoid assigning unnecessary fs to dict-type config items; change default value of the pattern argument from "filter_length|filt_size" to "filter_length|filter_size" to avoid unintended changement of configs for BlurPool (in torch_ecg.models._nets).

  • Enhance error message for BlurPool in torch_ecg.models._nets.

0.0.22 - 2022-08-05#

Changed#

  • Make utility function torch_ecg.utils.utils_data.default_collate_fn support dict type batched data.

  • Update docstrings of several metrics utility functions in torch_ecg.utils.utils_metrics.

0.0.21 - 2022-08-01#

Added#

  • Add utility function get_kwargs in torch_ecg.utils.misc for getting (keyword) arguments from a function/method.

  • Add AHA diagnosis statements in torch_ecg.databases.aux_data.

  • Add argument reset_index to the utility function torch_ecg.utils.utils_data.stratified_train_test_split.

  • Add typing-extensions as a dependency in the requirements.txt file.

  • Add database reader class QTDB for the QTDB database in torch_ecg.databases.physionet_databases.

Changed#

  • Enhance data handling (typicall when using the load_data method of the database reader classes) with precise dtypes via torch_ecg.cfg.DEFAUTLS.

  • Update the setup of optimizer for the base trainer class torch_ecg.components.trainers.BaseTrainer.

  • Update the DataBaseInfo class for the SPH database.

  • Update the README file in the torch_ecg/databases directory.

  • Update plotted figures of the benchmark studies.

  • Rename SequenceLabelingOutput to SequenceLabellingOutput (typo fixed) in the torch_ecg.components.outputs module.

  • Enhance docstring of LUDB database reader class via updating its DataBaseInfo class.

  • Append the _ls_rec method as the last step in the download method of the database reader classes.

  • Change torch_ecg.utils.utils_data.ECGWaveForm from a namedtuple to a dataclass.

Removed#

  • bib_lookup.py is removed from the project. It is now delivered in an isolated package bib_lookup published on PyPI, and added as a dependency in the requirements.txt file.

  • Remove unnecessary script exec_git.py.

  • Remove joblib in the requirements.txt file.

0.0.20 - 2022-06-15#

Added#

  • Add database reader class SPH for the SPH database in torch_ecg.databases.other_databases.

  • Add dataclass DataBaseInfo for storing information of a database. It has attributes title, about, note, usage, issues, reference, etc., and has a method format_database_docstring for formatting the docstring of a database reader class. The generated docstring can be assigned to corresponding database reader class via the add_docstring decorator (in torch_ecg.utils.misc).

  • Add default cache directory ~/.cache/torch_ecg for storing downloaded data files, model weight files, etc.

  • Add helper function is_compressed_file for checking if a file is compressed in torch_ecg.utils.download.

0.0.19 - 2022-06-09#

Added#

  • Add argument relative to the utility function get_record_list_recursive3.

  • Add attribute _df_records to the database reader classes. The attribute stores the DataFrame of the records of the database, containing paths to the records and other information (labels, demographics, etc.).

Fixed#

  • Fix bugs in the download utility function http_get.

  • Fix bugs in the database reader classe CPSC2021.

0.0.18 - 2022-06-05#

Added#

  • Add property in_channels for the models.The number of input channels is stored as a private attribute _in_channels, and the property in_channels makes it easier to access the value.

  • Add warning message to the download method of the CPSC2019 database reader class.

  • Add get_absolute_path method for the database reader classes to uniformly handle the path operations.

Changed#

  • All all absolute imports are replaced with relative imports.

  • Update citation and images for several benchmark studies

  • Update the downlaod link for the CPSC2019 database reader class (ref. property torch_ecg.databases.CPSC2019.url).

Removed#

  • Remove the torch_ecg.utils.misc.deprecate_kwargs decorator. It is delivered in an isolated package deprecate_kwargs published on PyPI, and added as a dependency in the requirements.txt file.

Fixed#

  • Fix errors in the _ls_rec method of the CPSC2019 database reader class.

  • Fix bugs in the torch_ecg.utils.misc.deprecate_kwargs decorator.

  • Fix the issue that tensorboardX is incompatible with the latest version of protobuf.

0.0.17 - 2022-05-03 [YANKED]#

This release was yanked.

0.0.16 - 2022-04-28#

Added#

  • Add method _categorize_records for the MITDB database reader class, categorize records by specific attributes. Related helper properties beat_types_records and rhythm_types_records are added.

  • Add method _aggregate_stats for the MITDB database reader class. Related helper properties df_stats and db_stats are added.

  • Add function cls_to_bin for converting categorical (typically multi-label) class labels to binary class labels (2D array with 0/1 values).

  • Add context manager torch_ecg.utils.misc.timeout for setting a timeout for a block of code.

  • Add context manager torch_ecg.utils.misc.Timer to time the execution of a block of code.

  • Add module torch_ecg.components.inputs for input data classes.

  • Add class Spectrogram (in torch_ecg.utils) for generating spectrogram input data. This class is modified from the torchaudio.transforms.Spectrogram.

  • Add decorator torch_ecg.utils.misc.deprecate_kwargs for deprecating keyword arguments of a function/method.

  • Top-level module torch_ecg.ssl for self-supervised learning methods and models is introduced, but not implemented yet.

  • Add helper function torch_ecg.utils.utils_nn.compute_sequential_output_shape to simplify the computation of output shape of sequential models.

  • mobilenet_v3 model is added to the torch_ecg.models module. It is now available as a cnn backbone choice for the ECG_CRNN model (and for other downstream task models).

Changed#

  • Use numpy’s default rng for random number generation in place of np.random and Python built-in random module.

  • Update the README file.

  • Move the function generate_weight_mask from CPSC2021 dataset to torch_ecg.utils.utils_data.

  • Database reader MITDB is enhanced: add properties df_stats_expanded; add arguments beat_types and rhythm_types to the data and annotation loading methods.

  • Downloading function http_get is enhanced to support downloading normal files other than compressed files.

  • Update __init__ file of the torch_ecg.utils module.

  • Database reader class CinC2017 is updated: add property _validation_set.

  • The ECG_UNET model is simplified by removing the unnecessary zero padding along the channel axis.

  • Update the README file.

Deprecated#

  • Keyword argument batch_norm in model building blocks (ref. torch_ecg.models) is deprecated. Use norm instead.

Removed#

  • Redundant functions in torch_ecg.utils.utils_interval are removed: diff_with_step, mask_to_intervals.

Fixed#

  • Remove redudant code for the ECG_UNET model which might cause error in computing output shapes.

0.0.15 - 2022-04-14#

Changed#

  • Use pathlib.Path.parents instead of sequence of pathlib.Path..parent to get the parent directory of a file path.

  • Type hints and docstrings of some database reader classes are enhanced: ApneaECG, CINC2020, CINC2021.

  • Update the README file: add citation information for the package.

0.0.14 - 2022-04-10#

Added#

  • Implements the lead-wise mechanism (as a method _assign_weights_lead_wise) for the Conv_Bn_Activation layer in the torch_ecg.models._nets module.

  • Implements assign_weights_lead_wise for model MultiScopicCNN (in torch_ecg.models).

  • Zenodo configuration file .zenodo.json is added.

Changed#

  • Update the README file: add :point_right: [Back to TOC](#torch_ecg) to the end of long sections.

Fixed#

  • Fix errors in the computation of classification metrics.

0.0.13 - 2022-04-09#

Added#

  • Add metrics computation class WaveDelineationMetrics for evaluating the performance of ECG wave delineation models.

  • Add methods for computing the metrics to the output classes (in the module torch_ecg.components.outputs).

  • Add script push2pypi.sh for pushing the package to PyPI.

  • Add attribute global_pool_size to the configuration of the classification models (torch_ecg.models.ECG_CRNN).

Changed#

  • flake8 check ignore list is updated.

  • README is updated.

Removed#

  • Usage of easydict is removed. Now we use torch_ecg.cfg.CFG for configuration.

Fixed#

  • Computation of the metric of mean_error for ECG wave delineation is corrected.

  • Fix bugs in SpaceToDepth layer (torch_ecg.models.resnet).

0.0.12 - 2022-04-05#

Changed#

  • Some out-of-date sample-data files are updated, unnecessary files are removed.

  • Passing a path that does not exist to a database reader class now raises no error, but a warning is issued instead.

  • Include isort and flake8 in the code formatting and linting steps. Code are reformatted and linted.

0.0.11 - 2022-04-03#

Changed#

  • Docstrings are cleaned up.

  • Unit tests are updated.

0.0.10 - 2022-04-01#

Added#

  • Add BibLookup class for looking up BibTeX entries from DOIs of papers related to datasets and models.

  • Add RPeaksDetectionMetrics class to the torch_ecg.components.metrics module for evaluating the performance of R-peaks detection models.

  • Add CI workflow for running tests via GitHub Actions.

Changed#

  • The loading methods (load_data, load_ann, etc.) of the database reader classes are enhanced to accept int type record name argument (rec), which redirects to the record with the corresponding index in the all_records attribute of the database reader class.

0.0.9 - 2023-03-30#

Added#

  • Add decorator add_docstring for adding/modifying docstrings of functions and classes.

  • Add method append for the BaseOutput class.

  • Add several metrics computation functions in torch_ecg/utils/utils_metrics.py:

    • confusion_matrix

    • ovr_confusion_matrix

    • auc

    • accuracy

    • f_measure

    • QRS_score

  • Add top-level module torch_ecg.components.

  • Add classes for metrics computation to the torch_ecg.components.metrics module.

Changed#

  • Dataset classes and corresponding config classes are added to the __init__.py file of the torch_ecg.databases.dataset module so that they can be imported directly from the module.

  • Logger classes, output classes, and trainer classes are moved to the new module torch_ecg.components.

  • Callbacks in BaseTrainer are enhanced, allowing empty monitor, and allowing non-positive number of checkpoints to be saved (i.e., no checkpoint is saved).

0.0.8 - 2022-03-29#

Fixed#

  • Bugs in extracting compressed files in the http_get function of the utils.download module.

Security#

0.0.7 - 2022-03-28#

Fixed#

  • Import errors for early versions of pytorch.

  • Cached table of PhysioNet databases is added as package_data in setup.py to avoid the error of missing the table file when installing the package.

Security#

0.0.6 - 2022-03-28#

Added#

  • Add methods __len__ and __getitem__ for the base class torch_ecg.databases.base._DataBase.

Changed#

  • The base class of CPSC2021 is changed from CPSCDataBase to PhysioNetDataBase.

  • Function compute_output_shape is enhanced to support different paddings in two ends of the input signal.

  • README is updated.

  • Docstrings of many classes and functions are updated.

  • black is used for code formatting.

0.0.5 - 2022-03-27#

Added#

  • Add cached table of PhysioNet databases as a data file stored in the package.

  • Add requests as a dependency in the requirements.txt file.

Changed#

  • An optional argument btype is added to the function butter_bandpass_filter to specify the type of the filter: "lohi", "hilo".

  • A compressed argument is added to the download method of the PhysioNetDataBase class to specify whether to download the compressed version of the database.

Fixed#

  • Fix bugs in the function preprocess_multi_lead_signal.

0.0.4 - 2022-03-26#

Added#

  • Add ReprMixin class for better representation of the classes (e.g., models, preprocessors, database readers, etc.).

  • Added model_dir to default config.

  • Add Dataset classes for generating input data for the models:

    • CINC2020

    • CINC2021

    • CPSC2019

    • CPSC2021

    • LUDB

  • Add sample-data directory for storing sample data for testing.

  • Add url property to the database classes.

  • Add utility functions for the computation of metrics.

  • Add BeatAnn class for better annotation of ECG beats.

  • Add download utility functions.

  • Add Output classes for the models. The output classes are used to store the output of the models and provide methods for post-processing.

Changed#

  • Manipulation of custom preprocessor classes is enhanced.

  • SizeMixin class is improved for better computation of the sizes of the models.

  • Replace os with pathlib, which is more flexible for path operations.

  • Several database reader classes are updated: mitdb, ltafdb.

  • Improve PhysioNetDataBase by using wfdb built-in methods of getting database version string and downloading the database.

  • Update the README file.

Removed#

  • Unnecessary imports are removed.

Fixed#

  • Fix bugs in the flush method of the TxtLogger.

0.0.3 - 2022-03-24 [YANKED]#

This release was yanked.

0.0.2 - 2022-03-04#

Added#

  • Add Preprocessor classes for ECG data preprocessing (ref. torch_ecg.preprocessors).

  • Add Augmenter classes for ECG data augmentation (ref. torch_ecg.augmenters).

  • Add database reader classes for reading ECG data from different sources (ref. torch_ecg.databases).

  • Add model classes for ECG signal analysis, including classification, segmentation (R-peak detection, wave delineation, etc., ref. torch_ecg.models).

  • Add several benchmark studies for ECG signal analysis tasks:

    • CinC2020, multi-label classification.

    • CinC2021, multi-label classification.

    • CPSC2019, QRS detection.

    • CPSC2020, single-label classification.

    • CPSC2021, single-label classification.

    • LUDB, wave delineation.

    ref. the benchmarks directory of the project.

  • Add documentation for the project (ref. docs directory).

  • Add CodeQL action for security analysis (ref. .github/workflows).

  • Add unit tests for the project (ref. test directory).

0.0.1 - 2022-03-03 [YANKED]#

This release was yanked.