Changelog

Version 3.11.0

Release Data: July 2026

Install: pip install netloader==3.11.0

Additions:

• Added MultiNetwork.

• Added error when importing NormFlow and NormFlowEncoder from architectures and SplineFlow without Zuko package.

• Added RNG state saving to BaseArchitecture.

• Added extra argument to _loss and _loss_tensor methods of BaseArchitecture.

• Added option for callable functions/objects for BaseArchitecture._loss_weights.

• Added BaseArchitecture._step method for each iteration.

• Added support for more schedulers in BaseArchitecture._update_scheduler.

• Added get_param_groups() method to define optimiser param groups.

• Added BaseArchitecture save frequency.

• Added description argument to BaseLayer.

• Added support for multiple layer outputs to BaseMultiLayer.

• Added forward pass to BaseMultiLayer to pack outputs.

• Added __add__() magic method.

• Added get_data() method.

• Added step() method.

• Added loader_init() function to initialise data loaders.

• Added activation_kwargs to Activation.

• Added GeneralLoss for any nn loss function.

• Added BaseNetwork.

• Added get_config() method.

• Added Branch.

• Added schedulers.

• Added BaseTypeTransform where input output has same type.

• Added support for loading networks with different weight names if weight shapes match.

• Added Network._logger attribute.

• Added Network._load_state_dict method.

• Added more type hints.

Changes:

• Changed Concatenate, Pack, Shortcut, Skip, & Unpack layers to be under multi_layer.

• Changed BaseArchitecture to only flag overwriting if base name is matching.

• Changed Data.__repr__ type to give specific type.

• Changed NormFlowEncoder to allow different learning rates for normalising flow and encoder.

• Changed BaseTransform to abstract class.

• Changed BaseTransform to allow different input output types.

• Changed safe_globals to allow only accepting classes defined in __all__.

• Changed BaseArchitecture to abstract class.

• Changed BaseLayer to abstract class.

• Changed Python version to 3.14.

• Changed package versions.

Removals:

• Removed support for networks saved in old deprecated weights format.

Deprecations:

• Deprecated networks for name change to architectures.

• Deprecated set_optimiser & set_scheduler for init_optimiser and init_scheduler methods.

Fixes:

• Fixed save() method not recovering on SystemExit.

• Fixed BaseArchitecture not changing Network device if compiled.

• Fixed BaseConfig.__setattr__ using incorrect type hints.

• Fixed back-propagation issue with Sample.

• Fixed loading RNG states if state was saved on GPU.

Version 3.10.1

Release Date: February 2026

Install: pip install netloader@git+https://github.com/EthanTreg/PyTorch-Network-Loader@2c1c4c3

Additions:

• Added Dropout layer.

• Added configs dataclasses for Network configuration.

• Added compare_versions() for version comparison.

• Added dict_list_append(), dict_list_convert(), and list_dict_convert().

• Added suppress_logger_warnings() wrapper to suppress warnings for a module.

• Added NetLoader version to BaseLayer.

• Added BaseLayer.__getstate__ magic method.

• Added compile() method.

• Added get_losses() method.

• Added loss weights to BaseArchitecture with get and set loss_weights methods and automatic calculation of total loss.

• Added ConvNeXt._stem private method.

• Added padding_mode keyword argument to convolutional layers.

• Added support for DataList inputs in data_collation().

• Added support for ModuleList with CompatibleNetwork.

• Added pool_kwargs keyword argument to Pool.

• Added packaging as requirement.

• Added Pad layer.

• Added get_save_path.

• Added apply() & apply() methods.

Changes:

• Changed BaseMultiLayer._target attribute to accept list[list[int]] shapes.

• Changed activation functions to use inplace where possible by default.

• Changed DataList length to return data length rather than list length.

• Changed Autoencoder to use a dictionary for loss terms.

• Changed Autoencoder to use a dictionary for loss weights.

• Changed NormFlowEncoder to use a dictionary for loss weights.

• Changed layers to be more compatible with compile().

• Changed package versions.

Removals:

• Removed deprecated BaseLayer.layers attribute.

Deprecations:

• Deprecated BaseArchitecture.save_path attribute and replaced with set_save_path() method.

• Deprecated Network.config attribute and replaced with get_config() method.

• Deprecated ArrayTC and replaced with ArrayCT.

• Deprecated Python versions < 3.12.

Fixes:

• Fixed utils not accessible from utils.

• Fixed batch printing from BaseArchitecture not using total loss.

• Fixed reduced precision training.

• Fixed total loss for loss functions returning dictionaries without total key.

• Fixed OrderedBottleneck blocking last feature.

• Fixed Checkpoint not saving combined outputs.

• Fixed to() method not working for non-tensors in checkpoints attribute.

• Fixed Normalise scale zero division.

• Fixed network name in BaseArchitecture.__repr__ magic method.

• Fixed formatting of BaseArchitecture.__repr__.

• Fixed BaseArchitecture._logger not initialised from saved state.

• Fixed getting BaseArchitecture.save_path deprecation.

• Fixed overwrite errors from set_save_path() if new save path is same as original.

• Fixed same asymmetric padding for convolutional and pooling layers.

• Fixed false deprecation warning in Composite layer

• Fixed incorrect attribute saving for AdaptivePool.

• Fixed type hints for Shapes.

Version 3.9.4

Release Date: October 2025

Install: pip install netloader@git+https://github.com/EthanTreg/PyTorch-Network-Loader@v3.9.4

Additions:

• Added base layers to layers.

• Added save_freq attribute to BaseArchitecture.

• Added keyword argument check to AdaptivePool.

• Added get_hyperparams() method.

Changes:

• Changed Network to not convert to CPU when saving.

• Changed some type hints.

Deprecations:

• Deprecated ArrayTC for ArrayCT.

Fixed:

• Fixed support for some old NetLoader versions.

• Fixed save corruption of BaseArchitecture during keyboard interrupt.

• Fixed keyword argument name clash in Index.

Version 3.9.1

Release Date: October 2025

Additions:

• Added BaseSingleLayer class.

• Added Shapes class.

• Added Data class for data & uncertainties.

• Added DataList class for storing multiple arrays or tensors.

• Added BaseArchitecture loss dictionary support.

• Added support to BaseArchitecture for list of transforms.

• Added DataList support to BaseArchitecture.

• Added BaseArchitecture._loss_tensor method.

• Added Data support to BaseTransform.

• Added ConvNeXt._head method.

• Added return_idxs to loader_init().

• Added data_collation() to collate lists of arrays, tensors, Data, and DataList objects.

• Added loss_func argument to BaseEncoder.

• Added GaussianNLLLoss class.

• Added Pack class.

• Added layer & checkpoint arguments to Unpack.

• Added support for DataList to several layers.

• Added CompatibleNetwork class.

• Added Network.version attribute.

• Added len, iter, and getitem magic methods to Network.

• Added label keyword argument to ascii_plot().

• Added several types in types module.

• Added BatchNorm layer.

• Added support to set total loss in BaseArchitecture if loss is a dictionary.

• Added option to only pack output from previous layer in Pack layer.

• Added root keyword argument to Network for loading networks and composite layers.

Changes:

• Changed layers to use Shapes class for layer shapes.

• Changed most methods and functions to not allow positional arguments for keyword arguments.

• Changed predict() to not return None values.

• Changed load_arch() to accept kwargs to pass to load().

• Changed ConvNeXtBlock to layers.

• Changed Index layer to use indexing or slicing.

• Changed type hints to be more consistent.

• Changed package versions.

Removals:

• Removed UNSET global variable.

Deprecations:

• Deprecated BaseLayer.layers attribute.

• Deprecated BaseLayer.forward method.

• Deprecated BaseArchitecture._loss_func method.

• Deprecated number and greater arguments in Index layer.

Fixes:

• Fixed ConvNeXt weight initialisation for biases that are None.

• Fixed Network ignoring layer_num attribute.

• Fixed Log uncertainty inplace operations.

• Fixed BaseDataset error message for incorrect attribute length.

• Fixed Concatenate layer when dim=-1.

• Fixed Index transform not detecting old state format.

• Fixed Normalise transform error for offset or scale ndarray keyword arguments.

• Fixed backwards compatibility with Python-3.11.

• Fixed type hint errors.

Version 3.8.0

Release Date: August 2025

Install: pip install netloader@git+https://github.com/EthanTreg/PyTorch-Network-Loader@v3.8.0

Additions:

• Added ConvNeXt implemented using classes.

• Added group attribute to BaseLayer.

• Added BaseArchitecture.losses functionality to save multiple loss values.

• Added print functions to BaseArchitecture for custom print logic during training & predicting.

• Added BaseArchitecture._loss_func method to return Tensor before BaseArchitecture._loss method returns float.

• Added extend() method to add multiple transforms.

• Added ASCII plot for loss curve during training.

• Added NetLoader version to BaseArchitecture save state.

• Added extra_repr method to DropPath.

• Added GitHub reference to timm for DropPath.

Changes:

• Changed loader_init() ratios argument to split the dataset after idxs.

• Changed Network forward pass to use BaseLayer rather than config for information.

• Changed Network to catch all exceptions during the forward pass.

Fixes:

• Fixed DropPath not working if prob is 0.

• Fixed NormFlowEncoder optimiser initialisation.

• Fixed optimiser and scheduler kwargs not being saved.

• Fixed Index weights only safe saving.

• Fixed Normalise uncertainty in-place operation.

• Fixed non-NetLoader classes being added to PyTorch safe globals.

• Fixed type hints.

Version 3.6.1

Release Date: June 2025

Install: pip install netloader@git+https://github.com/EthanTreg/PyTorch-Network-Loader@v3.6.1

Additions:

• Added BaseDataset for creating datasets.

• Added get_device & get_epochs to BaseArchitecture.

• Added overwrite keyword argument to BaseArchitecture to prevent unwanted file deletion.

• Added Jupyter Notebook package use example.

Changes:

• Changed input_ keyword argument to inputs in predict() method.

• Changed Network in_shape and out_shape arguments to accept tuples.

• Changed Index to allow negative in_shape dimension sizes for undefined lengths.

• Changed package requirements.

Fixes:

• Fixed docstrings & type hints.

Earlier Versions

See the GitHub Releases for complete version history.

Upgrade Guide

From 3.9.x to 3.11.x

No breaking changes. Existing code should be compatible.