torchmanager

Updates:

• Minor bugs fixed
• Typing improvement

API Updates:

• Add class index control for metrics.BinaryConfusionMetric
• Add EPOCH_START and BATCH_START to torchmanager_core.protocols.Frequency
• Add number of workers support for data.Dataset
• Add return_summary parameter for fit method to return training summary after training finished
• Basic configurations implementation with configs.Configs class
• Change metrics.ConfusionMetrics to abstract class
• Deprecated train package
• Implements losses.MAE, metrics.FeatureMetric, metrics.ExtractorScore, metrics.FID, metrics.F1, metrics.Precision, metrics.Recall, metrics.SSIM, metrics.Dice and metrics.PartialDice
• Introducing backward, forward function to do backward propagation and forward pass in Manager
• Introducing configs package
• Introducing random package with freezing/unfreezing random seeds
• Introducing a new version convert mechanism using torchmanager_core.protocols.VersionConvertible protocol
• Introducing data_parallel and reset method in Manager for multi-GPUs control
• Use forward method to calculate TP, FP, and FN and further calculate metric in forward_metric method in metrics.BinaryConfusionMetric

Other updates:

• Enables update on batch or epoch start for callbacks.FrequencyCallback
• Improved Exception handling during prediction
• Minor bugs fixed
• Removed deprecated methods

API Updates:

• Add class index control for metrics.BinaryConfusionMetric
• Add EPOCH_START and BATCH_START to torchmanager_core.protocols.Frequency
• Add number of workers support for data.Dataset
• Basic configurations implementation with configs.Configs class
• Change metrics.ConfusionMetrics to abstract class
• Deprecated train package
• Implements losses.MAE, metrics.FeatureMetric, metrics.ExtractorScore, metrics.FID, metrics.F1, metrics.Precision, metrics.Recall, metrics.SSIM, metrics.Dice and metrics.PartialDice
• Introducing backward, forward function to do backward propagation and forward pass in Manager
• Introducing configs package
• Introducing random package with freezing/unfreezing random seeds
• Introducing a new version convert mechanism using torchmanager_core.protocols.VersionConvertible protocol
• Introducing data_parallel and reset method in Manager for multi-GPUs control
• Use forward method to calculate TP, FP, and FN and further calculate metric in forward_metric method in metrics.BinaryConfusionMetric

Other updates:

• Enables update on batch or epoch start for callbacks.FrequencyCallback
• Improved Exception handling during prediction
• Minor bugs fixed
• Removed deprecated methods

RC Updates:

• Add return_summary parameter for fit method to return training summary after training finished

Updates:

• Minor bugs fixed

API Updates:

• Add EPOCH_START and BATCH_START to torchmanager_core.protocols.Frequency
• Add number of workers support for data.Dataset
• Basic configurations implementation with configs.Configs class
• Change metrics.ConfusionMetrics to abstract class
• Deprecated train package
• Implements losses.MAE
• Implements metrics.FeatureMetric, metrics.ExtractorScore, metrics.FID, metrics.F1, metrics.Precision, and metrics.Recall, metrics.SSIM, metrics.Dice and metrics.PartialDice
• Introducing backward, forward function to do backward propagation and forward pass in Manager
• Introducing configs package
• Introducing random package with freezing/unfreezing random seeds
• Introducing a new version convert mechanism using torchmanager_core.protocols.VersionConvertible protocol
• Introducing data_parallel and reset method in Manager for multi-GPUs control
• Use forward method to calculate TP, FP, and FN and further calculate metric in forward_metric method in metrics.BinaryConfusionMetric

Other updates:

• Enables update on batch or epoch start for callbacks.FrequencyCallback
• Improved Exception handling during prediction
• Minor bugs fixed
• Removed deprecated methods

RC Updates:

• Minor bugs fixed

API Updates:

• Add EPOCH_START and BATCH_START to torchmanager_core.protocols.Frequency
• Add number of workers support for data.Dataset
• Basic configurations implementation with configs.Configs class
• Change metrics.ConfusionMetrics to abstract class
• Implements losses.MAE
• Implements metrics.FeatureMetric, metrics.ExtractorScore, metrics.FID, metrics.F1, metrics.Precision, and metrics.Recall, metrics.SSIM.
• Introducing backward, forward function to do backward propagation and forward pass in Manager
• Introducing configs package
• Introducing random package with freezing/unfreezing random seeds
• Introducing a new version convert mechanism using torchmanager_core.protocols.VersionConvertible protocol
• Introducing data_parallel and reset method in Manager for multi-GPUs control
• Use forward method to calculate TP, FP, and FN and further calculate metric in forward_metric method in metrics.BinaryConfusionMetric

Other updates:

• Enables update on batch or epoch start for callbacks.FrequencyCallback
• Improved Exception handling during prediction
• Minor bugs fixed
• Removed deprecated methods

RC Updates:

• Deprecated _compile function in BasicManager
• Deprecated train package
• Implement metrics.Dice and metrics.PartialDice
• Minor bugs fixed

Updates:

• Minor bugs fixed

Updates:

• Add release candidate support for torchmanager_core.Version
• Minor bugs fixed

Updates:

• data.Dataset can now handle unsupervised dataset for reconstruction or dataset with packed dictionaries without further implementation of unpack_data
• Minor bugs fixed
• Move checkpiont property in callbacks._Checkpoint to read-only public
• Move from setup.py to pyproject.toml

Updates:

• Add Version class for better version checking
• Change metrics.ConfusionMetrics to abstract class
• Minor bugs fixed
• Move data.utils.sliding_window to data.sliding_window
• Removes data.utils package and redundant code

Updates:

• Extends losses.mse._ReductableLoss for losses.MSE
• Implements sliding_window function in data.utils package
• Minor bugs fixed
• Raise a TypeError when loading data.Dataset in a data.batched function instead of warning

API updates:

• Add softmax input and target control for torchmanager.losses.KLDiv loss
• Implement MSE loss with NAN replaced to zero
• Introducing errors package in torchmanager_core
• Introducing Resulting protocol
• Introducing torchmanager.callbacks.Experiment callback that combines torchmanager.callbacks.LastCheckpoint callback, multiple torchmanager.callbacks.BestCheckpoint callbacks, and torchmanager.callbacks.TensorBoard callback with formatted logs, which all saved in a wrapped .exp formatted folder
• Introducing torchmanager.data package
• Introducing torchmanager.losses.ParallelLoss for loss functions multi-GPUs support
• Option to replace NAN to zero after log calculation in KLDiv loss
• Support of listing checkpoints inside an experiment .exp folder using torchmanager.train.list_checkpoints method
• Support of loading a torchmanager.train.Checkpoint by .exp folder directory and checkpoint name using torchmanager.train.load method

Other updates:

• Performance improvement
• Minor bugs fixed

Updates:

• If a metric has not been called after reset, its result will return a nan.
• Introducing protected _metrics in EarlyStop callback
• Remove _update method implementation requirement for abstract class FrequencyCallback
• The loss weight can now be zero for non-activate

Updates:

• Add results property for the metrics to return all current results as a torch.Tensor
• Add threshold settings to MeanIoU metric
• Introducing FrequencyCallback with frequency support
• Introducing DynamicWeight callback that dynamically change the weight property in an object
• Introducing predict method in TestingManager to predict a dataset and returns a list of predictions
• Parsing device in fit and test as a list of specified devices

Updates:

• Managers now perform to torchmanager_core.devices.DeviceMovable and train.StateDictLoadable protocol
• Save weights only now supports any kind of objects that perform to StateDictLoadable protocol
• EarlyStop callback implementation to stop the training when monitored metrics not improvement for several epochs
• Loss now supports weight coefficient value
• Performance improved for multi-GPUs running
• Minor bugs fixed

Updates:

• Separate Manager into basic.BaseManager, basic.DataManager, testing.Manager, and training.Manager
• Add unpatch method (private) to unpatch data
• Add target support for multi outputs model in losses and metrics
• Minor bugs fixed

Updates:

• current_epoch index implementation
• Manager can be passed into Checkpoint as well as related callbacks to be saved
• LrScheduler callback now updates lr to fit the initial_epoch index
• Minor bugs fixed

Updates:

• Performance Improved
• Minor bugs fixed

Updates:

• Multi Losses support
• Multi Losses for multiple outputs support
• Logging support
• Minor bugs fixed

Main Release