New metric - TopKCategoricalAccuracy

• Added metric TopKCategoricalAccuracy
• Added Predictor.from_trainer() method to Predictor class
• Fixed loading predictor from_checkpoint if the checkpoint is not Full Trainer
• Added unittest for TopKCategoricalAccuracy

Verbosity

• Added verbosity support in Trainer.train(verbose=1) and Trainer.evaluate(verbose=1)
• StatsPrint validation bug fix
• Added unittest for StatsPrint validation

Predictor class

• Added Predictor class !! predicting was never easier, see README for more details
• Added example for train/save/load/predict using the new Predictor
• Added unittests for predictor
• Metrics optimizations

Predictor class

• Added Predictor class !! predicting was never easier, see README for more details
• Added example for train/save/load/predict using the new Predictor
• Added unittests for predictor

ModelCheckpoint and EarlyStopping using CallbackMonitors

• Added trainer validation for metric_name_to_func
• ModelCheckpoint args changed to accept CallbackMonitor
• EarlyStopping args changed to accept CallbackMonitor
• Adjusted examples and tests
• Added more unittests

num_epochs to Trainer.train(), better StatsPrint and unittest

• Added test for trainer save and load
• Moved LossOptimizerHandlerBase validation to Trainer.train() instead if Trainer.init
• Added name property to CallbackMonitorResult
• StatsPrint now accept list of monitors arguments for metrics
• StatsPrint will make validations on provided monitors
• Removed num_epochs from Trainer arguments, now its in Trainer.train(num_epochs)
• Adjusted all examples

Loss and Optimizer handling via callbacks

• Added predict_sample and predict_data_loader methods to Trainer
• Added LossOptimizerHandler and LossOptimizerHandlerBase to callbacks
• Trainer must have at least one callback of type LossOptimizerHandlerBase
• Removed optimizer_step_and_zero_grad_criteria argument from Trainer (use LossOptimizerHandler callback instead)
• Added optimizer, scheduler, and train_last_loss to CallbackContext properties for easier access
• CollectOutputs's arguments now must be explicitly provided
• CallbackBase will raise an exception if call not implemented
• Trainer has callbacks validation upon initialization, more validations will be added
• SchedulerStep's scheduler_parameters_func should accept CallbackContext instead of Trainer
• Added copy_model_weights to lpd.utils.torch_utils, (as requested, thank you for using lpd 🥳)
• Adjusted all examples

Predicting

• Added predict_batch
• Adjusted predict example
• Added threshold argument to BinaryAccuracy and BinaryAccuracyWithLogits

Metrics

• Custom metrics are now in a separate module
• Custom metrics now classes instead of functions
• Added unittests for metrics

Predicting

• Added predict method to trainer
• Added State.PREDICT enum
• Added Phase.PREDICT_BEGIN and Phase.PREDICT_END enums
• Added CollectOutputs callback
• Added predict example
• Moving some elements from nn.functional to nn
• Added sample count to save/load model