Fixes

• Minor change to MatMul2D, use torch.matmul instead of torch.bmm
• Fixed tqdm dependancy
• Fixed predict_sample when sample is an array of input

Bug fix and LossOptimizerHandlers

• Bug fix when saving full trainer that has tensorboard callback
• Added LossOptimizerHandlerAccumulateSamples
• Added LossOptimizerHandlerAccumulateBatches
• Added is_file_exists method to file_utils

TensorboardImage + code cleaning

• Added new callback - TensorboardImage
• Added example for TensorboardImage
• Added torchvision to requirements-dev.txt
• In Trainer - metric_name_to_func (dict) was changed to metrics (list)
• Trainer now holds _last_data property - a class InputOutputLabel that holds (inputs, outputs, labels)
• New method in file_utils ensure_folder_created

lpd-nodeps

• Added lpd-nodeps package (PyPI) in case you need to handle your own dependencies

Confusion Matrix - phase 3

• Improved handling of MetricConfusionMatrixBase with custom metrics

Previously on lpd:

• CallbackMonitor patience argument now optional for cleaner code
• Better handling for binary get_stats in confusion matrix based metric
• Added MetricConfusionMatrixBase for adding custom confusion matrix based metrics
• Added ConfusionMatrixBasedMetric Enum to get specific metrics such as tp,fp,fn,tn,precision,sensitivity,specificity,recall,ppv,npv,accuracy,f1score
• Added confusion matrix common metrics (TruePositives, TrueNegatives, FalsePositives, FalseNegatives)
• Added MetricMethod enum to pass to MetricBase, now you can define whether your metric is based on MEAN, SUM, or LAST of all batches
• StatsPrint callback now support "print_confusion_matrix" and "print_confusion_matrix_normalized" arguments in case MetricConfusionMatrixBase metric is found
• Added confusion matrix tests and example
• Some custom layers rename (breaking changes in this part)

Confusion Matrix - phase 2

• CallbackMonitor patience argument now optional for cleaner code
• Better handling for binary get_stats in confusion matrix based metric

Previously on lpd:

• Added MetricConfusionMatrixBase for adding custom confusion matrix based metrics
• Added ConfusionMatrixBasedMetric Enum to get specific metrics such as tp,fp,fn,tn,precision,sensitivity,specificity,recall,ppv,npv,accuracy,f1score
• Added confusion matrix common metrics (TruePositives, TrueNegatives, FalsePositives, FalseNegatives)
• Added MetricMethod enum to pass to MetricBase, now you can define whether your metric is based on MEAN, SUM or LAST of all batches
• StatsPrint callback now support "print_confusion_matrix" and "print_confusion_matrix_normalized" arguments in case MetricConfusionMatrixBase metric is found
• Added confusion matrix tests and example
• Some custom layers renames (breaking changes in this part)

Confusion Matrix

• Added MetricConfusionMatrixBase for adding custom confusion matrix based metrics
• Added ConfusionMatrixBasedMetric Enum to get specific metrics such as tp,fp,fn,tn,precision,sensitivity,specificity,recall,ppv,npv,accuracy,f1score
• Added confusion matrix common metrics (TruePositives, TrueNegatives, FalsePositives, FalseNegatives)
• Added MetricMethod enum to pass to MetricBase, now you can define whether your metric is based on MEAN, SUM or LAST of all batches
• StatsPrint callback now support "print_confusion_matrix" and "print_confusion_matrix_normalized" arguments in case MetricConfusionMatrixBase metric is found
• Added confusion matrix tests and example
• Some custom layers renames (breaking changes in this part)

StatsResult class

• Added StatsResult class
• Trainer.evaluate(...) now returns StatsResult instance with loss and metrics details

Bug fix in CollectOutputs on GPU

• Fixed CollectOutputs callbacks on GPU
• Pipfile explicit versioning

TopKCategoricalAccuracy & bugfixes

• Added metric TopKCategoricalAccuracy
• Added Predictor.from_trainer() method to Predictor class
• Fixed loading predictor from_checkpoint if the checkpoint is not Full Trainer
• Fixed loading Trainer & Predictor on CPU from GPU checkpoint
• Fixed saving/loading if scheduler is None
• Added unittest for TopKCategoricalAccuracy