0.8.9

This release:

• improves the functions for optimal batch size computation;
• allows for using dictionary for model outputs;
• adds a few new Monitor types;
• fixes Pipeline queue and allows for passing executors to the run;

Full Changelog: 0.8.8...0.8.9

0.8.8

This PR:

• improves Normalizer class;
• adds a way to get detailed activations of a TorchModel;
• adds a tutorial on segmentation with batchflow, as well as ADE20k dataset;
• adds delayed imports mechanism and applies it to rarely-used libraries or ones with optional dependencies;

Full Changelog: 0.8.7...0.8.8

0.8.7

Update package versions

0.8.6

Added Normalizer and Quantizer classes for convenience

0.8.5

Minor fixes

0.8.4

Fix both versions to be the same

0.8.3

Fix model decay behaviour

0.8.2a

Lots of incremental improvements. Main features are:

• fix memory leak in Pipeline.run with prefetch turned on;
• make Profiler a lot faster, which allows to turn it on by default
• make TorchModel work with outputs on the fly
• fixed release action

0.8.1

Fix small bug relating padding in ResBlock

0.8.0

This release fixes crop behavior of TorchModel, as well as adds new blocks and methods:

• InternBlock with deformable convolutions
• separate BottleneckBlock that extends the functionality of ResBlock
• method for getting a reference to the current TorchModel instance inside train/predict contexts
• mode parameter for train and predict methods to control nn.Module behavior.

Also, this is the first version after numpy deprecation of autocast to dtype=object of mishaped arrays, so this is fixed in some places.