Release 3.1

Version 3.1 (5 apr 2022)

• Add the "split" strategy in the PatchesSelection application. Now the application can select points for 3 independent data splits (train, valid, test).
• PatchesExtraction and PatchesSelection don't use the usenodata option anymore. Now to use nodata, one simply has to set a parameter value for the nodata (no more default value).
• New python examples in python/examples/tensorflow_v2x to illustrate how deterministic models can be quickly built using TensorFlow 2, exporte as SavedModel, and used with OTBTF.
• Change import gdal --> from osgeo import gdal in otbtf.py

Release 3.0 candidate 2

The following bugs were fixed since the v3.0-rc1:

• Exception message when last component size of output tensor shape is -1
• Handling tensors names in alphabetical order (signaturedef.inputs() and signaturedef.outputs() order is completely random!) now throws a warning when there are > 1 available tensor name in the graph. Add a blacklist to avoid confusion between users placeholders and input tensors.
• Fix: input users placeholders actual names are now resolved

Release 3.0 candidate 1

• Use Tensorflow 2 API everywhere. Everything is backward compatible (old models can still be used).
• Support models with no-named inputs and outputs. OTBTF now can resolve the names! :) Just in the same order as they are defined in the computational graph Sources inputs are feeding input tensors coming in their names alphabetical order.
• Support user placeholders of type vector (int, float or bool)
• More unit tests, spell check, better static analysis of C++ and python code
• Improve the handling of 3-dimensional output tensors, + more explanation in error messages about output tensors dimensions.
• Improve PatchesSelection to locate patches centers with corners or pixels centers depending if the patch size is odd or even.

Release 2.5

• fix bug in otbtf.py (see #68 )

Version 2.4

Bug fix

• The output image origin was sometimes shifted from a fraction of pixel. This issue happened only with multi-inputs models that have inputs of different spacing. This bug is now fixed!

Improvement

• The output image largest possible region is now computed on the maximum possible area within the expression field. Before that, the largest possible region was too much cropped when an expression field > 1 was used. Now output images are larger when a non unitary expression field is used.

Version 2.3

Now support the following tensor types:

• tensorflow::DT_FLOAT
• tensorflow::DT_DOUBLE
• tensorflow::DT_UINT64
• tensorflow::DT_INT64
• tensorflow::DT_UINT32
• tensorflow::DT_INT32
• tensorflow::DT_UINT16
• tensorflow::DT_INT16
• tensorflow::DT_UINT8

No more itkGenericExceptionMacro("TF DataType "<< dt << " not currently implemented !") for those 😉

Version 2.2

• Major enhancement of docker image build

r2.1

Introducing OTBTF python classes.

• PatchesReaderBase: base abstract class for patches readers. Users/developers can implement their own from it!
• PatchesImagesReader: a class implementing PatchesReaderBase to access the patches images, as they are produced by the OTBTF PatchesExtraction application.
• IteratorBase: base class to iterate on PatchesReaderBase-derived readers.
• RandomIterator: an iterator implementing IteratorBase designed to randomly access elements.
• Dataset: generic class to build datasets, consisting essentially of the assembly of a PatchesReaderBase-derived reader, and a IteratorBase-derived iterator. The Dataset handles the gathering of the data using a thread. It can be used as a tf.dataset to feed computational graphs.
• DatasetFromPatchesImages: a Dataset that uses a PatchesImagesReader to allow users/developers to stream their patches generated using the OTBTF PatchesExtraction through a tf.dataset which implements a streaming mechanism, enabling low memory footprint and high performance I/O thank to a threaded reading mechanism.

Version 2.0 release candidate 1

Uses TensorFlow 2

Version 1.8

New features:
-PatchesExraction supports no-data (a different value for each source can be set)
-New sampling strategy available in PatchesSelection (balanced strategy)