Changelog

Version 0.3.1

Fixes:

Fixed issue #2 "Error : expected an indented block after 'except' statement on line 240"
Updated version numbers and requirements_cu11.txt

Version 0.3.0

Improvements and bugfixes:

Fixed a bug in PyTorch-model loading
Improved loading and handling of TensorFlow models
Added support for Metal Performance Shaders (MPS)
Added requirement file for virtual python environment on mac os

Version 0.2.2

Fixed incorrect decoding of 1-dimensional tensors of size 1

Version 0.2.1

Fix:

Removed the import of an unused python-package which caused errors

Version 0.2.0

Bug fixes and improvements:

Fixed an issue, which caused a wrong return value when specifying return_decompressed_model in decompress_model
compress_model now also accepts arrays of type uint8, uint16, int8, int16 and float16 (but internally converts them into int32 or float32)
improved guessing of block_id and parameter_types for TensorFlow and PyTorch
NNCodec now accepts to only specify a subset of the tensors for block_id_and_param_type

Version 0.1.8

Integration of minor fixes:

The file 'imagenet_validation_files.txt' in framework/applications/datasets has not been copied during the installation process
The folder 'tuning' is not required to be present when using ImageNet. The data used for fine tuning is specifed in 'imagenet_validation_files.txt'

Version 0.1.7

Aligned requirements.txt with requirements_cu11.txt and updated handling of tensorflow state dictionaries such that it compatible with the current version of the "h5py" package.

Version 0.1.6

Bugfix for compression of arbitrary numpy dictionaries with 'compress':

The encoder crashed during encoding of arbitrary numpy dictionaries with the 'compress' function. This was caused by missing entries for the parameter types in the model_information dictionary.

Version 0.1.5

Fixed setting a qp per tensor using the encoder parameter qp_per_tensor. The encoder crashed when using this parameter.

Version 0.1.4

Updated the python requirement files. Now, instead of specifiying specific package versions, only minimum requirements are defined. This simplifies integration of NNCodec in existing python projects.