Changelog

Version 2.0

tag v2.0

Pico-CNN is now implemented in C++
- The set of implemented operations and test coverage is equivalent to v1.0.2.
- Data and its shape is now encapsulated in the pico_cnn::naive::Tensor class making the interface of all operations more similar and less cluttered. Moreover operations like Reshape or Flatten are now working more straightforward.
- Right now the pico_cnn::naive::Tensor class supports up to four shape dimensions.

Version 1.0.2

tag v1.0.2

Bugfixes
- Fixed bug in AveragePooling: When using asymmetric padding with count_include_pad == 0 the center/edge case computation assumed symmetric padding.
- Fixed typo in activation function test.
- Fixed bug using 'height' twice when generating random 4D input reference data.
Additional tests
- Added three more tests for the convolution layer covering multiple output channels.
Miscellaneous
- Added optimization flags to the library's and the generated Makefile.
- Added seeding option for the generation of reference data when using random input.
- Removed targets from CMakeLists.txt which have to be generated first (an example target for LeNet can be found commented). CMake should only be used for debugging. For all other uses please use the (generated) Makefiles.

Version 1.0.1

tag v1.0.1

Fixed typo in GlobalMaxPooling

Version 1.0

tag v1.0

Restructured C code (split in .c and .h)
Import of ONNX models and generation of C code
Tested neural networks:
- LeNet
- MNIST Multi-Layer-Perceptron (MLP)
- MNIST Perceptron
- AlexNet
- VGG-16
- VGG-19
- MobileNet-V2
- Inception-V3
- Inception-Resnet-V2
- TC-ResNet-8
Removed experimental hand-crafted ARM-NEON Code
Removed experimental OpenMP support

Version 0.1

tag v0.1

Tested neural networks:
- LeNet-5
- AlexNet
- VGG-16
Experimental support of ARM-NEON SIMD instructions for several layers
Experimental support of OpenMP on layer level