Changelog

This page collects all the "official" changes to CN24 in chronological order.

CN24 v2.0.5

• Switch to git-flow
• Preliminary support for jpg/png streaming

CN24 v2.0.4

• Introduced CTest suite to catch errors

CN24 v2.0.3

• Fixed memory corruption issue with AdvancedMaxPoolingLayer and non-square pooling regions

CN24 v2.0.2

• Fixed missing layer time measurement facilities
• classifyImage: Fixed missing spatial prior

CN24 v2.0.1

• Reduced number of compiler warnings significantly
• runBenchmark: Added random content to data for benchmarks and gradient check

CN24 v2.0.0

• Several bug fixes
• Introducing a compressed TensorStream format
• Improved GraphViz network structure output
• Better OpenCL and clBLAS support on Windows
• Complete statistics overhaul improving data analysis significantly
  • Modular architecture allowing any component to report data
  • Added a "set experiment name" command to trainNetwork
  • Improved scripts for experiments
• New layer InputDownSamplingLayer
• New tool tensorStreamStats: Reads a TensorStream and displays information on sample distribution between classes
• New tool makeCompressedTensorStream: New variant of makeTensorStream that uses compressed TensorStream format
• New tool compressTensorStream: Compresses an existing TensorStream
• New tool runBenchmark: Runs a repeatable benchmark for performance assessment

CN24 v1.3.2

• Fixed a bug that prevented loss sampling from working on OpenCL platforms

CN24 v1.3.1

• Improved startup times by memory mapping the dataset on supported systems (Linux,OS X)

CN24 v1.3.0

• Added QuickProp optimization algorithm
• New advanced max-pooling layer
• Performance improvements
• Network is now represented as a graph for increased flexibility
• Debug messages are now suppressed by default
• Several bug fixes
• trainNetwork: New command to output network structure as GraphViz file
• TensorMath: OpenCL code path now uses clBLAS
• ConvolutionLayer: Now supports stride, padding and grouping

CN24 v1.2.2

• Init: Fixed wrong delete operator, restricted OpenCL context to device
• Init: Fixed conditional includes

CN24 v1.2.1

• Enhanced OpenCL kernel search locations
• Made OpenCL platform and device configurable

CN24 v1.2.0

• Added dropout implementation
• Changed weight update formula, now independent of batch size
• Fixed slightly offset localized error function when using patchwise training
• Changed image size compensation method, hyperparameters are now comparable between FCN and patch training
• Fixed a bug that caused the first testing sample to be skipped
• CN24 now displays the current branch and commit SHA1 on startup
• Added support for Intel OpenCL SDK

CN24 v1.1.1

• Fix float parsing bug in ConfigParsing

CN24 v1.1.0

• Added preliminary support for hybrid training

CN24 v1.0.2

• Fixed gradient check for non-default localized error functions

CN24 v1.0.1

• Fixed issue #17: Path to OpenCL kernels invalid if working directory != build directory

CN24 v1.0.0

• This is the first public release of CN24. Enjoy!