v1.4

• Added support for slicing multi-channel input data (from numpy or HDF5)

v1.3

• Added new HDF5 data loader:
  • Supports files containing either single or multiple datasets. Each dataset becomes an input (or target) sample

v1.2

• Added support for half-precision (16-bit) floating-point format in output HDF5 datasets.
• Added capability to asynchronously write inference results to an HDF5 file:
  • It can be enabled by setting test_output_path to the directory where the output should be written. Inference results are saved as independent HDF5 datasets, each named after its corresponding input sample name.
  • This feature subsumes using the debug flag to write out inference results:
• Improvement: label discovery failures (due to unmatched class labels) now error out early.
  • Fixes issue where test sample with unmatched label would only cause error during test cycle (i.e., after a full epoch of training).

v1.1

(Released May 9, 2022)

Dependencies

• Upgraded CUDA from 10.2 to to 11.6, which adds support to NVIDIA A-series GPU architecture
• Upgraded cuDNN from 7.6 to 8.4
• Upgraded libtorch library from 1.7 to 1.11 (requires at least CUDA 11.3)

Features and capabilities

• Added support for depthwise separable convolution operations in one and two dimensions (SeparableConv1D/2D in Keras)

v1.0

• Initial release of rmldnn