SINGA is a distributed deep learning library.
This release includes following changes:
- Add one cifar-10 distributed CNN example for benchmarking the performance of the distributed training.
- Add one large CNN example for training with a dataset from the filesysetm.
Enhance distributed training
- Improve the data augmentation module for faster distributed training.
- Add device synchronization for more accurate time measurements during the distributed training.
Add Support for half-precision floating-point format (fp16) in deep learning models and computational kernels.
Update new onnx APIs and fix onnx examples accordingly, namely, DenseNet121, ShuffleNetv1, ShuffleNetv2, SqueezeNet, VGG19.
Add a new method to resize images by given width and height.
Use docusaurus versioning to simplify the process of generating the project homepage.
Promote code quality
- Unify the formats of docstrings that describe the contents and usage of the module.
- Unify the parameters of command-line arguments.
- Fix the CI build error by downloading the tbb binaries.
- Add disabling graph option for accessing parameter or gradient tensors during distributed
- Solve the warnings of deprecated functions in the distributed optimizer module.