An experimental deep learning library written in pure rust. Breakage expected on each release in the short term. See mnist.rs in examples or Rusty_SR for usage samples.
Issues are a great place for discussion, problems, requests, and coordinating future work.
Blatantly incorrect documentation contributions are encouraged as a way to guide efforts on docs, just submit a PR and fill a doc comment with anything from your best guess to passive aggressive nursery rhymes.
Patchy until the library settles down, particularly until the graph abstraction is finalised and the switch to ndarray is completed.
- Computation hypergraph
- Dense Connection and Bias operations
- Loss functions
- Mean Squared Error
- Categorical Cross Entropy
- SoftMax Cross Entropy
- Binary Cross Entropy
- SRGB Curves
- Spatial operations
- Shape constraint propagation
- N-dimensional Convolution
- Arbitrary padding
- N-dimensional AvgPooling
- N-dimensional spaxel shuffling for "Sub-pixel Convolution"
- N-dimensional Linear-Interpolation (backprop not finished)
- Global Pooling
- Data Loading
- Image Folders
- Imagenet (ILSVRC)
- Adaptive BatchSize
- Adaptive Learning Rate
- Adaptive Momentum
- Basic numerical tests
- Limit Optimiser evaluation batch size to stay within memory limits
- Selectively disable calculation of forward values, node derivatives and parameter derivatives
- Builder patterns for operation contruction
- Split Graph struct into mutable GraphBuilder and immutable Sub-Graphs
- Replace 'accidentally quadratic' graph algorithms
- Replace up-front allocation with Sub-Graph optimised allocation/deallocation patterns based on liveness analysis of nodes
- Overhaul data ingestion, particularly buffering input processing/reads.
- Move to bluss' ndarray where possible (long overdue)
- Improve naming inter/intra-library consistancy
- Complete Documentation
- Reduce ability to express illegal states in API
- Move from panics to error-chain
- Guard unsafe code rigourously
- Comprehensive tests
- Arrayfire as an option for sgemm on APUs
- Efficient probablistic structures (e.g. generative RNNs)
- Graph optimisation passes and inplace operations
- Support for both dynamic and static graphs