TensorFlake

TensorFlake is my own machine learning framework. This aims to be able to train several networks on the CPU fast.

Features

• Automatic differentiation
• Foundational NDArray operations
• Foundational functions for deep neural networks

TODO

• smallvec
• backward without create_graph
• Efficient Linear -> matmul_add
• Save & load
  • param_bin.rs
  • serde
    • Restore optimizers
    • Restore Fn (MLP::activation, etc...)
• Strong typing -> Functional API (tensorflake::function::chain)
  • Generic for dimension
• Multi thread
  • High level API
  • Synchronous update
  • Asynchronous update
  • Specify the number of thread
• Impl ops
• Lazy execution for optimization on the graph
• Regularization
• Tensordot -> ndarray_einsum_beta
• Transposed convolution
• Batch normalization
• Embedding
• Sequential
• Param creator -> Initializer
• Benchmarks
• Measure the execution time of functions and export it as dot file
• Tensor summarization
• Sparse tensor
• Examples
  • CNN
  • RNN
  • GAN
  • VAE
• wasm
• safe rust
• Compare performance against GPU Tensorflow
• Fitting report
• Generic ndarray
• functions
  • reduce_sum
  • signum
• Optimize consecutive element-wise operations

Benchmark

$ cargo +nightly bench -q > benches/result.txt

Author

• carrotflakes ([email protected])

Copyright

Copyright (c) 2022 carrotflakes ([email protected])

License

Licensed under the MIT License.