0b01 / tensorscript

shapechecking neural net DSL using Hindley-Milner type system(compiles to pytorch as proof of concept)

TensorScript

Dependently-typed tensor computation.

Features

• Parametric polymorphism
• Compile time type checking
• Dependently typed tensors
• Multiple targets(Tensorflow, PyTorch, more to come!)
• Pipes operator

Pipes operator

Pipes operator is a syntax sugar for chained function calls inspired by F#, Elixir and R. For example,

x |> lin1 |> leaky_relu(p=0.2) |> sigmoid

compiles to

x = lin1(x)
x = leaky_relu(x, p=0.2)
x = sigmoid(x)

Development

The language is not usable in production or development.

Todo

1. implement module pattern matching
2. type level computation (resolved tensor dimension)
3. BUG: dimension mismatch for mnist example need to create fresh type variables for different static forward functions
4. BUG: non-determinism
5. BUG: impl Hash, Eq for Type
6. set up examples and tests
7. set up commandline
8. more examples
9. better errors in parser
10. code gen: PyTorch
11. add more examples
12. lift dim and tsr to top level
13. add dim level computation dim1 * dim1
14. use Linear as L; aliasing
15. add binary ops (+, -, *, /, %)
16. add if else expression
17. add let binding
18. add more tests