Skip to content

Latest commit

 

History

History
155 lines (129 loc) · 7.19 KB

KNOWN_ISSUES.md

File metadata and controls

155 lines (129 loc) · 7.19 KB
Raw

Swift for TensorFlow Known Issues

This is a curated list of Swift for TensorFlow known issues and missing features. With every release, new issues are added and resolved issues are updated.

Legend:

  • Notable issues are marked in bold.
  • Issues under active development are marked with 🚧.

Please see the JIRA issue tracker for a full list of known issues.

Version 0.4

Automatic Differentiation

  • Marking stored properties as @differentiable may result in a crash. (TF-521)
  • Marking an intializer as @differentiable may result in a SIL verification error: operand of 'apply' doesn't match function input type. (TF-526)
  • Defining custom derivatives for @_alwaysEmitIntoClient declarations results in a crash. (TF-545)
  • Differentiating struct initializers may result in a SIL verification error: substituted callee type does not match substitutions. (TF-546)

Version 0.3

TensorFlow Library

  • Tensor advanced indexing and striding are not supported on GPU.

Version 0.2

Notebook Environments (Colab and Jupyter)

  • When a runtime error occurs or when you interrupt cell execution, resources (e.g. host memory, GPU memory) do not get released. This can lead to OOMs. (TF-338)
    • Workaround: Restart the runtime (Runtime > Restart Runtime in the Colab menu bar) to release all the resources.
  • If the last statement on a cell evaluates to a struct that was defined in the notebook, then you get an error ("use of undeclared type") instead of seeing the value of the statement. (TF-125)
    • Workaround: Wrap the last statement in print().
  • Using extensions to conform a type to a protocol (e.g. extension MyType: MyProtocol { ... }), often causes duplicate conformance errors. (TF-162)
    • Workaround: Add the conformance in the same cell where the type is defined.
  • If a cell that declares a type executes twice, then it creates two different types with the same name. Mixing these types can lead to confusing error messages like cannot convert value of type 'MyType' to expected argument type 'MyType'. (TF-156)
    • Workaround: Re-run all cells that use the declared type, so that they use the new type.
  • The autocomplete UI should show types and documentation. It should position your cursor in the first argument when you complete a function call.

Swift Standard Library Enhancements

  • The Differentiable protocol's allDifferentiableVariables requirement should not have a setter. Do not use this directly through a generic type with a Differentiable conformance constraint. (TF-208)

TensorFlow Library

  • 🚧 Model checkpointing and serialization APIs are missing. (TF-388)
  • TensorFlow runtime errors (e.g. shape mismatch errors) do not show useful source location information or useful stack traces. (TF-458)

TensorFlow Runtime

  • If you are using a CUDA build and you have an NVIDIA GPU with a compute capability other than 3.5 or 7.0, then you will experience a ~10 minute delay the first time you execute a TensorFlow operation, while TensorFlow compiles kernels for your GPU's compute capability. The program will not print anything out and it will appear to be frozen. (TF-461)
  • There is a long latency when allocating a Tensor for the first time on GPUs. Subsequent allocations are normal speed. (TF-460)
  • TensorFlow runtime errors (e.g. shape mismatch errors) do not show useful source location information or useful stack traces. (TF-458)

Swift for TensorFlow Deep Learning Library

  • Many Keras layers remain to be implemented, help wanted! (swift-apis#54)
  • Parameter sharing APIs (e.g. using the same Tensor weights in multiple layers) are missing.
  • The Parameter class does not yet conform to Differentiable, and is not recommended for general use.
  • The compiler errors displayed when a user-defined layer struct fails to fully satisfy the requirements of the Layer protocol are unclear.

Automatic Differentiation

  • 🚧 Differentiation does not yet support functions with control flow. ([TF-354])
  • 🚧 Higher-order differentiation is not yet supported.
  • Differentiating functions with respect to an inout parameter is not yet supported. (TF-357)
  • The compiler will only synthesize conformance requirements for Differentiable in struct types. (TF-37)
  • The @differentiable attribute incorrectly passes type-checking in some cases, when an error should be produced. This leads to compiler crashes. (TF-449)
  • The @differentiating attribute leads to a compiler crash when the derivative function is defined in a generic context that is more constrained than the original function's generic context. (TF-358)
    • Resolved (0.3). The @differentiating attribute can register derivatives with a generic context that is more constrained than the original function's generic context.
  • Referring to a @differentiable function using key paths leads to a compiler crash. (TF-123)

Python Interoperability

  • Python runtime errors do not show useful source location information. The Python call stack should be shown. (TF-150)
  • When an argument to a Python function cannot be converted to a Python object, the compiler wrongly claims that the function is of non-function type rather than pointing out that the argument doesn't conform to PythonConvertible. (TF-220)
  • Python TensorFlow cannot be imported because of various issues (binary incompatibility, symbol conflicts).