*   Modify custom export strategy to account for multidimensional sparse
        float splits.
    *   `Conv2D`, `Conv2DBackpropInput`, `Conv2DBackpropFilter` now supports
        arbitrary dilations with GPU and cuDNNv6 support.
    *   `Estimator` now supports `Dataset`: `input_fn` can return a `Dataset`
        instead of `Tensor`s.
    *   Add `RevBlock`, a memory-efficient implementation of reversible residual

  std::vector<TapeTensor> output_tensor_info;
  std::vector<int64_t> input_tensor_id;

  // TODO(apassos) consider narrowing down this interface.
  BackwardFunction* backward_function;

  // Should be called before deleting the backward function. TODO(apassos) use
  // unique_ptrs to ensure this happens.
  std::function<void(BackwardFunction*)> backward_function_deleter;
};

// * size_t is used to represent byte sizes of objects that are
//   materialized in the address space of the calling process.
// * int is used as an index into arrays.
// * Deletion functions are safe to call on nullptr.
//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//