`defun` uses dtypes and shapes instead of `Tensors` as cache keys. Dtypes are used because TensorFlow graphs are not parametric w.r.t. dtypes. Shapes are used for both performance reasons, as much TensorFlow code specializes on known shapes to produce slimmer graphs, and correctness, as some high-level APIs require shapes to be fully-known. `include_tensor_ranks_only` allows caching on arguments
source not stored for this graph (policy: none)
nothing calls this directly
no test coverage detected