Fetchable

Fetchables can be executed within a TensorFlow session and have their results returned from Session.run.

For example, the result of any mathematical operation can be fetched.

Currently supported fetchable types are:

• Single Output, OutputIndexedSlices, SparseOutput object.
• Sequences of other Fetchables (e.g., Seqs, Lists, etc.).
  • Sequences that are not homogeneous are not supported (e.g., Seq(output1, Seq(output1, output2))).
  • Note that, for that reason, even though Seq(List(output1), List(output1, output2)) is supported, Seq(Seq(output1), List(output1, output2)) is not.
  • A sequence containing both Outputs and SparseOutputs, for example, is considered heterogeneous. For such cases, it is advisable to use tuples.
• Arrays of other Fetchables.
• Maps with arbitrary key types and Fetchable value types.
• Products of other Fetchables (e.g., tuples).
  • Note that with tuples, heterogeneous types are supported, due to the tuple type being a kind of heterogeneous collection. Internally, the fetchable provided to a session will be de-duplicated to prevent redundant computation. This means that ops that appear more than once in the fetchable, will only be executed once by the session.

Fetchables guarantee that the returned result of a computation will match the structure of the provided fetchable. For example, if a Seq(List(output1), List(output1, output2)) is provided as the Session.run fetchable, then the result will have the following structure Seq(List(tensor1), List(tensor1, tensor2)).