C/Other Languages: Support serialization and deserialization of Tensors

[asimshankar]

In C++ and Python, Tensors can be serialized into a TensorProto using functions like Tensor::AsProtoTensorContents and make_tensor_proto

This is particularly useful when communicating tensors across process boundaries such as PredictRequest RPCs to a TensorFlow Model Server

Filing this to track any changes to the C API and language bindings.

Interested parties: @asimshankar @jhseu @nfiedel @josh11b

[yifeif]

@asimshankar, saw "Go: API functions to serialize/deserialize Tensors" was added with #6137. Are we expecting further update on this issue?

[asimshankar]

Nobody is actively working on this, this was more of a tracking bug. I think it's fine to close this for now (and re-open if and when the need is more urgent). The "need" would be to save/load tensors directly (as opposed to using things like the Restore operations or a SavedModel loader).

[samsends]

@asimshankar, could we reopen this? Tensor se/de operations in C would be extremely useful for multi-language interoperability.

[asimshankar]

@sbsends : Happy to reopen it, but would love to hear what specifically you're interested in. Could you elaborate on what you're trying to do and the problems being run into?

[samsends]

@asimshankar, we are interested in passing TensorProto's directly between languages. E.g. generating a packed TensorProto in Python and sending them over the wire to a program running in Golang. This is possible using Pythons implementation of make_proto and then building a custom handler in Golang to parse the TensorProto's contents/type/dim's. The tensor contents can be read using ReadTensor, but this only supports numeric data types. IMHO, it would be significantly cleaner (and less error prone) to handle ser/de in a standardized way across all client libraries using the C API.

Edit: grammar fix

[samsends]

@asimshankar, would you consider giving feedback on our stab at building a standard for transporting tensors across applications/languages? The goal is to support many frameworks and act as an intermediary between application specific protocols. What can we do to increase the likelihood of gaining TensorFlow support?

https://github.com/tensortask/ttp

[lyschoening]

@asimshankar Would it we possible to reopen this? In our case, we would like to serialize/deserialize tensors to communicate with Tensorflow Serving from Go. We can not use the Go library as it does not handle serialization of string tensors.

[blepfo]

Seconding the above comment -- I ran into this issue trying to use to Go API tf.ReadTensor() only to discover that DT_STRING is not supported.

In the meantime is there an easy workaround for parsing string TensorProto into Go tf.Tensor?