Support 'bytes' type in torchscript

[pritamdamania87]

🚀 Feature

Need support for 'bytes' type in torchscript

Motivation

While dealing with image data, we have use cases where we download a blob from a blobstore and need to pass it to a torchscript module for further processing (ex: torchvision.transforms).

Currently, there is no clean way to do this since pybind enforces that python strings be utf-8 encoded and as a result we can't pass in arbitrary bytes to a torchscript module.

Supporting the python 'bytes' type in torchscript would be a clean way to support use cases like this.

cc @suo

[pritamdamania87]

@dzhulgakov

[dzhulgakov]

Desired semantics for str and bytes:

• they behave like str and bytes types in Python3 (though encode/decode functions won't be part of script for a while likely)
• in Python3 str and bytes can be converted to the corresponding Python types only
• in Python2, both TorchScript str and bytes can accept python's str type as an input. Since TorchScript is statically typed at this level we can disambiguate which type they should become. When transferred to python land both of the types convert to regular str in py2.
• in Python2, we might consider accepting unicode type as source for str, but it might be cleaner and simpler to just forbid it

[apaszke]

Can't you just store the raw data in a byte tensor? I really think we should be careful about the limitations of what we implement, or we'll end up with a complete reimplementation of Python.

[pritamdamania87]

@apaszke Short term, we are using byte tensor for this. Although, I think the concern in general is if we store arbitrary blobs in bytetensor, its not very clear what individual elements in the tensor represent. For example, if I store a jpg blob in a bytetensor and then reshape it, the new tensor doesn't have much meaning (similarly for multiplying/adding tensors).

[apaszke]

And how does adding a byte type help with this? Sure, you can't resize/multiply them, but why would you do that to a tensor that has binary data in the first place?