/gopickle

Go library for loading Python's data serialized with pickle and PyTorch module files.

Primary LanguageGoBSD 2-Clause "Simplified" LicenseBSD-2-Clause

GoPickle

GoPickle is a Go library for loading Python's data serialized with pickle and PyTorch module files.

The pickle sub-package provides the core functionality for loading data serialized with Python pickle module, from a file, string, or byte sequence. All pickle protocols from 0 to 5 are supported.

The pytorch sub-package implements types and functions for loading PyTorch module files. Both the modern zip-compressed format and the legacy non-tar format are supported. Legacy tar-compressed files and TorchScript archives are not supported.

Project Status and Contributions

This project is currently in alpha development stage. While we provide better documentation, tests, and functionalities, above all we'd like to battle-test this library with real data and models. This would tremendously help us find tricky bugs, add more built-in types, or change the library API to make it easier to use.

The simplest and most useful and way to contribute is to try this library yourself and give us your feedback: report bugs, suggest improvements, or just tell us your opinion. And of course, if you feel like it, please go on with your own pull requests! We can discuss any issue and try to find the best solution for it together.

Usage

Pickle

Simple usage:

import "github.com/nlpodyssey/gopickle/pickle"

// ...

// from file
foo, err := pickle.Load("foo.p") 

// from string
stringDump := "I42\n."
bar, err := pickle.Loads(stringDump)

// ...

Advanced/custom usage:

import "github.com/nlpodyssey/gopickle/pickle"

var r io.Reader

// ...

u := pickle.NewUnpickler(r)

// Handle custom classes
u.FindClass = func(module, name string) (interface{}, error) {
    if module == "foo" && name == "Bar" {
        return myFooBarClass, nil
    }
    return nil, fmt.Errorf("class not found :(")
}

// Resolve objects by persistent ID
u.PersistentLoad = func(persistentId interface{}) (interface{}, error) {
    obj := doSomethingWithPersistentId(persistentId)
    return obj, nil
}

// Handle custom pickle extensions
u.GetExtension = func(code int) (interface{}, error) {
    obj := doSomethingToResolveExtension(code)
    return obj, nil
}

// Handle Out-of-band Buffers
// https://docs.python.org/3/library/pickle.html#out-of-band-buffers
u.NextBuffer = func() (interface{}, error) {
    buf := getMyNextBuffer()
    return buf, nil
}

// Low-level function to handle pickle protocol 5 READONLY_BUFFER opcode.
// By default it is completely ignored (sort of no-op); here you have the
// ability to manipulate objects as you need.
u.MakeReadOnly = func(obj interface{}) (interface{}, error) {
    newObj := myReadOnlyTransform(obj)
    return newObj, nil
}

data, err := u.Load()

// ...

PyTorch

The library currently provides a high-level function for loading a module file:

import "github.com/nlpodyssey/gopickle/pytorch"

// ...

myModel, err := pytorch.Load("module.pt")

// ...

More features will be provided in the future.

How it works

Pickle

Unlike more traditional data serialization formats, (such as JSON or YAML), a "pickle" is a program for a so-called unpickling machine, also known as virtual pickle machine, or PM for short. A program consists in a sequence of opcodes which instructs the virtual machine about how to build arbitrarily complex Python objects. You can learn more from Python pickletools module documentation.

Python PM implementation is straightforward, since it can take advantage of the whole environment provided by a running Python interpreter. For this Go implementation we want to keep things simple, for example avoiding dependencies or foreign bindings, yet we want to provide flexibility, and a way for any user to extend basic functionalities of the library.

This Go unpickling machine implementation makes use of a set of types defined in types. This sub-package contains Go types representing classes, instances and common interfaces for some of the most commonly used builtin non-scalar types in Python. We chose to provide only minimal functionalities for each type, for the sole purpose of making them easy to be handled by the machine.

Since Python's pickle can dump and load any object, the aforementioned types are clearly not always sufficient. You can easily handle the loading of any missing class by explicitly providing a FindClass callback to an Unpickler object. The implementation of your custom classes can be as simple or as sophisticated as you need. If a certain class is required but is not found, by default a GenericClass is used. In some circumstances, this is enough to fully load a pickle program, but on other occasions the pickle program might require a certain class with specific traits: in this case, the GenericClass is not enough and an error is returned. You should be able to fix this situation providing a custom class implementation, that jas to reflect the same basic behaviour you can observe in the original Python implementation.

A similar approach is adopted for other peculiar aspects, such as persistent objects loading, extensions handling, and a couple of protocol-5 opcodes: whenever necessary, you can implement custom behaviours providing one or more callback functions.

Once resolved, all representation of classes and objects are casted to interface{} type; then the machine looks for specific types or interfaces to be implemented on an object only where strictly necessary.

The virtual machine closely follows the original implementation from Python 3.8 - see the Unpickler class.

PyTorch

PyTorch machine learning framework allows you to save and load Python objects which include (or are) Tensors or other framework-specific elements. Tensors data are handled by the more primitive Storage classes, which are efficiently serialized as raw sequences of bytes. All the rest is dumped using pickle. When serializing, the programmer can choose any available pickle protocol, and whether to use zip compression.

The package pytorch implements loading functionalities for data files serialized with PyTorch (called modules). The Go implementation strictly follows the original Python (and C++) code. The pickle and types packages are used to read some parts of a given file. Other specific types are implemented in the pytorch module itself, most notably to reflect the content of PyTorch Tensor and Storage objects.

License

GoPickle is licensed under a BSD-style license. See LICENSE for the full license text.