README.md

TensorDict

Installation | General features | Tensor-like features | Distributed capabilities | TensorDict for functional programming | **TensorDict for parameter serialization | Lazy preallocation | Nesting TensorDicts | TensorClass

TensorDict is a dictionary-like class that inherits properties from tensors, such as indexing, shape operations, casting to device or point-to-point communication in distributed settings. Whenever you need to execute an operation over a batch of tensors, TensorDict is there to help you.

The primary goal of TensorDict is to make your code-bases more readable, compact, and modular. It abstracts away tailored operations, making your code less error-prone as it takes care of dispatching the operation on the leaves for you.

Using tensordict primitives, most supervised training loops can be rewritten in a generic way:

for i, data in enumerate(dataset):
    # the model reads and writes tensordicts
    data = model(data)
    loss = loss_module(data)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

With this level of abstraction, one can recycle a training loop for highly heterogeneous task. Each individual step of the training loop (data collection and transform, model prediction, loss computation etc.) can be tailored to the use case at hand without impacting the others. For instance, the above example can be easily used across classification and segmentation tasks, among many others.

Features

General principles

Unlike other pytrees, TensorDict carries metadata that make it easy to query the state of the container. The main metadata are the batch_size (also referred as shape), the device, the shared status (is_memmap or is_shared), the dimension names and the lock status.

A tensordict is primarily defined by its batch_size (or shape) and its key-value pairs:

>>> from tensordict import TensorDict
>>> import torch
>>> data = TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": torch.zeros(3, 4, 5, dtype=torch.bool),
... }, batch_size=[3, 4])

The batch_size and the first dimensions of each of the tensors must be compliant. The tensors can be of any dtype and device.

Optionally, one can restrict a tensordict to live on a dedicated device, which will send each tensor that is written there:

>>> data = TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": torch.zeros(3, 4, 5, dtype=torch.bool),
... }, batch_size=[3, 4], device="cuda:0")

When a tensordict has a device, all write operations will cast the tensor to the TensorDict device:

>>> data["key 3"] = torch.randn(3, 4, device="cpu")
>>> assert data["key 3"].device is torch.device("cuda:0")

Once the device is set, it can be cleared with the clear_device_ method.

TensorDict as a specialized dictionary

TensorDict possesses all the basic features of a dictionary such as clear, copy, fromkeys, get, items, keys, pop, popitem, setdefault, update and values.

But that is not all, you can also store nested values in a tensordict:

>>> data["nested", "key"] = torch.zeros(3, 4) # the batch-size must match

and any nested tuple structure will be unravelled to make it easy to read code and write ops programmatically:

>>> data["nested", ("supernested", ("key",))] = torch.zeros(3, 4) # the batch-size must match
>>> assert (data["nested", "supernested", "key"] == 0).all()
>>> assert (("nested",), "supernested", (("key",),)) in data.keys(include_nested=True)  # this works too!

You can also store non-tensor data in tensordicts:

>>> data = TensorDict({"a-tensor": torch.randn(1, 2)}, batch_size=[1, 2])
>>> data["non-tensor"] = "a string!"
>>> assert data["non-tensor"] == "a string!"

Tensor-like features

[Nightly feature] TensorDict supports many common point-wise arithmetic operations such as == or +, += and similar (provided that the underlying tensors support the said operation):

>>> td = TensorDict.fromkeys(["a", "b", "c"], 0)
>>> td += 1
>>> assert (td==1).all()

TensorDict objects can be indexed exactly like tensors. The resulting of indexing a TensorDict is another TensorDict containing tensors indexed along the required dimension:

>>> data = TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": torch.zeros(3, 4, 5, dtype=torch.bool),
... }, batch_size=[3, 4])
>>> sub_tensordict = data[..., :2]
>>> assert sub_tensordict.shape == torch.Size([3, 2])
>>> assert sub_tensordict["key 1"].shape == torch.Size([3, 2, 5])

Similarly, one can build tensordicts by stacking or concatenating single tensordicts:

>>> tensordicts = [TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": torch.zeros(3, 4, 5, dtype=torch.bool),
... }, batch_size=[3, 4]) for _ in range(2)]
>>> stack_tensordict = torch.stack(tensordicts, 1)
>>> assert stack_tensordict.shape == torch.Size([3, 2, 4])
>>> assert stack_tensordict["key 1"].shape == torch.Size([3, 2, 4, 5])
>>> cat_tensordict = torch.cat(tensordicts, 0)
>>> assert cat_tensordict.shape == torch.Size([6, 4])
>>> assert cat_tensordict["key 1"].shape == torch.Size([6, 4, 5])

TensorDict instances can also be reshaped, viewed, squeezed and unsqueezed:

>>> data = TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": torch.zeros(3, 4, 5, dtype=torch.bool),
... }, batch_size=[3, 4])
>>> print(data.view(-1))
torch.Size([12])
>>> print(data.reshape(-1))
torch.Size([12])
>>> print(data.unsqueeze(-1))
torch.Size([3, 4, 1])

One can also send tensordict from device to device, place them in shared memory, clone them, update them in-place or not, split them, unbind them, expand them etc.

If a functionality is missing, it is easy to call it using apply() or apply_():

tensordict_uniform = data.apply(lambda tensor: tensor.uniform_())

apply() can also be great to filter a tensordict, for instance:

data = TensorDict({"a": torch.tensor(1.0, dtype=torch.float), "b": torch.tensor(1, dtype=torch.int64)}, [])
data_float = data.apply(lambda x: x if x.dtype == torch.float else None) # contains only the "a" key
assert "b" not in data_float

Distributed capabilities

Complex data structures can be cumbersome to synchronize in distributed settings. tensordict solves that problem with synchronous and asynchronous helper methods such as recv, irecv, send and isend that behave like their torch.distributed counterparts:

>>> # on all workers
>>> data = TensorDict({"a": torch.zeros(()), ("b", "c"): torch.ones(())}, [])
>>> # on worker 1
>>> data.isend(dst=0)
>>> # on worker 0
>>> data.irecv(src=1)

When nodes share a common scratch space, the MemmapTensor backend can be used to seamlessly send, receive and read a huge amount of data.

TensorDict for functional programming

We also provide an API to use TensorDict in conjunction with FuncTorch. For instance, TensorDict makes it easy to concatenate model weights to do model ensembling:

>>> from torch import nn
>>> from tensordict import TensorDict
>>> import torch
>>> from torch import vmap
>>> layer1 = nn.Linear(3, 4)
>>> layer2 = nn.Linear(4, 4)
>>> model = nn.Sequential(layer1, layer2)
>>> params = TensorDict.from_module(model)
>>> # we represent the weights hierarchically
>>> weights1 = TensorDict(layer1.state_dict(), []).unflatten_keys(".")
>>> weights2 = TensorDict(layer2.state_dict(), []).unflatten_keys(".")
>>> assert (params == TensorDict({"0": weights1, "1": weights2}, [])).all()
>>> # Let's use our functional module
>>> x = torch.randn(10, 3)
>>> with params.to_module(model):
...     out = model(x)
>>> # an ensemble of models: we stack params along the first dimension...
>>> params_stack = torch.stack([params, params], 0)
>>> # ... and use it as an input we'd like to pass through the model
>>> def func(x, params):
...     with params.to_module(model):
...         return model(x)
>>> y = vmap(func, (None, 0))(x, params_stack)
>>> print(y.shape)
torch.Size([2, 10, 4])

Moreover, tensordict modules are compatible with torch.fx and (soon) torch.compile, which means that you can get the best of both worlds: a codebase that is both readable and future-proof as well as efficient and portable!

TensorDict for parameter serialization and building datasets

TensorDict offers an API for parameter serialization that can be >3x faster than regular calls to torch.save(state_dict). Moreover, because tensors will be saved independently on disk, you can deserialize your checkpoint on an arbitrary slice of the model.

>>> model = nn.Sequential(nn.Linear(3, 4), nn.Linear(4, 3))
>>> params = TensorDict.from_module(model)
>>> params.memmap("/path/to/saved/folder/", num_threads=16)  # adjust num_threads for speed
>>> # load params
>>> params = TensorDict.load_memmap("/path/to/saved/folder/", num_threads=16)
>>> params.to_module(model)  # load onto model
>>> params["0"].to_module(model[0])  # load on a slice of the model
>>> # in the latter case we could also have loaded only the slice we needed
>>> params0 = TensorDict.load_memmap("/path/to/saved/folder/0", num_threads=16)
>>> params0.to_module(model[0])  # load on a slice of the model

The same functionality can be used to access data in a dataset stored on disk. Soring a single contiguous tensor on disk accessed through the tensordict.MemoryMappedTensor primitive and reading slices of it is not only much faster than loading single files one at a time but it's also easier and safer (because there is no pickling or third-party library involved):

# allocate memory of the dataset on disk
data = TensorDict({
    "images": torch.zeros((128, 128, 3), dtype=torch.uint8),
    "labels": torch.zeros((), dtype=torch.int)}, batch_size=[])
data = data.expand(1000000)
data = data.memmap_like("/path/to/dataset")
# ==> Fill your dataset here
# Let's get 3 items of our dataset:
data[torch.tensor([1, 10000, 500000])]  # This is much faster than loading the 3 images independently

Preprocessing with TensorDict.map

Preprocessing huge contiguous (or not!) datasets can be done via TensorDict.map which will dispatch a task to various workers:

import torch
from tensordict import TensorDict, MemoryMappedTensor
import tempfile

def process_data(data):
    images = data.get("images").flip(-2).clone()
    labels = data.get("labels") // 10
    # we update the td inplace
    data.set_("images", images)  # flip image
    data.set_("labels", labels)  # cluster labels

if __name__ == "__main__":
    # create data_preproc here
    data_preproc = data.map(process_data, num_workers=4, chunksize=0, pbar=True)  # process 1 images at a time

Lazy preallocation

Pre-allocating tensors can be cumbersome and hard to scale if the list of preallocated items varies according to the script configuration. TensorDict solves this in an elegant way. Assume you are working with a function foo() -> TensorDict, e.g.

def foo():
    data = TensorDict({}, batch_size=[])
    data["a"] = torch.randn(3)
    data["b"] = TensorDict({"c": torch.zeros(2)}, batch_size=[])
    return data

and you would like to call this function repeatedly. You could do this in two ways. The first would simply be to stack the calls to the function:

data = torch.stack([foo() for _ in range(N)])

However, you could also choose to preallocate the tensordict:

data = TensorDict({}, batch_size=[N])
for i in range(N):
    data[i] = foo()

which also results in a tensordict (when N = 10)

TensorDict(
    fields={
        a: Tensor(torch.Size([10, 3]), dtype=torch.float32),
        b: TensorDict(
            fields={
                c: Tensor(torch.Size([10, 2]), dtype=torch.float32)},
            batch_size=torch.Size([10]),
            device=None,
            is_shared=False)},
    batch_size=torch.Size([10]),
    device=None,
    is_shared=False)

When i==0, your empty tensordict will automatically be populated with empty tensors of batch-size N. After that, updates will be written in-place. Note that this would also work with a shuffled series of indices (pre-allocation does not require you to go through the tensordict in an ordered fashion).

Nesting TensorDicts

It is possible to nest tensordict. The only requirement is that the sub-tensordict should be indexable under the parent tensordict, i.e. its batch size should match (but could be longer than) the parent batch size.

We can switch easily between hierarchical and flat representations. For instance, the following code will result in a single-level tensordict with keys "key 1" and "key 2.sub-key":

>>> data = TensorDict({
...     "key 1": torch.ones(3, 4, 5),
...     "key 2": TensorDict({"sub-key": torch.randn(3, 4, 5, 6)}, batch_size=[3, 4, 5])
... }, batch_size=[3, 4])
>>> tensordict_flatten = data.flatten_keys(separator=".")

Accessing nested tensordicts can be achieved with a single index:

>>> sub_value = data["key 2", "sub-key"]

TensorClass

Content flexibility comes at the cost of predictability. In some cases, developers may be looking for data structure with a more explicit behavior. tensordict provides a dataclass-like decorator that allows for the creation of custom dataclasses that support the tensordict operations:

>>> from tensordict.prototype import tensorclass
>>> import torch
>>>
>>> @tensorclass
... class MyData:
...    image: torch.Tensor
...    mask: torch.Tensor
...    label: torch.Tensor
...
...    def mask_image(self):
...        return self.image[self.mask.expand_as(self.image)].view(*self.batch_size, -1)
...
...    def select_label(self, label):
...        return self[self.label == label]
...
>>> images = torch.randn(100, 3, 64, 64)
>>> label = torch.randint(10, (100,))
>>> mask = torch.zeros(1, 64, 64, dtype=torch.bool).bernoulli_().expand(100, 1, 64, 64)
>>>
>>> data = MyData(images, mask, label=label, batch_size=[100])
>>>
>>> print(data.select_label(1))
MyData(
    image=Tensor(torch.Size([11, 3, 64, 64]), dtype=torch.float32),
    label=Tensor(torch.Size([11]), dtype=torch.int64),
    mask=Tensor(torch.Size([11, 1, 64, 64]), dtype=torch.bool),
    batch_size=torch.Size([11]),
    device=None,
    is_shared=False)
>>> print(data.mask_image().shape)
torch.Size([100, 6117])
>>> print(data.reshape(10, 10))
MyData(
    image=Tensor(torch.Size([10, 10, 3, 64, 64]), dtype=torch.float32),
    label=Tensor(torch.Size([10, 10]), dtype=torch.int64),
    mask=Tensor(torch.Size([10, 10, 1, 64, 64]), dtype=torch.bool),
    batch_size=torch.Size([10, 10]),
    device=None,
    is_shared=False)

As this example shows, one can write a specific data structures with dedicated methods while still enjoying the TensorDict artifacts such as shape operations (e.g. reshape or permutations), data manipulation (indexing, cat and stack) or calling arbitrary functions through the apply method (and many more).

Tensorclasses support nesting and, in fact, all the TensorDict features.

Installation

With Pip:

To install the latest stable version of tensordict, simply run

pip install tensordict

This will work with Python 3.7 and upward as well as PyTorch 1.12 and upward.

To enjoy the latest features, one can use

pip install tensordict-nightly

With Conda:

Install tensordict from conda-forge channel.

conda install -c conda-forge tensordict

Citation

If you're using TensorDict, please refer to this BibTeX entry to cite this work:

@misc{bou2023torchrl,
      title={TorchRL: A data-driven decision-making library for PyTorch}, 
      author={Albert Bou and Matteo Bettini and Sebastian Dittert and Vikash Kumar and Shagun Sodhani and Xiaomeng Yang and Gianni De Fabritiis and Vincent Moens},
      year={2023},
      eprint={2306.00577},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Disclaimer

TensorDict is at the beta-stage, meaning that there may be bc-breaking changes introduced, but they should come with a warranty. Hopefully these should not happen too often, as the current roadmap mostly involves adding new features and building compatibility with the broader PyTorch ecosystem.

License

TensorDict is licensed under the MIT License. See LICENSE for details.