tensordict-nightly 2023.3.1

No project description provided

TensorDict

Installation | General features | Tensor-like features | TensorDict for functional programming using FuncTorch | Lazy preallocation | Nesting TensorDicts | TensorClass

TensorDict is a dictionary-like class that inherits properties from tensors, such as indexing, shape operations, casting to device etc.

The main purpose of TensorDict is to make code-bases more readable and modular by abstracting away tailored operations:

for i, tensordict in enumerate(dataset):
    # the model reads and writes tensordicts
    tensordict = model(tensordict)
    loss = loss_module(tensordict)
    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.

Installation

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

Features

General

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

>>> from tensordict import TensorDict
>>> import torch
>>> tensordict = 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:

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

Tensor-like features

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

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

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

TensorDict for functional programming using FuncTorch

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
>>> from tensordict.nn import make_functional
>>> import torch
>>> from functorch import vmap
>>> layer1 = nn.Linear(3, 4)
>>> layer2 = nn.Linear(4, 4)
>>> model = nn.Sequential(layer1, layer2)
>>> # we represent the weights hierarchically
>>> weights1 = TensorDict(layer1.state_dict(), []).unflatten_keys(".")
>>> weights2 = TensorDict(layer2.state_dict(), []).unflatten_keys(".")
>>> params = make_functional(model)
>>> assert (params == TensorDict({"0": weights1, "1": weights2}, [])).all()
>>> # Let's use our functional module
>>> x = torch.randn(10, 3)
>>> out = model(x, params=params)  # params is the last arg (or kwarg)
>>> # 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
>>> y = vmap(model, (None, 0))(x, params_stack)
>>> print(y.shape)
torch.Size([2, 10, 4])

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

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:

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

However, you could also choose to preallocate the tensordict:

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

>>> tensordict = 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 = tensordict.flatten_keys(separator=".")

Accessing nested tensordicts can be achieved with a single index:

>>> sub_value = tensordict["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 many more features.

Disclaimer

TensorDict is at the alpha-stage, meaning that there may be bc-breaking changes introduced at any moment without warranty. Hopefully that should not happen too often, as the current roadmap mostly involves adding new features and building compatibility with the broader pytorch ecosystem.

License

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