tensordict 0.11.0

TensorDict is a pytorch dedicated tensor container.

📖 TensorDict

TensorDict is a dictionary-like class that inherits properties from tensors, making it easy to work with collections of tensors in PyTorch. It provides a simple and intuitive way to manipulate and process tensors, allowing you to focus on building and training your models.

Key Features | Examples | Installation | Citation | License

Key Features

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

The key features are:

• 🧮 Composability: TensorDict generalizes torch.Tensor operations to collection of tensors.
• ⚡️ Speed: asynchronous transfer to device, fast node-to-node communication through consolidate, compatible with torch.compile.
• ✂️ Shape operations: Perform tensor-like operations on TensorDict instances, such as indexing, slicing or concatenation.
• 🌐 Distributed / multiprocessed capabilities: Easily distribute TensorDict instances across multiple workers, devices and machines.
• 💾 Serialization and memory-mapping
• λ Functional programming and compatibility with torch.vmap
• 📦 Nesting: Nest TensorDict instances to create hierarchical structures.
• ⏰ Lazy preallocation: Preallocate memory for TensorDict instances without initializing the tensors.
• 📝 Specialized dataclass for torch.Tensor (@tensorclass)

Examples

This section presents a couple of stand-out applications of the library. Check our Getting Started guide for an overview of TensorDict's features!

Fast copy on device

TensorDict optimizes transfers from/to device to make them safe and fast. By default, data transfers will be made asynchronously and synchronizations will be called whenever needed.

# Fast and safe asynchronous copy to 'cuda'
td_cuda = TensorDict(**dict_of_tensor, device="cuda")
# Fast and safe asynchronous copy to 'cpu'
td_cpu = td_cuda.to("cpu")
# Force synchronous copy
td_cpu = td_cuda.to("cpu", non_blocking=False)

Coding an optimizer

For instance, using TensorDict you can code the Adam optimizer as you would for a single torch.Tensor and apply that to a TensorDict input as well. On cuda, these operations will rely on fused kernels, making it very fast to execute:

class Adam:
    def __init__(self, weights: TensorDict, alpha: float=1e-3,
                 beta1: float=0.9, beta2: float=0.999,
                 eps: float = 1e-6):
        # Lock for efficiency
        weights = weights.lock_()
        self.weights = weights
        self.t = 0

        self._mu = weights.data.clone()
        self._sigma = weights.data.mul(0.0)
        self.beta1 = beta1
        self.beta2 = beta2
        self.alpha = alpha
        self.eps = eps

    def step(self):
        self._mu.mul_(self.beta1).add_(self.weights.grad, 1 - self.beta1)
        self._sigma.mul_(self.beta2).add_(self.weights.grad.pow(2), 1 - self.beta2)
        self.t += 1
        mu = self._mu.div_(1-self.beta1**self.t)
        sigma = self._sigma.div_(1 - self.beta2 ** self.t)
        self.weights.data.add_(mu.div_(sigma.sqrt_().add_(self.eps)).mul_(-self.alpha))

Training a model

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.

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 uv + PyTorch nightlies:

If you're using a PyTorch nightly (e.g. installed from the PyTorch nightly wheel index), then for editable installs you should install tensordict with --no-deps.

This avoids uv re-resolving torch from its configured indexes (by default: PyPI), which can otherwise replace an existing nightly with the latest stable torch.

This is an uv-specific behavior; plain pip install -e . typically won’t replace an already-installed PyTorch nightly.

To keep a nightly torch with an editable install:

• install without resolving deps (recommended):

uv pip install -e . --no-deps

• or (less recommended) explicitly point uv at the PyTorch nightly wheel index (CPU shown; use the appropriate backend directory like cu126/):

uv pip install -e . --prerelease=allow -f "https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html"

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.