Getting Started

The path

Follow the pages in order. Each one builds on the previous.

1. Tensors — create, inspect, index, and do math with tensors.

2. Datasets and DataLoaders — organize data and stream it in batches.

3. Transforms — preprocess and augment data before it reaches a model.

4. Models — build a neural network with nn.Module.

5. Autograd — make tensors learnable and get gradients automatically.

6. Training — the loss + optimizer + loop that makes the model learn.

7. Save and Load — keep the trained weights and checkpoints.

After those seven lessons you can train, evaluate, and persist a real model.

What TensorPlay is

TensorPlay is a tensor library for deep learning. It gives you three things you will use on every line of code:

• Tensors — multi-dimensional arrays that carry a device, a data type, and support efficient vectorized math (tensorplay.Tensor).

• Automatic differentiation — every operation records itself into a graph, so for any scalar loss you can call loss.backward() and get the gradient of that loss with respect to every tensor that was involved (tensorplay.autograd).

• A neural-network toolkit — layers and models expressed as tensorplay.nn.Module, optimizers such as SGD and Adam (tensorplay.optim), vision transforms (tensorplay.vision.transforms), and data helpers such as DataLoader and TensorDataset (tensorplay.utils.data).

A defining design choice: the engine is explicit and readable. The graph that backward() walks is real code you can step through, which makes TensorPlay a good place to learn how the machinery works, not just to use it.

Install and check it works

You can install the latest release, or build from source. See the README for the full install options including GPU variants. Once installed, run this smoke test:

import tensorplay as tp

x = tp.tensor([1.0, 2.0, 3.0], requires_grad=True)
y = (x * x).sum()
y.backward()
print(x.grad)  # tensor([2., 4., 6.])

If that prints 2 4 6 as a gradient, your installation is working.

Where to go next

• API reference — every function and class, grouped by area. Start at tensorplay for the core tensor API, then nn for layers and losses, and optim for optimizers.

• Deep-dive notes — the notes section explains the concepts behind the code: how autograd builds its graph, broadcasting rules, serialization, randomness, and more.