A tensor is an array with any number of dimensions. In PyTorch, tensors hold both the data flowing through a neural network and the parameters that control it. This common representation makes tensor operations available to GPU computation and automatic differentiation. section 1.2 section 1.7
The number of axes tells you what kind of array you have:
- Zero axes: a scalar, such as `torch.tensor(3.14)`, with shape `torch.Size([])`. section 1.2
- One axis: a vector. `torch.tensor([1, 2, 3, 4])` has shape `torch.Size([4])`. section 1.2
- Two axes: a matrix. Four rows of two values have shape `torch.Size([4, 2])`. Adding an axis gives a three-dimensional array; the supplied example has shape `torch.Size([4, 2, 3])`. section 1.2
print(torch_scalar.shape) print(torch_vector.shape) print(torch_matrix.shape) print(torch_tensor3d.shape) torch.Size([]) torch.Size([4]) torch.Size([4, 2]) torch.Size([4, 2, 3])
Compare the four printed shapes: each added axis contributes another size to `torch.Size`. section 1.2
Shape names make code easier to check. For example, `(B, C, W, H)` names four axes; a weight matrix written `W` with shape `(d, C)` has `d` feature rows and `C` output columns. In a linear layer, `nn.Linear(d, C)` creates the corresponding input-to-output mapping. For matrix multiplication, shapes `(n, m)` and `(m, p)` produce `(n, p)`: the shared `m` dimension must line up. section 1.2 section 2.1.5 section 1.2.1
Move between NumPy and PyTorch
Given a NumPy array `x_np`, `x_pt = torch.tensor(x_np)` makes a PyTorch tensor. Check `x_pt.dtype` rather than assuming its numeric type: a NumPy `float64` array produces a `torch.float64` tensor in the example. To request single precision, use `torch.tensor(x_np, dtype=torch.float32)`. Familiar operations carry over: `torch.sum(x_pt)` sums values, while `torch.transpose(x_pt, 0, 1)` exchanges two axes. For a tensor of shape `(4, 2, 3)`, exchanging axes 0 and 2 yields `(3, 2, 4)`. section 1.2
Slices can select a row or column from a matrix, producing a smaller tensor. Which axis means what depends on the data: a matrix might hold a dataset, while a three-dimensional tensor might hold a dataset of sequences. section 1.2
A GPU can speed up large tensor calculations, but moving data between CPU and GPU has a cost. Benchmark small workloads rather than assuming a GPU is faster. Move a tensor with `x = x.to(device)`; tensors used together must be on the same device. A CUDA tensor must be copied back to CPU before NumPy can use it. section 1.2.1















































