PyTorch provides two useful functions, torch.zeros and torch.ones, to create tensors filled entirely with zeros or ones, respectively. These functions are commonly used for initializing tensors in machine learning and deep learning tasks.

Zero initialization is primarily utilized for defining biases or placeholders in machine learning models, providing a neutral starting point that does not affect initial computations. Conversely, ones initialization is valuable in scaling operations or additive computations, offering consistency and ensuring predictable behavior during the process.

Let’s create a tensor full of zeros using torch.zeros() :

# create a tensor of all zeros
zeros = torch.zeros(size = (3, 3))
zeros, zeros.dtype, zeros.ndim

(tensor([[0., 0., 0.],
         [0., 0., 0.],
         [0., 0., 0.]]),
 torch.float32,
 2)

Similarly, we can create a tensor of all ones using torch.ones()

# create a tensor of all ones
ones = torch.ones(size = (3, 3))
ones, ones.dtype, ones.ndim

(tensor([[1., 1., 1.],
         [1., 1., 1.],
         [1., 1., 1.]]),
 torch.float32,
 2)