Introduction

Matrices/Tensors

Construction

Matlab	Numpy	Pytorch
NaN(5, 3)	np.ones([5,3]) * np.nan	torch.empty(5, 3)
rand(5, 3)	np.random.rand(5,3)	torch.rand(5, 3)
zeros(5, 3)	np.zeros([5,3])	torch.zeros(5, 3)
ones(5, 3)	np.ones([5,3])	torch.ones(5, 3)

Convert from other types using using pre-existing tensor

Or from python list or np.array (MAKES COPY!)

torch.tensor([[1., -1.], [1., -1.]])
torch.tensor(np.array([[1, 2, 3], [4, 5, 6]]))

# Other option
a = np.ones(5)
torch.from_numpy(a)

Tensor copy with a new size

tensor = torch.ones((2,), dtype=torch.float64)
tensor.new_full((3, 4), 3.141592)

Option: requires_grad (Autograd)

Use autograd option (requires_grad=true) to add tensor in computational network (gradients)

t1 = torch.randn((3,3), requires_grad = True) 

#remove from compuations 
t1.detach()

Options: torch.device

Select device to store tensor

Option: torch.dtype

Torch	CPU tensor	Cuda Tensor
torch.bfloat16	torch.BFloat16Tensor	torch.cuda.BFloat16Tensor
torch.float	torch.FloatTensor	torch.cuda.FloatTensor
torch.uint8	torch.ByteTensor	torch.cuda.ByteTensor
torch.bool	torch.BoolTensor	torch.cuda.BoolTensor

Shapes

| Matlab | Numpy | Pytorch | |:-: |:-: |:-: | | reshape(x,3,4) | x.reshape(3,4,order=’F’).copy() | view |

Template

This document will be updated regulary

Matlab	Numpy	Pytorch

References

1. https://pytorch.org/tutorials/beginner/blitz/tensor_tutorial.html
2. https://blog.paperspace.com/pytorch-101-understanding-graphs-and-automatic-differentiation/
3. https://numpy.org/doc/stable/user/numpy-for-matlab-users.html

Last update:29 November 2020