/pytorch-useful-stuff

Pytorch useful stuff by me

Collecting from the Pytorch Tutorial and the other sources (updating)

Torch Operations:

I have been listing some useful and common operations here to easily keep traces when coding

  • Tensor:

    • Given x and y are Tensor matrices : Add operation:
      x + y
  torch.add(x,y, out=result)
  y.add_(x)

    Any operations that mutates a tensor in-place is post-fixed with and ' _ '. For example: x.copy_()y

    • More than 100 operations of torch in this doc
    • Converting Torch Tensor to Numpy and versa
      # torch tensor to numpy
  a = torch.ones(5)
  b = a.numpy()
  # numpy to torch tensor
  a = np.ones(5)
  b = torch.from_numpy(a)
    • Resize/reshape: use torch.view()-z = x.view(-1, n) #the size -1 inferred from other dimensons
    • Tensors can be moved onto GPU using .cuda() method
      if torch.cuda.is_available():
    x = x.cuda()
    y = y.cuda()
    x + y
    • torch.mm(mat1, mat2, output=None) multiplies two matrix mat1 and mat2
    • torch.mv(mat1, vec1, output=None) multiplies a matrix mat1 and vec1
    • torch.clamp(input, min, max, output) clamp all elements in input into the range [min, max] and return a resulting tensor

Autograd: Automatic differentiation

The autograd package provides automatic differentiation for all operations on Tensors

  • Variable: autograd.Variable is a wrapper of almost operations defined on Tensor. After all computation, call backward() to get all gradients computed automatically. Access the raw tensor through .data attributes and gradients through .grad
  • Variable