mlp_lesson2

Introduction to Deep Learning with PyTorch

Ch3 Tensor

• Initialisation
  • From list
    • torch.Tensor([1, 2], [3, 4])
  • From numpy array
    • torch.from_numpy(np.array([1, 2], [3, 4]))
  • Like numpy
    • torch.zeros(size=shape) / zeros_like(tensor)
    • torch.ones(size=shape) / ones_like(tensor)
    • torch.randn(size=shape) / rand_like(tensor)
• Mathematical operations
  • +, * - element to element operation
• .item()
  • Torch object to python type
• .view(*shape)
  • -1 shape flattens the tensor
  • Shallow copies original tensor
• Supports slicing

Ch4 PyTorch for Automatic Gradient Descent

• torch.nn.Parameter
  • Tensor with gradient
  • .backward()
    • Computes loss
• optim module
  • Optimiser
    • torch.optim.SGD(params=thetas, lr=...)
  • .zero_grad()
    • Reset tensor gradient to 0
  • .step()
    • One update step