Tutorial 1

Question 1

torch.tensor()

Answer 1

A torch.Tensor is a multi-dimensional matrix containing elements of a single data type

Question 2

How CPU and GPU torch.tensor() types exist?

Answer 2

8 CPU and 8 GPU

Question 3

Can a torch.tensor() be created from a Python list?

Answer 3

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

Question 4

Does torch.tensor() copy data?

Answer 4

Yes. If you want to change a numpy array to a tensor, use as_tensor()

Question 5

How do you use slice notation with torch.tensor()?

Answer 5

Like you would a numpy array.

Question 6

what is requires_grad argument in torch.tensor()?

Answer 6

It allows Pytorch to record operations including that tensor for automatic differentiation.

Question 7

How do you see the value of a torch.tensor()?

Answer 7

.item()

Question 8

Where do gradients accumulate when calling .backward()?

Answer 8

In the leaves

Question 9

What does a .backward() call do?

Answer 9

Computes the dloss/dx for every gradient enabled variable, and accumulates into the x.grad for every parameter x

Question 10

torch.randn(size)?

Answer 10

provides a pytorch tensor with each element sampled from the standard normal distribution

Question 11

torch.nn.Linear(in_features, out_features, bias=True)?

Answer 11

Applies a linear transformation to the incoming data: y = xW^T + b. W and b are randomly initialized

Question 12

what does nn.loss() create?

Answer 12

Creates a criterion, which you can then supply an input and target for that loss function.

Question 13

what does optimizer.step() do?

Answer 13

Updates the parameters based on current gradients stored in the .grad attribute of each paramter.

Question 14

What is a Pytorch parameter?

Answer 14

A parameter is still a Tensor, much like a normal variable, however when associated with a model, it becomes part of the model attributes. This means we can easily feed model parameters to the optimizer. Parameters automatically has requires_grad() set to true.

Question 15

What does sub() and add() do?

Answer 15

Add the argument to the tensor. If multiple elements, the argument is added to each element.

Question 16

What is the difference between sub() and sub_()?

Answer 16

sub_() is an in-place tensor operation.

Question 17

How do you convert from Pytorch tensor to a numpy array?

Answer 17

.numpy()

Question 18

What methods do you have to override to create a custom python dataset?

Answer 18

__len__(self) and __getitem__(self, index)

Question 19

What should __len__() do in a dataset?

Answer 19

Give the length of the dataset

Question 20

What should __getitem__(index) do in a dataset?

Answer 20

Give you the value in the dataset at index, after loading, transforming, etc.

Question 21

What should __init__(self) do in a dataset?

Answer 21

Initialize filepaths or list of filenames

Question 22

What does .state_dict() return

Answer 22

A dictionary containing the whole state of the module.

Question 23

How do you save a pytorch model?

Answer 23

You can do torch.save(model, ‘model.ckpt’), or torch.save(model.state_dict, ‘model.ckpt’). The latter is preferred as then intermediates are not saved.

Question 24

How do you load from state_dict?

Answer 24

model = TheModelClass(*args, **kwargs)

model.load_state_dict(torch.load(PATH)). Note you have to load the model first before you need to load_state_dict