Deep Learning Mastery Quiz — Quiz

Tensors, training, regularisation, CNNs, sequence models, transformers and deployment in PyTorch. · 10 questions

Question 1 of 10

What does calling `loss.backward()` do in PyTorch?

aUpdates the model weightsbComputes gradients of the loss with respect to parameters that require gradcResets all gradients to zerodRuns the forward pass again

Question 2 of 10

Which layer type is best suited to capturing local spatial patterns in images?

ann.LSTMbnn.Embeddingcnn.Conv2ddnn.LayerNorm

Question 3 of 10

Why is `optimizer.zero_grad()` called at the start of each training step?

aBecause PyTorch accumulates gradients across backward calls by defaultbTo reset the learning ratecTo re-initialise the weights randomlydTo free the model from GPU memory

Question 4 of 10

What does `model.eval()` change?

aIt disables gradient computation entirelybIt freezes all parameters so they cannot be trainedcIt moves the model to the CPUdIt switches layers like dropout and batch norm to inference behaviour

Question 5 of 10

What does this code print?

import torch
x = torch.tensor(3.0, requires_grad=True)
y = x ** 2 + 4 * x
y.backward()
print(x.grad)

atensor(21.)btensor(10.)ctensor(6.)dNone

Question 6 of 10

What is the output shape?

import torch, torch.nn as nn
conv = nn.Conv2d(3, 16, kernel_size=3, padding=1)
x = torch.randn(8, 3, 64, 64)
print(conv(x).shape)

atorch.Size([8, 16, 64, 64])btorch.Size([8, 16, 62, 62])ctorch.Size([8, 3, 64, 64])dtorch.Size([16, 8, 64, 64])

Question 7 of 10

What inputs does `nn.CrossEntropyLoss` expect for multi-class classification?

aSoftmax probabilities and one-hot float labelsbSigmoid outputs and float labelscRaw logits and integer class indicesdLog-probabilities and string labels

Question 8 of 10

You have 400 labelled images across 4 classes. What is the best first approach?

aTrain a deep CNN from scratch for 200 epochsbUse an RNN over the pixel rowscTrain a transformer from scratchdUse a pretrained backbone, freeze it, and train a new classification head

Question 9 of 10

Why are attention scores divided by the square root of the key dimension?

aTo stop large dot products from pushing softmax into saturated regions with tiny gradientsbTo make attention weights sum to 1cTo reduce memory usage from quadratic to lineardTo add positional information to the tokens

Question 10 of 10

What shapes are printed?

import torch, torch.nn as nn
lstm = nn.LSTM(input_size=10, hidden_size=32, num_layers=2, batch_first=True)
x = torch.randn(4, 15, 10)
out, (h, c) = lstm(x)
print(out.shape, h.shape)

a(4, 15, 32) and (4, 2, 32)b(15, 4, 32) and (2, 4, 32)c(4, 15, 32) and (2, 4, 32)d(4, 32) and (2, 32)