Figure from the original question paper

(B) Output size: 10 × 10; Effective receptive field: 5 × 5

(A) Output size: 8 × 8; Effective receptive field: 5 × 5

(C) Output size: 10 × 10; Effective receptive field: 3 × 3

(D) Output size: 12 × 12; Effective receptive field: 7 × 7

Consider the following PyTorch CNN model used for a classification task on a small dataset of 64 \times 64 RGB images: class SimpleCNN(nn.Module): def init(self): super(SimpleCNN, self).init() self.conv = nn.Sequential( nn.Conv2d(3, 64, kernel_size=3, padding=1), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, padding=1), nn.ReLU(), nn.MaxPool2d(2), nn.Conv2d(128, 256, kernel_size=3, padding=1), nn.ReLU(), nn.MaxPool2d(2) ) self.fc = nn.Sequential( nn.Flatten(), nn.Linear(2561616, 512), nn.ReLU(), nn.Linear(512, 10) ) def forward(self, x): x = self.conv(x) return self.fc(x) Suppose the model is underperforming and you consider replacing it with a deeper ResNet architecture. However, it turns out that this change does not improve performance significantly. Which of the following is the most likely reason why switching to ResNet does not help?

(B) Residual connections are only useful when training very deep networks, and this task does not require such depth.

(A) ResNet has too many parameters for this simple task.

(C) ResNet architectures perform poorly on low-resolution inputs due to their reliance on fixed feature scales.

(D) InceptionNet would have been better because it uses multiple receptive fields in parallel.

You are building an image classifier for the Nature12K dataset, which contains 12 categories of natural scenes (i.e., 12 output classes). You decide to use VGG-16 pretrained on ImageNet as a feature extractor and modify only the final layer for fine-tuning. Below is a partially completed PyTorch code snippet: import torch import torch.nn as nn from torchvision.models import vgg16 class Nature12KClassifier(nn.Module): def init(self): super(Nature12KClassifier, self).init() self.base = vgg16(pretrained=True) # [Missing Line] def forward(self, x): x = self.base.features(x) x = self.base.avgpool(x) x = torch.flatten(x, 1) x = self.base.classifier(x) return x model = Nature12KClassifier() Which of the following lines correctly fills in the missing line to adapt the VGG-16 model for Nature12K?

(A) Option figure from the original question paper

(B) Option figure from the original question paper

(C) Option figure from the original question paper

(D) Option figure from the original question paper

May 2025 term · Deep Learning Practice · BSDA5013

Deep Learning Practice End Term: 31 August 2025, Set QDB1 (May 2025 term)

The IIT Madras BS Deep Learning Practice (Deep Learning Practice) End Term paper sat on 31 Aug 2025, in the May 2025 term, set QDB1: 22 questions for 100 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.

Questions: 22
Marks: 100
Duration: 180 min
MCQ: 7
MSQ: 8
Numerical: 7

Take as mock test Practise with answers

Updated 25 Sept 2026

Official paper: IIT M DEGREE AN EXAM QDB3 31 Aug 2025 · No negative marking.

Question 1

+5 marksOne correct option

Consider the task of applying a convolutional layer with a stride of 1 to an image of size 227 × 227 (Grayscale), while ensuring that the output feature map retains the same spatial dimensions. Which of the following code snippets correctly implements this using PyTorch?

Show answer

Correct answer

Question 1 on its own page

Question 2

+5 marksOne correct option

A
Output size: 8 × 8; Effective receptive field: 5 × 5
B
Output size: 10 × 10; Effective receptive field: 5 × 5
C
Output size: 10 × 10; Effective receptive field: 3 × 3
D
Output size: 12 × 12; Effective receptive field: 7 × 7

Show answer

Correct answer

B
Output size: 10 × 10; Effective receptive field: 5 × 5

Question 2 on its own page

Question 3

+5 marksOne correct option

You are given the following CNN architectures:

AlexNet
VGG-19
Inception-v3
ResNet-50

Which of the following correctly orders these networks in terms of their increasing depth (i.e., from the fewest to the greatest number of layers)?

A
AlexNet < InceptionNet-v3 < VGG-19 < ResNet-50
B
AlexNet < VGG-19 < Inception-v3 < ResNet-50
C
Inception-v3 < AlexNet < VGG-19 < ResNet-50
D
AlexNet < Inception-v3 < ResNet-50 < VGG-19

Show answer

Correct answer

B
AlexNet < VGG-19 < Inception-v3 < ResNet-50

Question 3 on its own page

Question 4

+5 marksOne correct option

Consider the following PyTorch CNN model used for a classification task on a small dataset of $64 \times 64$ RGB images:

python

1class SimpleCNN(nn.Module):
2    def __init__(self):
3        super(SimpleCNN, self).__init__()
4        self.conv = nn.Sequential(
5            nn.Conv2d(3, 64, kernel_size=3, padding=1),
6            nn.ReLU(),
7            nn.Conv2d(64, 128, kernel_size=3, padding=1),
8            nn.ReLU(),
9            nn.MaxPool2d(2),
10            nn.Conv2d(128, 256, kernel_size=3, padding=1),
11            nn.ReLU(),
12            nn.MaxPool2d(2)
13        )
14        self.fc = nn.Sequential(
15            nn.Flatten(),
16            nn.Linear(256*16*16, 512),
17            nn.ReLU(),
18            nn.Linear(512, 10)
19        )
20
21    def forward(self, x):
22        x = self.conv(x)
23        return self.fc(x)

Suppose the model is underperforming and you consider replacing it with a deeper ResNet architecture. However, it turns out that this change does not improve performance significantly.

Which of the following is the most likely reason why switching to ResNet does not help?

A
ResNet has too many parameters for this simple task.
B
Residual connections are only useful when training very deep networks, and this task does not require such depth.
C
ResNet architectures perform poorly on low-resolution inputs due to their reliance on fixed feature scales.
D
InceptionNet would have been better because it uses multiple receptive fields in parallel.

Show answer

Correct answer

B
Residual connections are only useful when training very deep networks, and this task does not require such depth.

Question 4 on its own page

Question 5

+5 marksOne correct option

You are building an image classifier for the Nature12K dataset, which contains 12 categories of natural scenes (i.e., 12 output classes). You decide to use VGG-16 pretrained on ImageNet as a feature extractor and modify only the final layer for fine-tuning.

Below is a partially completed PyTorch code snippet:

python

1import torch
2import torch.nn as nn
3from torchvision.models import vgg16
4
5class Nature12KClassifier(nn.Module):
6    def __init__(self):
7        super(Nature12KClassifier, self).__init__()
8        self.base = vgg16(pretrained=True)
9        # [Missing Line]
10
11    def forward(self, x):
12        x = self.base.features(x)
13        x = self.base.avgpool(x)
14        x = torch.flatten(x, 1)
15        x = self.base.classifier(x)
16        return x
17
18model = Nature12KClassifier()

Which of the following lines correctly fills in the missing line to adapt the VGG-16 model for Nature12K?

Show answer

Correct answer

Question 5 on its own page

Question 6

+5 marksOne correct option

The diagram below represents the block-wise flow of an unsupervised depth estimation pipeline using left-right stereo image pairs.

$\boxed{A} \longrightarrow \boxed{B} \longrightarrow \boxed{C} \longrightarrow \boxed{D} \longrightarrow \boxed{E}$

The blocks (not in order) are described as:

(a) Image warping / reconstruction using disparity.
(b) Photometric loss computation
(c) Disparity map prediction
(d) Input stereo pair
(e) Left-right consistency loss

Match the blocks to the diagram and choose the correct sequence of labels corresponding to boxes A through E.

A
A–(e), B–(c), C–(a), D–(d), E–(b)
B
A–(e), B–(b), C–(c), D–(a), E–(d)
C
A–(e), B–(a), C–(c), D–(d), E–(b)
D
A–(c), B–(e), C–(a), D–(b), E–(d)

Show answer

Correct answer

A
A–(e), B–(c), C–(a), D–(d), E–(b)

Question 6 on its own page

Question 7

+5 marksOne correct option

Consider the original U-Net architecture.
Which of the following statements about the convolutional and pooling operations is correct?

A
U-Net uses 5 × 5 convolutional filters and 2 × 2 max pooling filters with stride 2 in the encoder path.
B
U-Net uses 3 × 3 convolutional filters and 2 × 2 average pooling filters with stride 1 in both encoder and decoder.
C
U-Net uses 3 × 3 convolutional filters (with padding) and 2 × 2 max pooling filters with stride 2 to halve the spatial resolution in the encoder.
D
U-Net uses 3 × 3 filters and 1 × 1 pooling to preserve spatial dimensions between contracting and expanding paths.

Show answer

Correct answer

C
U-Net uses 3 × 3 convolutional filters (with padding) and 2 × 2 max pooling filters with stride 2 to halve the spatial resolution in the encoder.

Question 7 on its own page

Question 8

+5 marksOne or more correct options

Which of the following are valid advantages of using Convolutional Neural Networks (CNNs) instead of fully connected Multilayer Perceptrons (MLPs) for image-based tasks?

Select all that apply.

A
CNNs require significantly fewer parameters than MLPs for high-dimensional image inputs.
B
CNNs are better at capturing spatial hierarchies due to weight sharing and local connectivity.
C
CNNs can approximate any function with fewer layers than MLPs due to the universal approximation theorem.
D
CNNs are translation invariant by design, making them robust to object position shifts in images.
E
CNNs do not require any activation functions since convolutions are linear operators.
F
CNNs are computationally faster than MLPs on image data due to sparse connections and weight sharing.

Show answer

Correct answers

A
CNNs require significantly fewer parameters than MLPs for high-dimensional image inputs.
B
CNNs are better at capturing spatial hierarchies due to weight sharing and local connectivity.
D
CNNs are translation invariant by design, making them robust to object position shifts in images.
F
CNNs are computationally faster than MLPs on image data due to sparse connections and weight sharing.

Question 8 on its own page

Question 9

+5 marksOne or more correct options

In the Faster R-CNN architecture, the Region Proposal Network (RPN) plays a critical role in object detection.
Which of the following statements correctly describe the role of the RPN?

Select all that apply.

A
The RPN generates a set of class-agnostic region proposals likely to contain objects.
B
The RPN performs classification and bounding box regression to output object classes and their exact locations.
C
The RPN uses anchor boxes of different scales and aspect ratios to slide over convolutional feature maps.
D
The RPN replaces traditional region proposal methods like selective search, making the pipeline faster and end-to-end trainable.
E
The RPN outputs fixed-size feature maps that are directly used for object classification.
F
The RPN filters out background regions by assigning objectness scores to anchors.

Show answer

Correct answers

A
The RPN generates a set of class-agnostic region proposals likely to contain objects.
C
The RPN uses anchor boxes of different scales and aspect ratios to slide over convolutional feature maps.
D
The RPN replaces traditional region proposal methods like selective search, making the pipeline faster and end-to-end trainable.
F
The RPN filters out background regions by assigning objectness scores to anchors.

Question 9 on its own page

Question 10

+5 marksOne or more correct options

Which of the following are valid advantages of YOLO compared to Fast R-CNN or Faster R-CNN?

Select all that apply.

A
YOLO performs object classification and localization in a single forward pass, making it faster than two-stage detectors like Fast R-CNN and Faster R-CNN.
B
YOLO does not require region proposal generation, making the entire pipeline fully end-to-end trainable.
C
YOLO achieves higher localization accuracy than Faster R-CNN on small objects due to its coarse grid-based prediction.
D
YOLO treats bounding box coordinates and objectness score as regression outputs, while computing class probabilities through classification.
E
YOLO achieves higher precision by using a large number of anchor boxes per region.
F
YOLO treats object detection as a regression problem, which simplifies the detection pipeline and reduces computational overhead.

Show answer

Correct answers

A
YOLO performs object classification and localization in a single forward pass, making it faster than two-stage detectors like Fast R-CNN and Faster R-CNN.
B
YOLO does not require region proposal generation, making the entire pipeline fully end-to-end trainable.
D
YOLO treats bounding box coordinates and objectness score as regression outputs, while computing class probabilities through classification.
F
YOLO treats object detection as a regression problem, which simplifies the detection pipeline and reduces computational overhead.

Question 10 on its own page

Question 11

+5 marksOne or more correct options

Which of the following are advantages of ResNet over VGG-19?

Select all that apply.

A
ResNet requires fewer parameters than VGG-19 while achieving higher accuracy.
B
ResNet’s residual connections help mitigate vanishing gradients in very deep networks.
C
On the CIFAR-10 dataset, ResNet achieves higher training and validation accuracy using significantly fewer convolutional layers than VGG-19.
D
ResNet models can be trained with depths exceeding 100 layers without suffering from degradation problems.
E
ResNet blocks require more memory and training time due to the skip connections compared to VGG-19.
F
VGG-19 is easier to deploy on edge devices because of its lightweight residual architecture.

Show answer

Correct answers

A
ResNet requires fewer parameters than VGG-19 while achieving higher accuracy.
B
ResNet’s residual connections help mitigate vanishing gradients in very deep networks.
D
ResNet models can be trained with depths exceeding 100 layers without suffering from degradation problems.

Question 11 on its own page

Question 12

+5 marksOne or more correct options

Which of the following code snippets can be used for data augmentation during training in PyTorch?

Select all that apply.

Show answer

Correct answers

Question 12 on its own page

Question 13

+5 marksOne or more correct options

Which of the following statements are correct regarding Generative Adversarial Networks (GANs)?

Select all that apply.

A
The generator tries to minimize the probability that the discriminator correctly classifies fake samples.
B
The discriminator is trained to maximize the probability of correctly classifying real and generated (fake) data.
C
The generator directly receives labeled training data to learn the target distribution.
D
During training, the generator is updated based on the discriminator’s feedback to improve sample quality.
E
GAN training can suffer from instability due to the adversarial objective, leading to issues like mode collapse.
F
A perfectly trained discriminator ensures faster convergence of the generator.

Show answer

Correct answers

A
The generator tries to minimize the probability that the discriminator correctly classifies fake samples.
B
The discriminator is trained to maximize the probability of correctly classifying real and generated (fake) data.
D
During training, the generator is updated based on the discriminator’s feedback to improve sample quality.
E
GAN training can suffer from instability due to the adversarial objective, leading to issues like mode collapse.

Question 13 on its own page

Question 14

+5 marksOne or more correct options

SRGAN and ESRGAN are both deep learning architectures designed for single image super- resolution. Which of the following statements correctly describe the key differences between ESRGAN and SRGAN?

Select all that apply.

A
ESRGAN replaces the standard residual blocks in SRGAN with Residual-in- Residual Dense Blocks (RRDB) for better performance.
B
SRGAN employs Perceptual Loss, while ESRGAN removes it to focus only on pixel-wise MSE loss for training stability.
C
ESRGAN produces sharper and more realistic textures than SRGAN due to architectural and loss function improvements.
D
Both SRGAN and ESRGAN use a VGG-based perceptual loss computed on feature maps extracted from deeper and shallower layers, respectively, of a pretrained classifier.
E
SRGAN is trained using only adversarial and perceptual losses, whereas ESRGAN includes a content loss computed using pixel-wise L1 loss.

Show answer

Correct answers

A
ESRGAN replaces the standard residual blocks in SRGAN with Residual-in- Residual Dense Blocks (RRDB) for better performance.
C
ESRGAN produces sharper and more realistic textures than SRGAN due to architectural and loss function improvements.
E
SRGAN is trained using only adversarial and perceptual losses, whereas ESRGAN includes a content loss computed using pixel-wise L1 loss.

Question 14 on its own page

Question 15

+5 marksOne or more correct options

The Multi-stage Progressive Image Restoration Network (MPRNet) is a state-of-the-art architecture for tasks like image deblurring and deraining. It follows a multi-stage design where features are progressively refined across stages. how many of the above components are explicitly part of the original MPRNet architecture?

Select all that apply.

A
a single layer neural network.
B
UNet
C
Supervised Attention Module
D
Channel Attention Block
E
Fast RCNN block

Show answer

Correct answers

B
UNet
C
Supervised Attention Module
D
Channel Attention Block

Question 15 on its own page

Question 16

+5 marksNumerical answer

You are given a feature map of size $32 \times 32 \times 128$ (height $\times$ width $\times$ channels). Three convolution operations are applied in parallel on this input with stride=1 and padding=1, maintaining the spatial dimensions:

$1 \times 1$ convolution with 256 filters
$3 \times 3$ convolution with 128 filters
$5 \times 5$ convolution with 64 filters

Each convolution operates on all input channels (i.e., $C_{\text{in}} = 128$ ). Compute the total number of operations performed across all three convolutions.[Assume bias=0]

Show answer

Correct answer: 3029908880

Question 16 on its own page

Question 17

+5 marksNumerical answer

Show answer

Correct answer: 0.63 (accepted within ±0.02)

Question 17 on its own page

Question 18

+5 marksNumerical answer

You are evaluating an object detection model using the Mean Intersection over Union (mIoU) metric.

The model predicts the following 3 bounding boxes for a given image:

Predicted Box 1: $(10, 10)$ to $(30, 30)$
Predicted Box 2: $(40, 40)$ to $(60, 60)$
Predicted Box 3: $(70, 70)$ to $(90, 90)$

The ground truth boxes for these objects are:

Ground Truth Box 1: $(15, 15)$ to $(35, 35)$
Ground Truth Box 2: $(40, 40)$ to $(60, 60)$
Ground Truth Box 3: $(80, 80)$ to $(100, 100)$

The Mean IoU is defined as:

$\text{Mean IoU} = \frac{1}{N} \sum_{i=1}^{N} \frac{\text{Area of } B_i^{\text{pred}} \cap B_i^{\text{gt}}}{\text{Area of } B_i^{\text{pred}} \cup B_i^{\text{gt}}}$

Compute the Mean IoU (rounded to two decimal places) over the 3 box pairs.

Show answer

Correct answer: 0.51 (accepted within ±0.02)

Question 18 on its own page

Question 19

+5 marksNumerical answer

Show answer

Correct answer: 64

Question 19 on its own page

Question 20

+1.5 marksNumerical answer

Two models, Model 1 and Model 2, are used to reconstruct grayscale images with pixel intensity values in the range $[0, 255]$ . Their performance is evaluated by computing the Mean Squared Error (MSE) between the reconstructed images and their corresponding ground truth images.

The reported MSE values are:

Model 1: $\text{MSE}_1 = 205.80$
Model 2: $\text{MSE}_2 = 180.50$

The Peak Signal-to-Noise Ratio (PSNR) is calculated using the formula:

$\text{PSNR} = 10 \cdot \log_{10}\left(\frac{MAX^2}{\text{MSE}}\right), \quad \text{where } MAX = 255$

Based on the above data, answer the given subquestions.

Compute the PSNR for Model 1 (in dB) (rounded to 2 decimal places) :

Show answer

Correct answer: 25.00

Question 20 on its own page

Question 21

+1.5 marksNumerical answer

The reported MSE values are:

Model 1: $\text{MSE}_1 = 205.80$
Model 2: $\text{MSE}_2 = 180.50$

The Peak Signal-to-Noise Ratio (PSNR) is calculated using the formula:

$\text{PSNR} = 10 \cdot \log_{10}\left(\frac{MAX^2}{\text{MSE}}\right), \quad \text{where } MAX = 255$

Based on the above data, answer the given subquestions.

Compute the PSNR for Model 2 (in dB) (rounded to 2 decimal places):

Show answer

Correct answer: 25.57 (accepted within ±0.02)

Question 21 on its own page

Question 22

+2 marksNumerical answer

The reported MSE values are:

Model 1: $\text{MSE}_1 = 205.80$
Model 2: $\text{MSE}_2 = 180.50$

The Peak Signal-to-Noise Ratio (PSNR) is calculated using the formula:

$\text{PSNR} = 10 \cdot \log_{10}\left(\frac{MAX^2}{\text{MSE}}\right), \quad \text{where } MAX = 255$

Based on the above data, answer the given subquestions.

Which model produces an image more similar to the ground truth?

Show answer

Correct answer: 2

Question 22 on its own page