Figure from the original question paper

(A) Output size: 15 × 15; Effective receptive field: 5 × 5

(B) Output size: 17 × 17; Effective receptive field: 5 × 5

(C) Output size: 13 × 13; Effective receptive field: 5 × 5

(D) Output size: 15 × 15; Effective receptive field: 3 × 3

Figure from the original question paper

(A) ResNet-50 > Inception-v3 > VGG-19 > AlexNet

(B) ResNet-50 > VGG-19 > Inception-v3 > AlexNet

(C) Inception-v3 > ResNet-50 > VGG-19 > AlexNet

(D) VGG-19 > ResNet-50 > Inception-v3 > AlexNet

Consider the following PyTorch CNN model used for a classification task on a small dataset of 32 x 32 RGB images: class SmallCNN(nn.Module): def init(self): super(SmallCNN, self).init() self.conv = nn.Sequential( nn.Conv2d(3, 32, kernel_size=3, padding=1), nn.BatchNorm2d(32), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2), nn.Conv2d(64, 128, kernel_size=3, padding=1), nn.BatchNorm2d(128), nn.ReLU(), nn.MaxPool2d(2) ) self.fc = nn.Sequential( nn.Flatten(), nn.Linear(12888, 256), nn.ReLU(), nn.Linear(256, 10) ) def forward(self, x): x = self.conv(x) return self.fc(x) You replace this model with a much deeper VGG-19 network, hoping to improve performance. However, you observe that the performance remains the same or even slightly worsens. Which of the following is the most likely reason why this replacement did not help?

(C) The dataset is too small and simple for a deep architecture like VGG-19 to show its advantages, and may even overfit.

(A) VGG-19 uses too many skip connections, which interfere with shallow patterns.

(B) The original model had batch normalization, and VGG-19 lacks it by default.

(D) VGG-19 performs better only on grayscale images, not RGB images.

You are building a classifier for a medical imaging dataset with 5 categories (e.g., types of skin lesions). You decide to use ResNet-18 pretrained on ImageNet and fine-tune only the final layer. Below is a partially completed PyTorch code snippet: import torch import torch.nn as nn from torchvision.models import resnet18 class MedicalImageClassifier(nn.Module): def init(self): super(MedicalImageClassifier, self).init() self.base = resnet18(pretrained=True) # [Missing Line] def forward(self, x): return self.base(x) model = MedicalImageClassifier() Which of the following lines correctly fills in the missing line to adapt the ResNet-18 model for 5-class classification?

(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 QDB3 (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 QDB3: 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 111 × 111 (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: 15 × 15; Effective receptive field: 5 × 5
B
Output size: 17 × 17; Effective receptive field: 5 × 5
C
Output size: 13 × 13; Effective receptive field: 5 × 5
D
Output size: 15 × 15; Effective receptive field: 3 × 3

Show answer

Correct answer

A
Output size: 15 × 15; Effective receptive field: 5 × 5

Question 2 on its own page

Question 3

+5 marksOne correct option

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

Show answer

Correct answer

A
ResNet-50 > Inception-v3 > VGG-19 > AlexNet

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 32 x 32 RGB images:

python

1class SmallCNN(nn.Module):
2    def __init__(self):
3        super(SmallCNN, self).__init__()
4        self.conv = nn.Sequential(
5            nn.Conv2d(3, 32, kernel_size=3, padding=1),
6            nn.BatchNorm2d(32),
7            nn.ReLU(),
8            nn.Conv2d(32, 64, kernel_size=3, padding=1),
9            nn.BatchNorm2d(64),
10            nn.ReLU(),
11            nn.MaxPool2d(2),
12            nn.Conv2d(64, 128, kernel_size=3, padding=1),
13            nn.BatchNorm2d(128),
14            nn.ReLU(),
15            nn.MaxPool2d(2)
16        )
17        self.fc = nn.Sequential(
18            nn.Flatten(),
19            nn.Linear(128*8*8, 256),
20            nn.ReLU(),
21            nn.Linear(256, 10)
22        )
23
24    def forward(self, x):
25        x = self.conv(x)
26        return self.fc(x)

You replace this model with a much deeper VGG-19 network, hoping to improve performance. However, you observe that the performance remains the same or even slightly worsens.

Which of the following is the most likely reason why this replacement did not help?

A
VGG-19 uses too many skip connections, which interfere with shallow patterns.
B
The original model had batch normalization, and VGG-19 lacks it by default.
C
The dataset is too small and simple for a deep architecture like VGG-19 to show its advantages, and may even overfit.
D
VGG-19 performs better only on grayscale images, not RGB images.

Show answer

Correct answer

C
The dataset is too small and simple for a deep architecture like VGG-19 to show its advantages, and may even overfit.

Question 4 on its own page

Question 5

+5 marksOne correct option

You are building a classifier for a medical imaging dataset with 5 categories (e.g., types of skin lesions). You decide to use ResNet-18 pretrained on ImageNet and fine-tune only the final layer.

Below is a partially completed PyTorch code snippet:

python

1import torch
2import torch.nn as nn
3from torchvision.models import resnet18
4
5class MedicalImageClassifier(nn.Module):
6    def __init__(self):
7        super(MedicalImageClassifier, self).__init__()
8        self.base = resnet18(pretrained=True)
9        # [Missing Line]
10
11    def forward(self, x):
12        return self.base(x)
13
14model = MedicalImageClassifier()

Which of the following lines correctly fills in the missing line to adapt the ResNet-18 model for 5-class classification?

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 the Faster R-CNN object detection pipeline:

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

The blocks (not in order) are described as:

(a) Image warping / reconstruction using disparity.
(b) RoI Pooling
(c) Classification and Bounding Box Regression
(d) Input image
(e) Backbone CNN (e.g., ResNet)

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

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

Show answer

Correct answer

A
A–(d), B–(e), C–(a), D–(b), E–(c)

Question 6 on its own page

Question 7

+5 marksOne correct option

In the original U-Net architecture, which of the following statements about the decoder (expanding) path is correct?

A
The decoder upsamples the feature maps using 2 x 2 transposed convolutions and concatenates them with corresponding encoder features.
B
The decoder uses 3 x 3 max pooling to increase the spatial resolution before feature fusion.
C
The decoder applies 1 x 1 convolutions to increase spatial dimensions and reduce channel depth.
D
The decoder merges encoder and decoder features using element-wise addition instead of concatenation.

Show answer

Correct answer

A
The decoder upsamples the feature maps using 2 x 2 transposed convolutions and concatenates them with corresponding encoder features.

Question 7 on its own page

Question 8

+5 marksOne or more correct options

Which of the following are valid reasons why Convolutional Neural Networks (CNNs) are preferred over fully connected Multilayer Perceptrons (MLPs) for image processing tasks?

Select all that apply.

A
CNNs reduce the number of parameters by using local connections and weight sharing.
B
CNNs can process images of arbitrary size without any architectural changes.
C
CNNs exploit the spatial structure of images, making them more efficient at recognizing patterns.
D
CNNs perform better than MLPs because they always use deeper networks.
E
CNNs maintain spatial hierarchies by stacking convolution and pooling layers.
F
CNNs can generalize better to unseen data due to their built-in translational invariance.

Show answer

Correct answers

A
CNNs reduce the number of parameters by using local connections and weight sharing.
C
CNNs exploit the spatial structure of images, making them more efficient at recognizing patterns.
E
CNNs maintain spatial hierarchies by stacking convolution and pooling layers.
F
CNNs can generalize better to unseen data due to their built-in translational invariance.

Question 8 on its own page

Question 9

+5 marksOne or more correct options

In the Faster R-CNN architecture, the Region of Interest (RoI) Pooling layer plays a key role in the object detection pipeline.
Which of the following statements correctly describe the function and properties of RoI Pooling?

Select all that apply.

A
RoI Pooling converts variable-sized region proposals into fixed-size feature maps for the classifier head.
B
RoI Pooling crops the region proposals from the input image and resizes them to a fixed dimension.
C
RoI Pooling enables batching and fully connected layers by standardizing the input dimensions.
D
RoI Pooling uses bilinear interpolation to produce smoother feature maps for each proposal.
E
RoI Pooling operates on convolutional feature maps, not directly on the raw image.
F
RoI Pooling discards proposals with low objectness scores before passing them to the classifier.

Show answer

Correct answers

A
RoI Pooling converts variable-sized region proposals into fixed-size feature maps for the classifier head.
C
RoI Pooling enables batching and fully connected layers by standardizing the input dimensions.
E
RoI Pooling operates on convolutional feature maps, not directly on the raw image.

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 is faster at inference time because it predicts bounding boxes and class probabilities in a single pass.
B
YOLO is fully end-to-end trainable and does not rely on external region proposal mechanisms.
C
YOLO achieves higher recall than Faster R-CNN on all types of objects, especially small and overlapping ones.
D
YOLO formulates detection as a combination of regression (for bounding box coordinates and objectness score) and classification (for discrete object categories).
E
YOLO uses multi-scale feature fusion to improve detection of small objects.
F
YOLO’s architecture allows it to be deployed more easily on real-time and edge devices due to its compactness.

Show answer

Correct answers

A
YOLO is faster at inference time because it predicts bounding boxes and class probabilities in a single pass.
B
YOLO is fully end-to-end trainable and does not rely on external region proposal mechanisms.
D
YOLO formulates detection as a combination of regression (for bounding box coordinates and objectness score) and classification (for discrete object categories).
F
YOLO’s architecture allows it to be deployed more easily on real-time and edge devices due to its compactness.

Question 10 on its own page

Question 11

+5 marksOne or more correct options

Which of the following are NOT advantages of AlexNet over InceptionNet?

Select all that apply.

A
AlexNet is simpler and easier to implement compared to the complex multi- branch design of InceptionNet.
B
AlexNet requires fewer parameters than InceptionNet due to its shallow architecture.
C
AlexNet has better accuracy on ImageNet than InceptionNet due to its early adoption of ReLU and dropout.
D
AlexNet is faster to train on small datasets because of its shallow depth and simpler structure.
E
InceptionNet extracts multi-scale features using filters of different sizes in parallel, which improves accuracy.
F
AlexNet lacks 1 x 1 convolutions, which are used in InceptionNet for dimensionality reduction and increased efficiency.

Show answer

Correct answers

C
AlexNet has better accuracy on ImageNet than InceptionNet due to its early adoption of ReLU and dropout.
E
InceptionNet extracts multi-scale features using filters of different sizes in parallel, which improves accuracy.
F
AlexNet lacks 1 x 1 convolutions, which are used in InceptionNet for dimensionality reduction and increased efficiency.

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 the training dynamics and properties of Generative Adversarial Networks (GANs)?

Select all that apply.

A
The generator’s objective is to produce data that the discriminator cannot distinguish from real data.
B
The discriminator is trained to minimize the distance between the generated data and the real data distributions.
C
GAN training is a min-max game where the generator and discriminator have opposing goals.
D
Mode collapse refers to the generator producing diverse outputs that cover multiple modes of the data distribution.
E
A balance between generator and discriminator learning rates is important to ensure stable training.
F
Conditional GANs extend standard GANs by incorporating label information into both the generator and discriminator.

Show answer

Correct answers

A
The generator’s objective is to produce data that the discriminator cannot distinguish from real data.
C
GAN training is a min-max game where the generator and discriminator have opposing goals.
E
A balance between generator and discriminator learning rates is important to ensure stable training.
F
Conditional GANs extend standard GANs by incorporating label information into both the generator and discriminator.

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 about the differences between ESRGAN and SRGAN are NOT correct?

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 layers 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

B
SRGAN employs Perceptual Loss, while ESRGAN removes it to focus only on pixel-wise MSE loss for training stability.
D
Both SRGAN and ESRGAN use a VGG-based perceptual loss computed on feature maps extracted from deeper layers of a pretrained classifier.

Question 14 on its own page

Question 15

+5 marksOne or more correct options

The Multi-stage Progressive Image Restoration Network (MPRNet) is designed to progressively restore degraded images across multiple stages. It integrates several architectural innovations to enhance image restoration performance. How many of the above components are explicitly part of the original MPRNet architecture?

Select all that apply.

A
UNet-style encoder-decoder at each stage
B
Residual Channel Attention Network (RCAN)
C
Supervised Attention Module (SAM)
D
Skip connections across stages
E
Non-local Attention Module

Show answer

Correct answers

A
UNet-style encoder-decoder at each stage
C
Supervised Attention Module (SAM)
D
Skip connections across stages

Question 15 on its own page

Question 16

+5 marksNumerical answer

You are given a feature map of size $64 \times 64 \times 64$ (height $\times$ width $\times$ channels). Three convolution operations are applied in parallel on this input, maintaining the spatial dimensions:

$1 \times 1$ convolution with 128 filters
$3 \times 3$ convolution with 64 filters
$5 \times 5$ convolution with 32 filters

Each convolution operates on all input channels (i.e., $C_{\text{in}} = 64$ ). Compute the total number of operations performed across all three convolutions.

Show answer

Correct answer: 394264576

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: $(20, 20)$ to $(50, 50)$
Predicted Box 2: $(60, 60)$ to $(80, 80)$
Predicted Box 3: $(100, 100)$ to $(130, 130)$

The ground truth boxes for these objects are:

Ground Truth Box 1: $(30, 30)$ to $(60, 60)$
Ground Truth Box 2: $(55, 55)$ to $(75, 75)$
Ground Truth Box 3: $(110, 110)$ to $(140, 140)$

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: 0.32 (accepted within ±0.02)

Question 18 on its own page

Question 19

+5 marksNumerical answer

Show answer

Correct answer: 25

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: $\text{MSE}_1 = 150.25$
Model 2: $\text{MSE}_2 = 210.90$

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: 26.35 (accepted within ±0.02)

Question 20 on its own page

Question 21

+1.5 marksNumerical answer

The reported MSE values are:

Model 1: $\text{MSE}_1 = 150.25$
Model 2: $\text{MSE}_2 = 210.90$

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: 24.88 (accepted within ±0.02)

Question 21 on its own page

Question 22

+2 marksNumerical answer

The reported MSE values are:

Model 1: $\text{MSE}_1 = 150.25$
Model 2: $\text{MSE}_2 = 210.90$

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: 1

Question 22 on its own page