January 2026 term · Deep Learning Practice · BSDA5013

Deep Learning Practice End Term: 10 May 2026 (January 2026 term)

The IIT Madras BS Deep Learning Practice (Deep Learning Practice) End Term paper sat on 10 May 2026, in the January 2026 term: 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
Numerical: 7
MSQ: 8
MCQ: 7

Take as mock test Practise with answers

Updated 25 Sept 2026

Official paper: Deep Learning Practice 06 May 26 · No negative marking.

Question 1

+1 markNumerical answer

Read the following configuration file and code snippet related to training and annotations in a YOLO model, and answer the given subquestions.
A dataset configuration file (YAML) is defined as:
path: /data/helmet_dataset train: images/train val: images/val nc: 5 names: [’helmet’, ’no_helmet’, ’person’,’car’, ’truck’] A sample annotation in YOLO format is given below: 2 0.25 0.60 0.40 0.20 This represents: classid, xcenter, ycenter,width, height. Assume the corresponding image has dimensions: 640 (width) × 480 (height).

Based on the YAML configuration file, what is the total number of object classes the model is trained to detect?

Show answer

Correct answer: 5

Question 1 on its own page

Question 2

+2 marksNumerical answer

What is the x-coordinate (in pixels) of the center of the bounding box?

Show answer

Correct answer: 160

Question 2 on its own page

Question 3

+2 marksNumerical answer

What is the x-coordinate (in pixels) of the top-left corner of the bounding box?

Show answer

Correct answer: 32

Question 3 on its own page

Question 4

+5 marksNumerical answer

Consider a convolutional neural network with an input image of size 84 × 84 × 3. The architecture consists of the following layers:
1.A convolutional layer with 16 filters of size 4 × 4, a stride of 2, and no padding. 2. A ReLU activation layer. 3. A second convolutional layer with 32 filters of size 2×2, a stride of 2, and no padding. 4. A flattening layer.
What is the dimension of the resulting flattened layer?

Show answer

Correct answer: 12800 (accepted within ±100)

Question 4 on its own page

Question 5

+5 marksNumerical answer

A bottleneck layer in a ResNet receives an input feature map of dimensions 14×14×1024. We apply a 1×1 convolution layer with 256 filters to reduce dimensionality. Calculate the total number of multiplications required for this operation. (Give answer in millions, e.g., if 65,452,123, write 65.45).

Show answer

Correct answer: 51.35 (accepted within ±0.15)

Question 5 on its own page

Question 6

+5 marksNumerical answer

A modified YOLO-style architecture divides an image into a 13 × 13 grid (S = 13). If the model is designed to predict B = 3 bounding boxes per grid cell and is trained to recognize C = 80 distinct classes (similar to COCO dataset), what is the depth (number of channels) of the final output tensor?

Show answer

Correct answer: 95

Question 6 on its own page

Question 7

+5 marksNumerical answer

In the VGG16 architecture, consider the second convolutional layer of the first block. It takes an input with 64 channels and applies 64 filters of size 3 × 3. Calculate the total number of weights (excluding bias) for this specific layer.

Show answer

Correct answer: 36864

Question 7 on its own page

Question 8

+5 marksOne or more correct options

When fine-tuning a large vision model (e.g., a pre-trained ResNet or Vision Transformer) for a specific competition dataset, which of the following techniques are most effective for improving training efficiency (speed and memory usage)?

Select all that apply.

A
Mixed Precision Training (FP16): Using 16-bit floating-point numbers instead of 32-bit for certain operations to speed up computation and reduce GPU memory consumption.
B
Freeze the Backbone: Keeping the early layers of the pre-trained model non- trainable (requires grad=False) and only training the final classification head.
C
Increase Batch Size: Utilizing larger batches improves training efficiency by lowering memory usage.
D
Data Prefetching: Using multiple CPU workers (e.g., num workers > 0 in PyTorch) to load and augment data in the background while the GPU processes the current batch.

Show answer

Correct answers

A
Mixed Precision Training (FP16): Using 16-bit floating-point numbers instead of 32-bit for certain operations to speed up computation and reduce GPU memory consumption.
B
Freeze the Backbone: Keeping the early layers of the pre-trained model non- trainable (requires grad=False) and only training the final classification head.
D
Data Prefetching: Using multiple CPU workers (e.g., num workers > 0 in PyTorch) to load and augment data in the background while the GPU processes the current batch.

Question 8 on its own page

Question 9

+5 marksOne or more correct options

When training a deep learning model for a vision competition (such as a ”Cat vs Dog” classifier or a YOLO detector), which of the following statements correctly describe the technical practices for managing hardware and model state?

Select all that apply.

A
To utilize a GPU in PyTorch, you must explicitly move both the model parameters and the input data tensors to the same device (e.g.,.to(’cuda’)).
B
Saving a model using torch.save(model.state dict(), PATH) is generally preferred over saving the entire model object because it only stores the learnable parameters (weights and biases).
C
When resuming training or performing inference, you must first instantiate the model architecture and then load the weights using model.load state dict(torch.load(PATH, map location=device)).
D
The state dict of a model includes the architecture's source code, allowing the weights to be loaded onto a completely different model class without errors.

Show answer

Correct answers

A
To utilize a GPU in PyTorch, you must explicitly move both the model parameters and the input data tensors to the same device (e.g.,.to(’cuda’)).
B
Saving a model using torch.save(model.state dict(), PATH) is generally preferred over saving the entire model object because it only stores the learnable parameters (weights and biases).
C
When resuming training or performing inference, you must first instantiate the model architecture and then load the weights using model.load state dict(torch.load(PATH, map location=device)).

Question 9 on its own page

Question 10

+5 marksOne or more correct options

Consider the following Python code snippet using the Ultralytics YOLOv8 API:
from ultralytics import YOLO model = YOLO("yolov8n.pt")
for name, module in model.model.named_modules(): print(name, type(module))
Based on the architecture of YOLOv8, which of the following statements correctly describe the model?

Select all that apply.

A
The model predicts object locations directly without using predefined anchor boxes.
B
The model uses separate components within its detection head for classification and bounding box regression.
C
The model performs detection in a single forward pass without a separate proposal stage.
D
The model internally generates region proposals using a Region Proposal Network (RPN) before classification.

Show answer

Correct answers

A
The model predicts object locations directly without using predefined anchor boxes.
B
The model uses separate components within its detection head for classification and bounding box regression.
C
The model performs detection in a single forward pass without a separate proposal stage.

Question 10 on its own page

Question 11

+5 marksOne or more correct options

Which of the following loss components are typically used in SRGAN (Super-Resolution GAN) to achieve photo-realistic results?

Select all that apply.

A
Adversarial Loss: To encourage the generator to produce solutions on the natural image manifold.
B
Content (Perceptual) Loss: Based on feature maps from a pre-trained VGG network.
C
Pixel-wise MSE only: To strictly maximize PSNR.
D
Classification Loss: To classify the image into 1000 classes.

Show answer

Correct answers

A
Adversarial Loss: To encourage the generator to produce solutions on the natural image manifold.
B
Content (Perceptual) Loss: Based on feature maps from a pre-trained VGG network.

Question 11 on its own page

Question 12

+5 marksOne or more correct options

Regarding the U-Net architecture, which is widely used for tasks like medical image segmentation and can be adapted for depth estimation:

Select all that apply.

A
It features a symmetric architecture consisting of a contracting path (encoder) and an expansive path (decoder).
B
Skip connections concatenate high-resolution features from the contracting path directly to the up sampled features in the expansive path.
C
The expansive path uses transposed convolutions (or up-convolutions) to increase the spatial resolution of the feature maps.
D
It relies on Global Average Pooling at every layer to ensure that spatial information is discarded in favor of global context.

Show answer

Correct answers

A
It features a symmetric architecture consisting of a contracting path (encoder) and an expansive path (decoder).
B
Skip connections concatenate high-resolution features from the contracting path directly to the up sampled features in the expansive path.
C
The expansive path uses transposed convolutions (or up-convolutions) to increase the spatial resolution of the feature maps.

Question 12 on its own page

Question 13

+5 marksOne or more correct options

Consider the following code:
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True) images = [torch.randn(3, 600, 800)] outputs = model(images)
Unlike YOLO-based models, the number of bounding boxes in outputs[0]['boxes'] varies for each image.
What architectural characteristic of the model explains this behavior?

Select all that apply.

A
The model generates a variable number of region proposals based on the image content.
B
The model filters region proposals using confidence scores and non maximum suppression.
C
The model uses a fixed grid structure to predict bounding boxes.
D
The model predicts a constant number of bounding boxes for every image.

Show answer

Correct answers

A
The model generates a variable number of region proposals based on the image content.
B
The model filters region proposals using confidence scores and non maximum suppression.

Question 13 on its own page

Question 14

+5 marksOne or more correct options

Which of the following statements correctly describe the characteristics and motivations of the Inception (GoogLeNet) architecture?

Select all that apply.

A
It uses Inception modules that apply multiple filter sizes (1×1, 3×3, 5×5) in parallel to capture multi-scale features.
B
It utilizes 1×1 convolutions as bottleneck layers to reduce dimensionality before computationally expensive operations.
C
It incorporates Auxiliary Classifiers in intermediate layers to inject additional gradient signal and combat the vanishing gradient problem.
D
It relies exclusively on stacking 3 × 3 filters to achieve its receptive field, similar to the VGGNet design philosophy.

Show answer

Correct answers

A
It uses Inception modules that apply multiple filter sizes (1×1, 3×3, 5×5) in parallel to capture multi-scale features.
B
It utilizes 1×1 convolutions as bottleneck layers to reduce dimensionality before computationally expensive operations.
C
It incorporates Auxiliary Classifiers in intermediate layers to inject additional gradient signal and combat the vanishing gradient problem.

Question 14 on its own page

Question 15

+5 marksOne or more correct options

In the context of Convolutional Neural Networks (CNNs), what is the primary mechanism that allows the network to handle inputs of varying spatial sizes while maintaining a fixed number of parameters, a characteristic that distinguishes them from standard Multi-Layer Perceptrons (MLPs)?

Select all that apply.

A
Neurons are connected only to a local region of the input.
B
Max Pooling: Reducing the spatial resolution to a single pixel before the first hidden layer.
C
Parameter Sharing: Using the same set of weights (filters) across different spatial locations of the input.
D
Fully Connected Layers: Replacing all convolutional layers with dense layers to increase the receptive field.

Show answer

Correct answers

A
Neurons are connected only to a local region of the input.
C
Parameter Sharing: Using the same set of weights (filters) across different spatial locations of the input.

Question 15 on its own page

Question 16

+5 marksOne correct option

The first layer of the VGG16 model is a Conv2d with kernel = 3, stride = 1, padding = 1.
Consider the following Python code snippet using PyTorch and a pre-trained VGG16 model to process an input image:
import torch import torch.nn as nn from torchvision import models
model = models.vgg16(pretrained=True) feature_extractor = model.features[0] input_image = torch.randn(1, 3, 224, 224) output = feature_extractor(input_image) print(output.shape)
What will be the shape of the output tensor printed by this code?

A
torch.Size([1, 3, 224, 224])
B
torch.Size([1, 64, 112, 112])
C
torch.Size([1, 64, 224, 224])
D
torch.Size([1, 64, 222, 222])

Show answer

Correct answer

C
torch.Size([1, 64, 224, 224])

Question 16 on its own page

Question 17

+5 marksOne correct option

In the ResNet (Residual Network) architecture, 1 × 1 convolutions are frequently used in the ”bottleneck” building block. Beyond dimensionality reduction, what is an additional benefit of using these layers compared to a standard building block?

A
They allow the network to increase the spatial resolution of feature maps to recover lost details.
B
They allow for the addition of non-linearity (via activation functions) without increasing the receptive field or computational cost excessively.
C
They are used to perform Max Pooling operations within the residual mapping.
D
They eliminate the need for skip connections by allowing gradients to flow through the 1 × 1 filters instead.

Show answer

Correct answer

B
They allow for the addition of non-linearity (via activation functions) without increasing the receptive field or computational cost excessively.

Question 17 on its own page

Question 18

+5 marksOne correct option

In the YOLO (You Only Look Once) framework, if multiple bounding boxes are predicted by a single grid cell for the same object, how does the algorithm determine which specific bounding box is responsible for that prediction during the training process?
It is observed that the number of bounding boxes printed varies for different input images.
Which of the following statements best explains this behavior?

A
All predicted boxes for that cell are responsible and updated simultaneously.
B
The bounding box that has the highest class probability score.
C
The bounding box that has the highest Intersection over Union (IoU) with the ground truth.
D
The bounding box that is physically closest to the center of the image.

Show answer

Correct answer

C
The bounding box that has the highest Intersection over Union (IoU) with the ground truth.

Question 18 on its own page

Question 19

+5 marksOne correct option

Consider the following code snippet using a YOLO-based object detection model:
from ultralytics import YOLO
model = YOLO("yolov8n.pt") results = model("image.jpg") boxes = results[0].boxes.xyxy
print(boxes.shape)
It is observed that the number of bounding boxes printed varies for different input images.
Which of the following statements best explains this behavior?

A
The model divides the image into a fixed grid and always predicts a fixed number of boxes.
B
The model filters predictions based on confidence scores and non maximum suppression, resulting in a variable number of final detections.
C
The model uses a Region Proposal Network (RPN) to generate candidate regions before prediction.
D
The number of bounding boxes is fixed by the number of classes in the dataset.

Show answer

Correct answer

B
The model filters predictions based on confidence scores and non maximum suppression, resulting in a variable number of final detections.

Question 19 on its own page

Question 20

+5 marksOne correct option

In the SRResNet and SRGAN architectures for image super-resolution, which specialized layer is used to upsample the feature maps by rearranging elements from the channel dimension into the spatial dimension?

A
The Max Pooling layer with a stride of 2.
B
The Bilinear Interpolation layer for smooth scaling.
C
The Pixel Shuffle (Sub-pixel Convolution) layer.
D
The Global Average Pooling layer to reduce spatial dimensions.

Show answer

Correct answer

C
The Pixel Shuffle (Sub-pixel Convolution) layer.

Question 20 on its own page

Question 21

+5 marksOne correct option

In the Fast R-CNN architecture, which component is responsible for extracting a fixed-size feature vector from a shared convolutional feature map for each region proposal?

A
The Region Proposal Network (RPN) which generates anchors.
B
The Region of Interest (RoI) Pooling layer.
C
A stack of three 3 × 3 convolutional layers.
D
The Global Average Pooling layer used in GoogLeNet.

Show answer

Correct answer

B
The Region of Interest (RoI) Pooling layer.

Question 21 on its own page

Question 22

+5 marksOne correct option

In the GoogLeNet (Inception v1) architecture, 1 × 1 convolutions (bottleneck layers) are applied before larger 3 × 3 and 5 × 5 convolutions. What is the primary motivation for including these 1 × 1 filters within an Inception module?

A
Transposed Convolution.
B
To reduce the dimensionality (depth) of feature maps to manage computational complexity.
C
To replace the need for skip connections and solve the degradation problem.
D
To strictly enforce a mean of 0 and variance of 1 across the channel dimension.

Show answer

Correct answer

B
To reduce the dimensionality (depth) of feature maps to manage computational complexity.

Question 22 on its own page