Scribd
Upload
42 views1 page

Indian Institute of Technology Bombay September 29, 2023 EE782 Advanced Topics in Machine Learning Assignment 2: Metric Learning and Generative AI

The document outlines an assignment for an advanced machine learning course. It involves two parts: [1] training a Siamese neural network on labeled face images to determine if images are of the same person, and [2] training a generative adversarial network (GAN) to generate new face images. For part two, students can optionally implement a conditional GAN that generates new images of the same person. The assignment requires commenting code, submitting work, and creating an explanatory video.

Uploaded by

dhruvkumar.moodi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
42 views1 page

Indian Institute of Technology Bombay September 29, 2023 EE782 Advanced Topics in Machine Learning Assignment 2: Metric Learning and Generative AI

The document outlines an assignment for an advanced machine learning course. It involves two parts: [1] training a Siamese neural network on labeled face images to determine if images are of the same person, and [2] training a generative adversarial network (GAN) to generate new face images. For part two, students can optionally implement a conditional GAN that generates new images of the same person. The assignment requires commenting code, submitting work, and creating an explanatory video.

Uploaded by

dhruvkumar.moodi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Indian Institute of Technology Bombay; September 29, 2023

EE782 Advanced Topics in Machine Learning

Assignment 2: Metric Learning and Generative AI

Instructions:
 Assignment has to be done individually, in Google CoLab using python and pytorch
 Write copious comments using text cells and in the code cells to explain your intent and
demonstrate understanding
 End with a cell that lists sources of starter code (e.g., github repo links, or chatGPT) for all
modules
 Submit the (1) ipynb file, (2) its py version, (3) link to a 10 minute video explaining the code, your
intent, and your observations and decisions on Moodle
 CODE WITHOUT COMMENTS FOR EACH LINE OF CODE WILL NOT BE GRADED
 VIDEO WITHOUT VIEW PERMISSIONS WILL NOT BE GRADED. Test your video by sending link to a
friend
 Deadline is Oct 15, 2023, 11:59pm

Objectives:

1. Download labeled faces in the wild (LFW) dataset: http://vis-www.cs.umass.edu/lfw/


2. Get the number of persons who have more than one image

Part-A

3. Split into training, validation, and testing by person (not by image)


4. Start with a network that is pre-trained on ImageNet and appropriate for your computational
resources
5. Appropriately crop and resize the images based on your computational resources
6. Setup and train a Siamese network to predict whether a pair of images are of the same person
or not using a metric learning scheme (cosine similarity or Euclidean distance; paired with cross-
entropy or hinge loss with a margin), and validate on the validation split [3 for basic set up]
a) Use image augmentation [1 for including this]
b) Experiment with a regularization technique and its hyper-parameter [1 for including this]
7. Experiment with at least two learning rate schedulers and comment on what works, doesn’t
work, and potential reason why [1]
8. Experiment with at least two different optimizers and comment on what works, doesn’t work,
and potential reason why [1]
9. Test on the test split [1]
10. Gather a few images of yourself and your friends (with their consent) and check if the model
works well on this data [1]

Part-B

11. Train a generative model for generating face images, using a GAN. The generator takes a
Gaussian noise vector as input, and tries to output a face image, while the discriminator
distinguishes between real and fake face images. See, for example:
https://www.youtube.com/watch?v=_pIMdDWK5sc [2]
12. (Bonus) Modify the GAN to become a conditional GAN, where the condition itself is (features of)
a real face image. The CGAN should generate another image of the same person, and you can
use the Siamese network from Part-A as an additional discriminator for person matching. You
will need to ensure that it does not simply show the same image as the one used for conditional
input. See, for instance, https://www.youtube.com/watch?v=Hp-jWm2SzR8 [2]

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy