Q1. Use CIFAR-10 dataset, [80 marks + 20 Bonus] a. Implement a 6-layer CNN network (Choose your own architecture) to perform classification, use the same training and testing data split as given in the dataset, Report the following

i. Accuracy on test data by varying the Optimization Techniques

1. Vanilla SGD [5 marks] 2. SGD with momentum. [5 marks] 3. Adam [5 marks]

ii. Accuracy on test data by varying the Normalization Techniques

1. Dropout [10 marks] 2. Batch Normalization. [10 marks]

iii. Accuracy on test data by varying activation function in between CNN layers,

1. Identity. [5 marks] 2. Sigmoid. [5 marks] 3. ReLU. [5 marks] 4. Tanh. [5 marks]

iv. Accuracy on test data by varying the loss functions

1. Cross-Entropy Loss [5 marks] 2. MSELoss [5 marks] 3. L1 Loss

Find the best configuration for your CNN (a combination of Optimization Technique, Normalization technique, activation function and Loss function), support your claim. [30 marks] Bonus:- Train above model for Image Embeddings with triplet loss, use these embeddings and train an SVM, report accuracy and compare it with the best model obtained in the previous exercise [20 marks]

Q2. [Noise to Image Generation] Train a DCGAN to generate images from noise. Use the MNIST database to learn the GAN network.

[Discriminator in DCGAN:- i. if roll no. % 2 == 0: use VGG16 as a discriminator. ii. if roll no. % 2 == 1: use Resnet 18 as a discriminator.]

Perform the following tasks: [35 marks for training GAN] b. Uniformly generate ten noise vectors that act as latent representation vectors, and generate the images for these noise vectors, and visualize them at [5 + 5 + 5 marks]

i. After the first epoch.

ii. After n/2 th epoch.

iii. After your last epoch. (say n epochs in total) and comment on the image interpretation at (i), (ii) and (iii), can you identify the images? Pass these images generated at each epoch to the model youve trained in (Q1) and report their predictions with confidence scores. Comment on them. [5 + 15 marks]

c. Plot generator and discriminator losses for all the iterations. [One iteration = forward pass of a mini-batch] [10 marks]

d. [Bonus] Do we have control of what class image the generator will generate, given a noise vector? Suppose, we are interested in generating only 4 images, will the GAN you trained in (a) can do that? Explain why. If not, modify the GAN trained above to do this. [10 + 40 marks]

Q3. [Image to Image Translation] Your task is to train a GAN that generates maps given satellite images. [Dataset Link]. [40 marks for training GAN] a. Report visualized image translations [satellite image -ground truth map - map generated]. [at least 10 examples] [10 marks] b. Plot generator and discriminator losses for all the iterations. [One iteration = forward pass of a mini-batch]. [10 marks] c. Use the pre-trained sat2map generator model from here [Pytorch - Link, TensorFlow - Link] and generate images for the same ten examples in (a). Compare and comment on the images generated with a pre-trained generator with your trained generator in (a). Support your claim with SSIM. [15 + 5 marks]

3. Choose any GAN paper from top venues (with h5-index > 50, for example, CVPR, ECCV, ICCV, ICML, etc.) in the last two years [2019-20] [Reference:- https://github.com/hindupuravinash/the-gan-zoo], reproduce the results in the paper. You can use the authors code (mostly provided via GitHub) or any other reimplementation available on the internet. [Please cite the source]. You should be able to understand and explain the code that you are using. [80 marks]