Machine Learning Generative Models Question 1 - Generative models In the following exercise, you wi ...

---

Machine Learning Generative Models Question 1 - Generative models In the following exercise, you will deal with a generative learning problem, precisely, VAE and GAN. You should write your training code and meet the following constraints. In this exercise, you will create a generative model: choose VAE or GAN implement and train your model: The decoder/generator should get as input a vector from the latent space and produce an image. For convolutions that upscale the input’s spatial size (for the decoder/generator), use nn.ConvTranspose2d . Output visualization: Generate images from your model and visualize its latent spaces. You can compare different architectures for this purpose (e.g., low/high dimension of the latent spaces, etc.). Dataset: You will use the 102 Category Flower Dataset dataset Images can be resized for efficiency but not smaller than 64x64. https://www.robots.ox.ac.uk/~vgg/data/flowers/102/ You should provide: Code (python file) able to reproduce your results. The trained network with trained weights (.pkl file). If the model size is less than 500MB, you should submit it on Moodle. Otherwise, upload it to your Google-Drive. A function called "reproduce_hw4()". This function should be able to reproduce the results that you reported. Discussion: Discuss your results. You should provide the following: Model architecture description and illustration, training procedure (hyperparameters, optimization details, etc.). Training convergence plots as a function of training time: GAN: discriminator and generator losses VAE: reconstruction loss, and KL divergence. Summary of your attempts and conclusions. Your conclusions and explanations should be based on the actual results you received during your attempts. Include 1-2 pages of visualizations (the images your model produces). The discussion should be typed. Hand-written submissions won’t be accepted. submission You should submit a ZIP (not RAR!) file containing: Code - as many files as you need (one of them should be “main.py,” which will include the running process). One pdf file (discussion of Q1). The .pkl file (If the file is too big for the Moodle, upload it to your Google-Drive and copy the link to your pdf report). Run ‘pip freeze > requirements.txt’ and attach it to your submission. Using chatgpt is allowed, a reference from another year is provided for the same question but on a different data set.