Assignment 4 - trustagi # Part 1 ## Content Reconstruction These are the reconstructed images for different layers. Earlier layers render images more closely resembling the input image. From conv4 onwards, we see the model preserve structure, but lose fine tuned details.         ## Content Loss for different random noise I wanted to use a layer that removed sytle, but preserved the main content of the input image. I chose conv_7. These are 2 images with random initialized loss.   # Part 2 - Texture Synthesis To preserve both high level and low level features in texture synthesis, i chose a combination of conv layers to see their effect on texture synthesis loss.     I chose conv1 to conv5 combination as it leads to the best texture, preserving both color and shapes. Here are 2 images sampled from random noise.   # Part 3 - Style Transfer For Style Transfer, I used the following parameters Content Weight = 1 Style Weight = 100000 Steps = 500 Style Layers = Conv 1 to Conv 8 Content Layers = Conv 4 These are some sample outputs with content image as the input image.      Let's also visualize this using random noise as the input image.    After running these experiments, we see that both of the image inputs close to the same runtime, but the one with random noise had a slightly higher training time. Images with the content image produces significantly better results, as the texture is more holistically infused in the image. Furthermore, the loss decreases faster with input image as the content image. Here are some images i really like to apply style transfer on. I chose images of my pets.   # Bells and Whistles.