Lesson 23: Deep Learning Foundations to Stable Diffusion

(All lesson resources are available at course.fast.ai.) In this lesson, we work with Tiny Imagenet to create a super-resolution U-Net model, discussing dataset creation, preprocessing, and data augmentation. The goal of super-resolution is to scale up a low-resolution image to a higher resolution. We train the model using AdamW optimizer and mixed precision, achieving an accuracy of nearly 60%. We also explore the potential for improvement by examining the results of other models on Tiny Imagenet from the Papers with Code website.
We discuss the limitations of using a convolutional neural network for image super-resolution and introduce the concept of U-net, a more efficient architecture for this task. We implement perceptual loss, which involves comparing the features of the output image and the target image at an intermediate layer of a pre-trained classifier model. After training the U-net model with the new loss function, the output images are less blurry and more similar to the target images.
Finally, we discuss the challenges of comparing different models and their outputs. We demonstrate how perceptual loss has improved the results significantly, but also note that there isn't a clear metric to use for comparison. We then move on to gradually unfreezing pre-trained networks, a favorite trick at fast.ai. We copy the weights from the pre-trained model into our model and train it for one epoch with frozen weights for the down path. This results in a significant improvement in loss.

[@satirthapaulshyam7769]

1:06:00 modulelist is just like sequential but it doesnt autofwd we need to define the fwd method

[@thehigheststateofsalad]

I see you are trying to help. However, can you combine all your comments into one and delete the others? It's just messy.

[@satirthapaulshyam7769]

38:00 no need of augmentation calback

[@mattzucca4102]

😮

[@satirthapaulshyam7769]

1:09:00 weight initialization of unet

[@satirthapaulshyam7769]

49:00 superres

[@satirthapaulshyam7769]

1:28:00 replacing some layers of unet with pre trained classifier

[@satirthapaulshyam7769]

57:00 why in superres squeez and again unsquezz why not we use stride1

[@satirthapaulshyam7769]

34:00 trivial augmentations

[@satirthapaulshyam7769]

59:00

[@satirthapaulshyam7769]

1:22:00 comb loss