Hi, my name is Jonathan Borg

I'm a

We looked into SwinIR: Image Restoration Using Swin Transformer, a model for image restoration based on the Swin Transformer. Their model combines a traditional Convolutional approach with Swin Transformer layers, which is a residual transformer approach to image processing.

The contribution of this blog post is twofold. First of all, we explained the technical details of the SwinIR paper in our own words, providing ample detail to understand the authors’ contribution and algorithm. Secondly, we explored modifying the architecture used in the paper to allow it to run using reduced resources, 1 RTX 3070 instead of the 8 RTX 2080TIs used in the paper. Our primary focus is to investigate whether the model could be reduced and achieve results comparable to the original model.

Python PyTorch scikit-learn Matplotlib Transformers

We created a Generative Adversarial Network (GAN) capable of generating animated faces based on the data set provided during training (specific to a particular drawing style — cartoon or anime).

GANs are composed of two primary components: The generator, a neural network tasked with generating images appearing to be as ‘real’ as possible. The discriminator is another neural network that determines if the supplied image is real or fake (i.e., generated by the other network). These two networks compete against each other to form a zero-sum game, where one agent’s gain is another agent’s loss.

Following the advancements of GANs in several fields, we decided to implement a model capable of generating animated faces. Apart from the model generation, ample tests, such as different model architectures and different datasets with diverse animation styles, were conducted to develop a robust and stable model.

Python PyTorch NumPy SciPy Matplotlib GAN