Relearning Deep Learning with PyTorch - Understanding the Principles and Implementing Them Yourself

This course consists of 47 lessons, totaling approximately 39 hours, in which you build deep learning from the ground up with PyTorch. There are many deep learning courses, but most fall into one of two categories. Either they skim only the basics of using libraries, leaving you unsure why the code is written the way it is, or they start with equations and leave you giving up halfway through. This course takes a different approach. First, you create a result; then you dig into what happened under the hood. The course begins with pretrained models. You will run image classification, try inpainting with Stable Diffusion, and even practice image captioning with BLIP. Because you produce tangible results from the very first lesson, you gain the motivation to learn the theory that follows. Next comes the core of this course. Tensors are covered over seven lessons. In addition to indexing and broadcasting, you will explore memory structures such as storage and strides, memory contiguity management, GPU operations, and conversions between tensors and NumPy arrays. It is rare to find a Korean-language course that covers tensors in this much depth. Once you understand this part, you will read all subsequent code differently. You will then learn preprocessing that converts real-world data into tensors. The course covers different data types, including image color channels, tabular data and one-hot encoding, time series, and text embeddings. Moving on to the principles of training, you will implement loss functions and gradient descent yourself using numerical differentiation, and then see how Autograd automates the process. Only those who have built it themselves can truly appreciate the convenience of automatic differentiation. The second half focuses on neural network architectures. You will build deep neural networks with nn.Module and classify CIFAR-10, encounter the limitations of fully connected neural networks, and naturally arrive at why CNNs are necessary. Finally, you will move on to text generation, covering everything from bigram models to embeddings and MLP-based generative models, and then the principles of Attention’s Query, Key, and Value. If you know how to use Python and want to properly understand deep learning, what you will take away from this course is not example code, but the ability to read and understand the structure of any model you encounter.