What I studied today:

  1. AutoGrad & Optimizer
    • nn.Parameter
    • Inherits tensor object
    • If an attribute of nn.Module, required_grad = True
    • Usually inside the layers (not usually delcared independently) - Backward Process from the Scratch
    • optimizer automatically does the job
    • Can get deriviative manually
  2. Dataset & DataLoader
    • ToTensor(): to numerical format
    • transforms: preprocessing
    • Dataset: object that decides how to load a single data
    • DataLoader: providing data to the model via batch, suffle etc.
  3. ViT Paper Review