ml-theory.md

Variational Autencoders vs GANs

Generative AI Models
├── Explicit Density Models
│   ├── Autoregressive Models
│   │   ├── PixelRNN/PixelCNN
│   │   └── Transformer-based Models
│   ├── Flow-based Models
│   │   ├── Normalizing Flows
│   │   └── Real NVP
│   └── Variational Autoencoders (VAEs)
│       ├── Conditional VAEs
│       └── Beta-VAEs
└── Implicit Density Models
    └── Generative Adversarial Networks (GANs)
        ├── DCGANs
        ├── WGANs
        └── StyleGANs

Convolution

Ouptut Size of Convolution is the integer

output_width = (input_width - kernel_size + 2*padding_size) / stride + 1

• Valid padding means no padding

  padding_size = 0
  

• Same padding preserves the input size

  padding_size can be calculated by solving with output_width = input_width
  padding_size = (input_width * (stride - 1) + kernel_size - stride) / 2
  
  for stride = 1
  padding_size = (kernel_size - 1)/2
  
  padding_size will always be an integer
  
  In tensorflow if padding size is even, the padding is distributed evenly on both sides.
  In tensorflow if padding size is odd, the one extra padding is added to the right-most or bottom-most side.
  

padding visualiztion