GAN Tutorial

Arvin Liu @ AISS 2020

url: slides.com/arvinliu/aiss_gan

DL Review

General Backbone in DL

Data
(Input)

Model
(Function)

Loss
Function

Optimizer

Output
(Output)

Expected
(Answer)

  • We'll not mention model &
    optimizer.

GAN Review

GAN's GOAL

Generate a new fake data

i.e., pictures, music, voice, etc.

Generative Adversarial Network

Data
(Input)

Model
(Function)

Output
(Output)

Expected
(Answer)

???

hyper-params

real data

L

---

???

random vector (z)

First ???

Model / 

Generator

Skill Point -G-> Character

???

image resource: 李宏毅老師的投影片

Given skill points to generate something.

Generative Adversarial Network

Data
(Input)

Model
(Function)

Output
(Output)

Expected
(Answer)

random

hyper-params

real data

---

Second ??? - How 2 Measure?

Generative Adversarial Network

???

Model

Model /
Discriminator

How real is the data

Data
(Input)

Score
(Output)

L

???

Generative Adversarial Network

Model /
Discriminator

(How real is the data,
[0, 1])

Data
(Input)

Score
(Output)

Model /
Discriminator

0.01

0.5

0.9

GAN Framework

Generator
(G)

Generated

Data

Real
Data

random vector (z)

Discriminator
(D)

Score

Loss + Optimizer(G)

Loss+
Optimizer(D)

GAN Intuition (D)

Generator
(G)

Generated

Data

Real
Data

random vector (z)

Discriminator
(D)

Score

Loss+

Optimizer(G)

Loss+

Optimizer(D)

D's target: beat G

GAN Intuition (G)

Generator
(G)

Generated

Data

Real
Data

random vector (z)

Discrimintaor
(D)

Score

Loss+

Optimizer(G)

Loss+

Optimizer(D)

G's target: fool D

GAN vs RL

RL /GAN by DL framework

Data
(Input)

Model
(Function)

Output
(Output)

Environment

Action

Reward

Loss Function

Agent

Optimizer

Random vector

Generator

Fake Data

Discriminator's score

  • discriminator's score in GAN ~= environment's reward in RL.
  • GAN need to update discriminator.

Hand-by-hand GAN

colab link: reurl.cc/VXapby

(If cannot open successfully, use incognito mode to open it.)

Task

Generate MNIST Data

image shape: (28, 28) gray scale

GAN Framework

Generator
(G)

Generated

Data

Real
Data

random vector (z)

Discrimintaor
(D)

Score

Loss+Opt(G)

Loss+Opt(D)

Data Pre-process (cell 1) 

from torchvision import transforms
from torchvision.datasets import MNIST
from torch.utils.data import DataLoader

transform = transforms.Compose([
  transforms.ToTensor(),
  transforms.Normalize(mean=(0.5,), std=(0.5,))])

mnist = MNIST(root='./data/', train=True, transform=transform, download=True)
data_loader = DataLoader(dataset=mnist, batch_size=32, shuffle=True)
  1. torchvision package has MNIST dataset.
  2. Use dataloader to get batched data.
  3. Use transform to cast PIL images to tensor 

Generator & Discriminator (cell 2-1) 

import torch
import torch.nn as nn

# Discriminator
D = nn.Sequential(
  nn.Linear(28*28, 256),
  nn.LeakyReLU(0.2),
  nn.Linear(256, 256),
  nn.LeakyReLU(0.2),
  nn.Linear(256, 1),
  nn.Sigmoid()).cuda()

# Generator 
G = nn.Sequential(
  nn.Linear(64, 256),
  nn.ReLU(),
  nn.Linear(256, 256),
  nn.ReLU(),
  nn.Linear(256, 28*28),
  nn.Tanh()).cuda()

ReLU

Leaky ReLU

tanh

sigmoid

value in (-1, 1)

value in (0, 1)

Note: You can try CNN!

Optimizer (cell 2-2) 

# Binary cross entropy loss and optimizer
criterion = nn.BCELoss()
d_optimizer = torch.optim.Adam(D.parameters(), lr=0.0002)
g_optimizer = torch.optim.Adam(G.parameters(), lr=0.0002)
  1. Use BCELoss as loss function.
  2. Don't ask me why I use Adam optimizer.

Start Training! (cell 3-1) 

for epoch in range(100):
  for images, _ in data_loader:
    batch_size = images.shape[0]
    images = images.view(batch_size, 784).cuda()
    real_labels = torch.ones(batch_size, 1).cuda()
    fake_labels = torch.zeros(batch_size, 1).cuda()
	
    train_discriminator()
    train_generator()
  1. tensor.view([the shape you wish])
  2. real_labels: ones_tensor, fake_labels: zeros_tensor.
    These two are later used as input for BCE Loss.
  3. train your discriminator.
  4. train your generator.
  5. repeat all of steps above.

Start Training! (G) (cell 3-2) 

def train_generator():
    z = torch.randn(batch_size, 64).cuda()
    fake_image = G(z)
    g_loss = criterion(D(fake_image), real_labels)

    g_optimizer.zero_grad()
    g_loss.backward()
    g_optimizer.step()

Generator
(G)

Generated

Data

random vector (z)

Discrimintaor
(D)

Score

Optimizer(G)

  1. torch.randn([shape you want]) : generate random
    matrix from normal distribution(0, 1).

loss

Start Training! (D) (cell 3-3) 

def train_discriminator():
  z = torch.randn(batch_size, 64).cuda()
  d_loss_fake = criterion(D(G(z)), fake_labels)  
  d_loss_real = criterion(D(images), real_labels)
  d_loss = d_loss_real + d_loss_fake

  d_optimizer.zero_grad()
  d_loss.backward()
  d_optimizer.step()

Generated

Data

Real
Data

Discrimintaor
(D)

Score

Optimizer(D)

loss

Result

More?

I think it's useless...

conditional GAN

Generator
(G)

Generated

Data

random vector (z)

condition
(c)

Voice

Generated Data

Condition

安安

TTS (Text2Speech)

Style Transfer

Q&A?

Made with Slides.com