Generative adversarial networks for text generation

Deep Learning for Text with PyTorch

Shubham Jain

Instructor

GANs and their role in text generation

GANs can generate new content that seems original
- Preserves statistical similarities
Can replicate complex patterns unachievable by RNNs
Can emulate real-world patterns

Orignal Data to Synthetic Data

Structure of a GAN

A GAN has two components:
- Generator: creates fake samples by adding noise
- Discriminator: differentiates between real and generated text data

Structure of GANs

¹ https://www.sciencefocus.com/future-technology/how-do-machine-learning-gans-work/

Building a GAN model in PyTorch: Generator

# Embedding reviews
# Convert reviews to tensors

class Generator(nn.Module):
    def __init__(self):
        super().__init__()

        self.model = nn.Sequential(
            nn.Linear(seq_length, seq_length),
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.model(x)

Building the discriminator network

class Discriminator(nn.Module):
    def __init__(self):
        super().__init__()

        self.model = nn.Sequential(
            nn.Linear(seq_length, 1),
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.model(x)

Initializing networks and loss function

generator = Generator()

discriminator = Discriminator()


criterion = nn.BCELoss()


optimizer_gen = torch.optim.Adam(generator.parameters(), lr=0.001)
optimizer_disc = torch.optim.Adam(discriminator.parameters(), lr=0.001)

Training the discriminator

num_epochs = 50
for epoch in range(num_epochs):

    for real_data in data:
        real_data = real_data.unsqueeze(0)

        noise = torch.rand((1, seq_length))

        disc_real = discriminator(real_data)

        fake_data = generator(noise)
        disc_fake = discriminator(fake_data.detach())

        loss_disc = criterion(disc_real, torch.ones_like(disc_real)) + 
                    criterion(disc_fake, torch.zeros_like(disc_fake))

        optimizer_disc.zero_grad()

        loss_disc.backward()

        optimizer_disc.step()

Training the generator

       # ... (continued from last slide)
        disc_fake = discriminator(fake_data)

        loss_gen = criterion(disc_fake, torch.ones_like(disc_fake))

        optimizer_gen.zero_grad()

        loss_gen.backward()

        optimizer_gen.step()


    if (epoch+1) % 10 == 0:

        print(f"Epoch {epoch+1}/{num_epochs}:\t 
                Generator loss: {loss_gen.item()}\t 
                Discriminator loss: {loss_disc.item()}")

Printing real and generated data

print("\nReal data: ")
print(data[:5])


print("\nGenerated data: ")
for _ in range(5):
    noise = torch.rand((1, seq_length))
    generated_data = generator(noise)
    print(torch.round(generated_data).detach())

GANs: generated synthetic data

Epoch 10/50: Generator loss: 0.8992824673652 Discriminator loss: 1.37682652473
Epoch 20/50: Generator loss: 0.7347183227539 Discriminator loss: 1.390102505683
...
Epoch 50/50: Generator loss: 0.7019854784011 Discriminator loss: 1.3501529693603

Generated data

Real data: 
tensor([[1., 0., 0., 1., 1.], 
        [0., 0., 1., 0., 0.], 
        [1., 0., 1., 1., 1.], 
        [1., 0., 1., 0., 0.], 
        [1., 1., 1., 1., 1.]])

Evaluation metric: correlation matrix

Generated data: 
tensor([[0., 1., 1., 0., 0.]]), 
tensor([[0., 1., 1., 1., 1.]])
tensor([[1, 1., 1., 0., 0.]]), 
tensor([[1., 1., 1., 0., 0.]]) 
tensor([[0., 1., 1., 1., 1.]])

Let's practice!

Deep Learning for Text with PyTorch