source/model.py

from __future__ import annotations

import torch
from torch import nn
from torch.nn import functional as F


class VectorQuantizedAutoencoder(nn.Module):
    def __init__(
        self,
        codebook_size: int = 32,
        embedding_dimensions: int = 16,
    ) -> None:
        super().__init__()
        self.codebook_size = codebook_size
        self.embedding_dimensions = embedding_dimensions
        self.encoder = nn.Sequential(
            nn.Conv2d(1, 32, kernel_size=4, stride=2, padding=1),
            nn.SiLU(),
            nn.Conv2d(32, embedding_dimensions, kernel_size=3, padding=1),
        )
        self.codebook = nn.Embedding(codebook_size, embedding_dimensions)
        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(
                embedding_dimensions,
                32,
                kernel_size=4,
                stride=2,
                padding=1,
            ),
            nn.SiLU(),
            nn.Conv2d(32, 1, kernel_size=3, padding=1),
            nn.Sigmoid(),
        )
        nn.init.uniform_(
            self.codebook.weight,
            -1 / codebook_size,
            1 / codebook_size,
        )

    def encode(self, pixels: torch.Tensor) -> torch.Tensor:
        return self.encoder(pixels)

    def quantize(
        self,
        encoded: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        channels_last = encoded.permute(0, 2, 3, 1).contiguous()
        flat = channels_last.reshape(-1, self.embedding_dimensions)
        distances = (
            flat.square().sum(1, keepdim=True)
            + self.codebook.weight.square().sum(1)
            - 2 * flat @ self.codebook.weight.t()
        )
        indices = distances.argmin(dim=1)
        quantized = self.codebook(indices).reshape(channels_last.shape)
        quantized = quantized.permute(0, 3, 1, 2).contiguous()
        straight_through = encoded + (quantized - encoded).detach()
        return straight_through, indices.reshape(encoded.shape[0], 4, 4)

    def decode(self, latent: torch.Tensor) -> torch.Tensor:
        return self.decoder(latent)

    def decode_indices(self, indices: torch.Tensor) -> torch.Tensor:
        quantized = self.codebook(indices)
        quantized = quantized.permute(0, 3, 1, 2).contiguous()
        return self.decode(quantized)

    def forward(
        self,
        pixels: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        encoded = self.encode(pixels)
        quantized_st, indices = self.quantize(encoded)
        quantized = self.codebook(indices).permute(0, 3, 1, 2).contiguous()
        reconstruction = self.decode(quantized_st)
        codebook_loss = F.mse_loss(quantized, encoded.detach())
        commitment_loss = F.mse_loss(encoded, quantized.detach())
        return reconstruction, indices, codebook_loss, commitment_loss


class ConditionalCodePrior(nn.Module):
    def __init__(
        self,
        codebook_size: int = 32,
        token_dimensions: int = 32,
        hidden_dimensions: int = 64,
    ) -> None:
        super().__init__()
        self.codebook_size = codebook_size
        self.start_token = codebook_size
        self.token_embedding = nn.Embedding(codebook_size + 1, token_dimensions)
        self.label_embedding = nn.Embedding(10, 16)
        self.position_embedding = nn.Embedding(16, 16)
        self.recurrent = nn.GRU(
            token_dimensions + 32,
            hidden_dimensions,
            batch_first=True,
        )
        self.output = nn.Linear(hidden_dimensions, codebook_size)

    def forward(self, input_tokens: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
        batch, length = input_tokens.shape
        positions = torch.arange(length, device=input_tokens.device)
        token_features = self.token_embedding(input_tokens)
        condition = self.label_embedding(labels)[:, None, :].expand(batch, length, -1)
        position = self.position_embedding(positions)[None, :, :].expand(batch, -1, -1)
        hidden, _ = self.recurrent(
            torch.cat([token_features, condition, position], dim=2)
        )
        return self.output(hidden)

    @torch.inference_mode()
    def generate(
        self,
        labels: torch.Tensor,
        *,
        seed: int,
        temperature: float = 1.0,
    ) -> torch.Tensor:
        generator = torch.Generator(device=labels.device).manual_seed(seed)
        sequence = torch.full(
            (len(labels), 1),
            self.start_token,
            dtype=torch.long,
            device=labels.device,
        )
        for _ in range(16):
            logits = self(sequence, labels)[:, -1]
            if temperature <= 0.05:
                token = logits.argmax(dim=1, keepdim=True)
            else:
                probabilities = torch.softmax(logits / temperature, dim=1)
                token = torch.multinomial(
                    probabilities,
                    1,
                    generator=generator,
                )
            sequence = torch.cat([sequence, token], dim=1)
        return sequence[:, 1:].reshape(-1, 4, 4)


class TinyVisionJudge(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(1, 8, kernel_size=3, padding=1),
            nn.GELU(),
            nn.Conv2d(8, 8, kernel_size=3, padding=1, groups=8),
            nn.GELU(),
            nn.Conv2d(8, 12, kernel_size=1),
            nn.GELU(),
            nn.MaxPool2d(2),
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(12 * 4 * 4, 10),
        )

    def forward(self, pixels: torch.Tensor) -> torch.Tensor:
        return self.classifier(self.features(pixels))


def parameter_count(model: nn.Module) -> int:
    return sum(parameter.numel() for parameter in model.parameters())