models.py

"""
Model architectures for the SURPRISE backend.

These match exactly what was trained in the Colab notebook
(lewm_starter.ipynb), so the saved checkpoint will load cleanly.
"""

import torch
import torch.nn as nn
from einops import rearrange


class TinyEncoder(nn.Module):
    """
    Small CNN encoder. Maps (B, 3, 64, 64) → (B, embed_dim).
    Mirrors the Colab definition exactly — do not modify without retraining.
    """

    def __init__(self, embed_dim: int = 64):
        super().__init__()
        self.embed_dim = embed_dim
        self.net = nn.Sequential(
            nn.Conv2d(3, 32, kernel_size=4, stride=2, padding=1),   # 64 → 32
            nn.GELU(),
            nn.Conv2d(32, 64, kernel_size=4, stride=2, padding=1),  # 32 → 16
            nn.GELU(),
            nn.Conv2d(64, 128, kernel_size=4, stride=2, padding=1), # 16 → 8
            nn.GELU(),
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
            nn.Linear(128, embed_dim),
        )
        self.norm = nn.BatchNorm1d(embed_dim, affine=False)

    def forward(self, x):
        z = self.net(x)
        z = self.norm(z)
        return z


class TinyPredictor(nn.Module):
    """
    Predicts the next embedding from the current one.
    No actions (passive video), no temporal context beyond t-1.
    """

    def __init__(self, embed_dim: int = 64, hidden_dim: int = 128):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(embed_dim, hidden_dim),
            nn.GELU(),
            nn.Linear(hidden_dim, hidden_dim),
            nn.GELU(),
            nn.Linear(hidden_dim, embed_dim),
        )

    def forward(self, z_t):
        return self.net(z_t)