jepa.py

"""
Joint Embedding Predictive Architecture (JEPA) for PDE dynamics.

Spatial JEPA: encoder produces spatial feature maps, predictor operates
on spatial features, loss computed on spatial latent representations.
Prevents collapse via VICReg regularization.
"""
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F


# ---------------------------------------------------------------------------
# Building blocks
# ---------------------------------------------------------------------------


class ConvBlock(nn.Module):
    def __init__(self, in_ch, out_ch, stride=1):
        super().__init__()
        self.conv1 = nn.Conv2d(in_ch, out_ch, 3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(out_ch)
        self.conv2 = nn.Conv2d(out_ch, out_ch, 3, padding=1)
        self.bn2 = nn.BatchNorm2d(out_ch)
        self.skip = (
            nn.Sequential(nn.Conv2d(in_ch, out_ch, 1, stride=stride), nn.BatchNorm2d(out_ch))
            if in_ch != out_ch or stride != 1
            else nn.Identity()
        )

    def forward(self, x):
        h = F.gelu(self.bn1(self.conv1(x)))
        h = self.bn2(self.conv2(h))
        return F.gelu(h + self.skip(x))


# ---------------------------------------------------------------------------
# Spatial Encoder (outputs feature maps, not vectors)
# ---------------------------------------------------------------------------


class SpatialEncoder(nn.Module):
    """ResNet-style encoder outputting spatial latent maps.

    Input:  [B, C_in,  H,   W]
    Output: [B, lat_ch, H/8, W/8]
    """

    def __init__(self, in_channels, latent_channels=128, base_ch=32):
        super().__init__()
        self.stem = nn.Sequential(
            nn.Conv2d(in_channels, base_ch, 3, padding=1),
            nn.BatchNorm2d(base_ch),
            nn.GELU(),
        )
        self.layer1 = ConvBlock(base_ch, base_ch * 2, stride=2)        # /2
        self.layer2 = ConvBlock(base_ch * 2, base_ch * 4, stride=2)    # /4
        self.layer3 = ConvBlock(base_ch * 4, latent_channels, stride=2) # /8

    def forward(self, x):
        x = self.stem(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        return x


# ---------------------------------------------------------------------------
# Spatial Predictor (conv-based, operates on feature maps)
# ---------------------------------------------------------------------------


class SpatialPredictor(nn.Module):
    """Lightweight CNN predictor on spatial latent maps.

    Input/Output: [B, lat_ch, H', W']
    """

    def __init__(self, latent_channels=128, hidden_channels=256):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(latent_channels, hidden_channels, 3, padding=1),
            nn.BatchNorm2d(hidden_channels),
            nn.GELU(),
            nn.Conv2d(hidden_channels, hidden_channels, 3, padding=1),
            nn.BatchNorm2d(hidden_channels),
            nn.GELU(),
            nn.Conv2d(hidden_channels, latent_channels, 3, padding=1),
        )

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


# ---------------------------------------------------------------------------
# VICReg-style loss (prevents representation collapse)
# ---------------------------------------------------------------------------


def vicreg_loss(z_pred, z_target, sim_w=25.0, var_w=25.0, cov_w=1.0):
    """VICReg loss on spatial features (flattened to [B, D]).

    Args:
        z_pred:   [B, D] predicted latent.
        z_target: [B, D] target latent (detached).
        sim_w, var_w, cov_w: loss weights.

    Returns:
        total loss, dict of components.
    """
    # Invariance
    sim_loss = F.mse_loss(z_pred, z_target)

    # Variance
    std_p = torch.sqrt(z_pred.var(dim=0) + 1e-4)
    std_t = torch.sqrt(z_target.var(dim=0) + 1e-4)
    var_loss = F.relu(1 - std_p).mean() + F.relu(1 - std_t).mean()

    # Covariance
    B, D = z_pred.shape
    zp = z_pred - z_pred.mean(0)
    zt = z_target - z_target.mean(0)
    cov_p = (zp.T @ zp) / max(B - 1, 1)
    cov_t = (zt.T @ zt) / max(B - 1, 1)
    mask = ~torch.eye(D, device=z_pred.device).bool()
    cov_loss = cov_p[mask].pow(2).sum() / D + cov_t[mask].pow(2).sum() / D

    total = sim_w * sim_loss + var_w * var_loss + cov_w * cov_loss
    return total, {"sim": sim_loss.item(), "var": var_loss.item(), "cov": cov_loss.item()}


# ---------------------------------------------------------------------------
# Full JEPA model
# ---------------------------------------------------------------------------


class JEPA(nn.Module):
    """Spatial JEPA for PDE dynamics prediction.

    Online encoder + predictor learn to predict the target encoder's
    representation of the next frame.  The target encoder is an EMA
    copy of the online encoder.

    Args:
        in_channels: number of input field channels.
        latent_channels: spatial latent feature map channels.
        base_ch: encoder base width.
        pred_hidden: predictor hidden channels.
        ema_decay: starting EMA decay.
    """

    def __init__(
        self,
        in_channels,
        latent_channels=128,
        base_ch=32,
        pred_hidden=256,
        ema_decay=0.996,
    ):
        super().__init__()
        self.online_encoder = SpatialEncoder(in_channels, latent_channels, base_ch)
        self.predictor = SpatialPredictor(latent_channels, pred_hidden)
        self.target_encoder = copy.deepcopy(self.online_encoder)
        self.ema_decay = ema_decay

        # Freeze target
        for p in self.target_encoder.parameters():
            p.requires_grad_(False)

    @torch.no_grad()
    def update_target(self):
        """EMA update of target encoder."""
        for pt, po in zip(self.target_encoder.parameters(), self.online_encoder.parameters()):
            pt.data.lerp_(po.data, 1 - self.ema_decay)

    def set_ema_decay(self, decay):
        """Update EMA decay (e.g. cosine schedule from 0.996 to 1.0)."""
        self.ema_decay = decay

    def forward(self, x_input, x_target):
        """
        Args:
            x_input:  current frame(s) [B, C, H, W]
            x_target: next frame(s)    [B, C, H, W]

        Returns:
            z_pred:   predicted spatial latent [B, lat_ch, H', W']
            z_target: target spatial latent    [B, lat_ch, H', W']
        """
        z_online = self.online_encoder(x_input)
        z_pred = self.predictor(z_online)

        with torch.no_grad():
            z_target = self.target_encoder(x_target)

        return z_pred, z_target

    def compute_loss(self, x_input, x_target):
        """Full forward + loss computation.

        VICReg is computed on channel vectors after spatial averaging
        to keep the covariance matrix small (D = latent_channels).

        Returns:
            loss: scalar.
            metrics: dict.
        """
        z_pred, z_target = self(x_input, x_target)

        # Spatial MSE loss (pixel-level prediction quality)
        spatial_mse = F.mse_loss(z_pred, z_target.detach())

        # VICReg on spatially-averaged channel vectors [B, C]
        zp_avg = z_pred.mean(dim=(-2, -1))    # [B, lat_ch]
        zt_avg = z_target.mean(dim=(-2, -1))  # [B, lat_ch]

        vicreg, vicreg_m = vicreg_loss(zp_avg, zt_avg.detach())

        # Combine: spatial MSE drives prediction, VICReg prevents collapse
        loss = spatial_mse + 0.1 * vicreg
        metrics = {
            "sim": vicreg_m["sim"],
            "var": vicreg_m["var"],
            "cov": vicreg_m["cov"],
            "spatial_mse": spatial_mse.item(),
        }
        return loss, metrics