import torch
import torch.nn as nn
from transformers import ViTMAEModel, ViTMAEConfig

class VibeMAE(nn.Module):
    def __init__(self, pretrained=False):  # changed default to False
        super().__init__()
        if pretrained:
            self.mae = ViTMAEModel.from_pretrained("facebook/vit-mae-base")
        else:
            # Just set up the architecture, weights loaded from .pth
            config = ViTMAEConfig()
            self.mae = ViTMAEModel(config)
        
        self.vibe_head = nn.Sequential(
            nn.Linear(768, 256),
            nn.ReLU(),
            nn.Linear(256, 3),
            nn.Sigmoid()
        )

    def forward(self, x):
        outputs = self.mae(x)
        latent = outputs.last_hidden_state[:, 0, :]
        return self.vibe_head(latent)