import json
import torch
import torch.nn as nn
import os
from pathlib import Path
from typing import Optional, Union, Dict
from huggingface_hub import snapshot_download
import warnings

class ConvVAE(nn.Module):
    def __init__(self, latent_size):
        super(ConvVAE, self).__init__()

        # Encoder
        self.encoder = nn.Sequential(
            nn.Conv2d(3, 64, 3, stride=2, padding=1),  # (batch, 64, 64, 64)
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 128, 3, stride=2, padding=1),  # (batch, 128, 32, 32)
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.Conv2d(128, 256, 3, stride=2, padding=1),  # (batch, 256, 16, 16)
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.Conv2d(256, 512, 3, stride=2, padding=1),  # (batch, 512, 8, 8)
            nn.BatchNorm2d(512),
            nn.ReLU()
        )

        self.fc_mu = nn.Linear(512 * 8 * 8, latent_size)
        self.fc_logvar = nn.Linear(512 * 8 * 8, latent_size)

        self.fc2 = nn.Linear(latent_size, 512 * 8 * 8)

        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(512, 256, 4, stride=2, padding=1),  # (batch, 256, 16, 16)
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.ConvTranspose2d(256, 128, 4, stride=2, padding=1),  # (batch, 128, 32, 32)
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1),  # (batch, 64, 64, 64)
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.ConvTranspose2d(64, 3, 4, stride=2, padding=1),  # (batch, 3, 128, 128)
            nn.Tanh()  
        )

    def forward(self, x):
        mu, logvar = self.encode(x)
        z = self.reparameterize(mu, logvar)
        decoded = self.decode(z)
        return decoded, mu, logvar

    def encode(self, x):
        x = self.encoder(x)
        x = x.view(x.size(0), -1) 
        mu = self.fc_mu(x)
        logvar = self.fc_logvar(x)
        return mu, logvar

    def reparameterize(self, mu, logvar):
        std = torch.exp(0.5 * logvar)
        eps = torch.randn_like(std)
        return mu + eps * std

    def decode(self, z):
        x = self.fc2(z)
        x = x.view(-1, 512, 8, 8)
        decoded = self.decoder(x)
        return decoded

    @classmethod
    def from_pretrained(
        cls,
        model_id: str,
        revision: Optional[str] = None,
        cache_dir: Optional[Union[str, Path]] = None,
        force_download: bool = False,
        proxies: Optional[Dict] = None,
        resume_download: bool = False,
        local_files_only: bool = False,
        token: Union[str, bool, None] = None,
        map_location: str = "cpu",
        strict: bool = False,
        **model_kwargs,
    ):
        """
        Load a pretrained model from a given model ID.
        Args:
            model_id (str): Identifier of the model to load.
            revision (Optional[str]): Specific model revision to use.
            cache_dir (Optional[Union[str, Path]]): Directory to store downloaded models.
            force_download (bool): Force re-download even if the model exists.
            proxies (Optional[Dict]): Proxy configuration for downloads.
            resume_download (bool): Resume interrupted downloads.
            local_files_only (bool): Use only local files, don't download.
            token (Union[str, bool, None]): Token for API authentication.
            map_location (str): Device to map model to. Defaults to "cpu".
            strict (bool): Enforce strict state_dict loading.
            **model_kwargs: Additional keyword arguments for model initialization.
        Returns:
            An instance of the model loaded from the pretrained weights.
        """
        model_dir = Path(model_id)
        if not model_dir.exists():
            model_dir = Path(
                snapshot_download(
                    repo_id=model_id,
                    revision=revision,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    resume_download=resume_download,
                    token=token,
                    local_files_only=local_files_only,
                )
            )

        config_file = model_dir / "config.json"
        with open(config_file, 'r') as f:
            config = json.load(f)

        latent_size = config.get('latent_size')
        if latent_size is None:
            raise ValueError("The configuration file is missing the 'latent_size' key.")

        model = cls(latent_size, **model_kwargs)

        model_file = model_dir / "model_conv_vae_256_epoch_304.pth"
        if not model_file.exists():
            raise FileNotFoundError(f"The model checkpoint '{model_file}' does not exist.")
        
        state_dict = torch.load(model_file, map_location=map_location)

        new_state_dict = {}
        for k, v in state_dict.items():
            if k.startswith('_orig_mod.'):
                new_state_dict[k[len('_orig_mod.'):]] = v
            else:
                new_state_dict[k] = v

        model.load_state_dict(new_state_dict, strict=strict)
        model.to(map_location)

        return model


model = ConvVAE.from_pretrained(
    model_id="BioMike/classical_portrait_vae",
    cache_dir="./model_cache",
    map_location="cpu",
    strict=True).eval()

Training Details

Training Data

The model was trained on the Portrait Dataset

Training Procedure

The model was trained into two steps, in the first the model vgg16 was employed in the perceptual loss, to train our model to extract general features, and the model vggface2 was used to train VAE to decode faces accurately.

Model Card Authors

Mykhailo Shtopko, werent4