Source code for cca_zoo.deep._generative._dvcca

from typing import Iterable

import torch

from .._base import BaseDeep
from .._generative._base import _GenerativeMixin




class DVCCA(BaseDeep, _GenerativeMixin):
    """
    A class used to fit a DVCCA model.

    References
    ----------
    Wang, Weiran, et al. 'Deep variational canonical correlation analysis.' arXiv preprint arXiv:1610.03454 (2016).
    https: // arxiv.org / pdf / 1610.03454.pdf
    https: // github.com / pytorch / examples / blob / master / vae / main.py

    """

    def __init__(
        self,
        latent_dimensions: int,
        encoders=None,
        decoders=None,
        private_encoders: Iterable = None,
        latent_dropout=0,
        img_dim=None,
        recon_loss_type="mse",
        **kwargs,
    ):
        super().__init__(latent_dimensions=latent_dimensions, **kwargs)
        self.img_dim = img_dim
        self.latent_dropout = torch.nn.Dropout(p=latent_dropout)
        self.encoders = torch.nn.ModuleList(encoders)
        self.decoders = torch.nn.ModuleList(decoders)
        if private_encoders:
            self.private_encoders = torch.nn.ModuleList(private_encoders)
        else:
            self.private_encoders = None
        self.recon_loss_type = recon_loss_type



    def forward(self, views, mle=True, **kwargs):
        """
        Forward method for the model. Outputs latent encoding for each view

        :param views:
        :param kwargs:
        :return:
        """
        z = {}
        # Used when we get reconstructions
        z["mu_shared"], z["logvar_shared"] = self._encode(views)
        z["shared"] = self._sample(z["mu_shared"], z["logvar_shared"], mle)
        if self.private_encoders is not None:
            z["mu_private"], z["logvar_private"] = self._encode_private(views)
            z["private"] = [
                self._sample(mu_, logvar_, mle)
                for mu_, logvar_ in zip(z["mu_private"], z["logvar_private"])
            ]
        return z


    def _sample(self, mu, logvar, mle):
        """

        :param mu:
        :param logvar:
        :param mle: whether to return the maximum likelihood (i.e. mean) or whether to sample
        :return: a sample from latent vector
        """
        if mle:
            return mu
        else:
            return mu + torch.randn_like(logvar) * torch.exp(0.5 * logvar)

    def _encode(self, views):
        mu = []
        logvar = []
        for i, encoder in enumerate(self.encoders):
            mu_i, logvar_i = encoder(views[i])
            mu.append(mu_i)
            logvar.append(logvar_i)
        return torch.stack(mu).sum(dim=0), torch.stack(logvar).sum(dim=0)

    def _encode_private(self, views):
        mu = []
        logvar = []
        for i, private_encoder in enumerate(self.private_encoders):
            mu_i, logvar_i = private_encoder(views[i])
            mu.append(mu_i)
            logvar.append(logvar_i)
        return mu, logvar

    def _decode(self, z, uncertainty=False, **kwargs):
        x = []
        if uncertainty:
            z["shared"] = z["logvar_shared"]
            if self.private_encoders is not None:
                z["private"] = z["logvar_private"]
        for i, decoder in enumerate(self.decoders):
            if "private" in z:
                x_i = decoder(
                    torch.cat(
                        (
                            self.latent_dropout(z["shared"]),
                            self.latent_dropout(z["private"][i]),
                        ),
                        dim=-1,
                    )
                )
            else:
                x_i = decoder(self.latent_dropout(z["shared"]))
            x.append(x_i)
        return x



    def loss(self, batch, **kwargs):
        z = self(batch["views"], mle=False)
        recons = self._decode(z)
        loss = dict()
        loss["reconstruction"] = torch.stack(
            [
                self.recon_loss(x, recon, loss_type=self.recon_loss_type)
                for x, recon in zip(batch["views"], recons)
            ]
        ).sum()
        loss["kl shared"] = (
            self.kl_loss(z["mu_shared"], z["logvar_shared"]) / batch["views"][0].numel()
        )
        if "private" in z:
            loss["kl private"] = torch.stack(
                [
                    self.kl_loss(mu_, logvar_) / batch["views"][0].numel()
                    for mu_, logvar_ in zip(z["mu_private"], z["logvar_private"])
                ]
            ).sum()
        loss["objective"] = torch.stack(tuple(loss.values())).sum()
        return loss




    def transform(
        self,
        loader: torch.utils.data.DataLoader,
    ):
        """
        :param loader: a dataloader that matches the structure of that used for training
        :return: transformed representations
        """
        with torch.no_grad():
            z_shared = []
            z_private = []
            for batch_idx, batch in enumerate(loader):
                views = [view.to(self.device) for view in batch["views"]]
                z_ = self(views)
                z_shared.append(z_["shared"].cpu())
                if "private" in z_:
                    z_private.append(self.detach_all(z_["private"]))
        z = {"shared": torch.vstack(z_shared).cpu().numpy()}
        if "private" in z_:
            z["private"] = [torch.vstack(i).cpu().numpy() for i in zip(*z_private)]
        return z