import torch from torch import nn from diffcve.modules.vae_blocks import Encoder, PriorConditionedVideoDecoder from diffcve.modules.distributions import DiagonalGaussianDistribution class VideoAutoencoder(nn.Module): def __init__(self, ddconfig, embed_dim, fusion_w=1.0): super().__init__() self.encoder = Encoder(**ddconfig) self.decoder = PriorConditionedVideoDecoder(**ddconfig) self.decoder.fusion_w = fusion_w self.quant_conv = nn.Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1) self.post_quant_conv = nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) self.embed_dim = embed_dim def encode(self, x, return_features=True): if return_features: h, encoder_features = self.encoder(x, return_features=True) else: h = self.encoder(x) posterior = DiagonalGaussianDistribution(self.quant_conv(h)) return (posterior, encoder_features) if return_features else posterior def decode(self, z, encoder_features, prior_features, qp): return self.decoder(self.post_quant_conv(z), encoder_features, prior_features, qp)