o }oi @s8ddlZddlmZddlmZGdddejjZdS)N) AutoencoderKL) rearrangecs0eZdZdZfddZeddZZS)AutoencoderKLVAEz A class that wraps the AutoencoderKL model and provides a decode method. Attributes: vae (AutoencoderKL): The underlying AutoencoderKL model loaded from a pretrained path. cs ttj|tjd|_dS)z Initialize the AutoencoderKLVAE instance. Args: path (str): The path to the pretrained AutoencoderKL model. ) torch_dtypeN)super__init__rfrom_pretrainedtorchbfloat16vae)selfpath __class__`/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/diffusion/vae/diffusers_vae.pyrs zAutoencoderKLVAE.__init__cCs^|j\}}}}}|dkrt|d}||jjj}|jj|ddd}|dkr-t|dddS|S)a Decode a latent representation using the underlying VAE model. This method takes a latent tensor `x` and decodes it into an image. If `x` has a temporal dimension `T` of 1, it rearranges the tensor before and after decoding. Args: x (torch.Tensor): A tensor of shape (B, C, T, H, W), where: B = batch size C = number of channels T = temporal dimension H = height W = width Returns: torch.Tensor: Decoded image tensor with the same shape as the input (B, C, T, H, W). zb c t h w -> (b t) c h wF) return_dictrz(b t) c h w -> b c t h w)t)shaperr configscaling_factordecode)r xBCTHWoutrrrr&s zAutoencoderKLVAE.decode) __name__ __module__ __qualname____doc__rr no_gradr __classcell__rrrrrs   r)r diffusersreinopsrnnModulerrrrrs