o }oi@sddlmZmZddlZddlZddlmZddlm Z ddl m Z m Z m Z mZedgedgedgddd d fd ejd ejd ejdedededejdejfddZGdddeZdS))DictUnionN)Dataset) PinholeCamera)compute_look_at_vectorsconstruct_posesget_raysget_view_direction皙 @<Fcudaradiusthetaphiangle_overhead angle_front return_dirsdevicereturnc Cs|dtj}|dtj}|dtj}|dtj}tj|t|t||t||t|t|gdd}t||d\}} } t|| | ||d} |rZt||||nd} | | fS)a Generate camera poses based on a circular arrangement. Parameters: radius: torch.Tensor - Radii for the camera positions. theta: torch.Tensor - Theta angles for the camera positions. phi: torch.Tensor - Phi angles for the camera positions. angle_overhead: float - Angle range of the overhead view. angle_front: float - Angle range of the front view. return_dirs: bool - Whether to return the view directions. device: str - The device to allocate the tensor on (e.g., 'cuda' or 'cpu'). Returns: Tuple: Contains the following: - poses (torch.Tensor): Generated poses, shape [size, 4, 4]. - dirs (torch.Tensor, optional): View directions, if requested. dim)centersr)r right_vector up_vectorforward_vectorrN) nppitorchstacksincosrrr ) rrrrrrrrrrrposesdirsr&f/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/multimodal/data/nerf/circle_poses.py circle_posess"   r(cseZdZdZ          d&d ededededededededededededejddffdd ZdefddZ d ede e e eej fffd!d"Zd#ede e e eej fffd$d%ZZS)'CirclePosesDatasetz3 A dataset class to generate circle poses. d4@r V@>@N@{Gz?@@cpusizeheightwidth default_fovx default_fovydefault_radius default_polardefault_azimuthrrnearfarrrNcszt||_||_||_||_||_||_||_||_ | |_ | |_ | |_ | |_ | |_t|j|j|j |j |jd|_dS)a6 Initializes a new CirclePosesDataset instance. Parameters: size (int): Number of samples in the dataset. height (int): Height of the image. width (int): Width of the image. default_fovx (float): Default field of view in x-direction. default_fovy (float): Default field of view in y-direction. default_radius (float): Default radius of the circle. default_polar (float): Default polar angle. default_azimuth (float): Default azimuth angle. angle_overhead (float): Overhead angle. angle_front (float): Frontal angle. near (float): Near clipping distance. far (float): Far clipping distance. device (torch.device): Device to generate data on. )r6r5r<r=rN)super__init__r4r5r6r7r8r9r:r;rrr<r=rrcamera)selfr4r5r6r7r8r9r:r;rrr<r=r __class__r&r'r?Zs" " zCirclePosesDataset.__init__cCs|jS)z-Returns the number of samples in the dataset.)r4)rAr&r&r'__len__szCirclePosesDataset.__len__idxc Cst|jg|j}t||jdg|j}t|jg|j}t||||j|j d|jd\}}|j j |j |j d}|j j|d|dd}|t|} t|||j|j|jd\} } ||j} | | d kd8<|j|j| | || | d } | S) aGet an item from the dataset. Args: idx (int): Index of the item to retrieve. Returns: dict: Data dictionary containing the following: - height (int): Height of the image. - width (int): Width of the image. - rays_o (torch.Tensor): Ray origins, shape [height, width, 3]. - rays_d (torch.Tensor): Ray directions, shape [height, width, 3]. - dir (torch.Tensor): View direction, shape [3]. - mvp (torch.Tensor): Model-view-projection matrix, shape [4, 4]. - azimuth (torch.Tensor): Azimuth angle, shape [1]. ihT)rrrrrrr)fovxfovyr)focal_xfocal_y)r$ intrinsicsr5r6rr)r5r6rays_orays_ddirmvpazimuth)r FloatTensorr:torr4r9r(rrr@compute_intrinsicsr7r8compute_projection_matrixinverserr5r6r;)rArEthetasphisrr$r%rK projectionrOrLrM delta_azimuthdatar&r&r' __getitem__s:    zCirclePosesDataset.__getitem__batchc Cs||j|jtjdd|Dddtjdd|Dddtjdd|Dddtjdd|Dddtjdd|Dddd S) zCollate function to combine multiple data points into batches. Args: batch (list): List of data dictionaries. Returns: dict: Collated data. cSg|]}|dqS)rLr&.0itemr&r&r' z1CirclePosesDataset.collate_fn..rrcSr])rMr&r^r&r&r'rarbcSr])rOr&r^r&r&r'rarbcSr])rNr&r^r&r&r'rarbcSr])rPr&r^r&r&r'rarb)r5r6rLrMrOrNrP)r5r6r cat)rAr\r&r&r' collate_fns zCirclePosesDataset.collate_fn) r*r+r+r,r,r r-r.r/r0r1r2r3)__name__ __module__ __qualname____doc__intfloatr rr?rDrstrrTensorr[listrd __classcell__r&r&rBr'r)Us^     9$,s>     7