o i @sddlmZmZddlZddlmZddlmmZd#ddZ Gdddej Z d$d ej d e d ed ej fddZd$d e dej d ed ej fddZ   d%de de dedejdejd ej f ddZ    d&dej deee e fdfdeedfded ed ej f d!d"ZdS)')TupleUnionNnorm_expexpp1c Cs|dddd}|durdn|}|ddddddd| f}|dkr3|dddddddfn|dddddd| df}|dkr_|jddd jd d }||}|t|} n:|d krm||jddd } n,|d krwt|} n"|dkrt|} n|dkrt|} n|dkr|} nt d||dkrd|} | | fS|dkr|} | | fS|dkrt|} | | fS|dkr|} | | fSt d|)aq Process network output to extract GS params and density values. Density could be view-dependent as SH coefficient Args: out: Network output tensor (B, C, H, W) activation: Activation type for 3D points conf_activation: Activation type for confidence values Returns: Tuple of (3D points tensor, confidence tensor) rNrT)dimkeepdimg:0yE>)minnormexprelusigmoidlinearzUnknown activation: rZexpp0zUnknown conf_activation: ) permuter clamptorchexpm1rFrr ValueError) out activationconf_activationZconf_dimfmapxyzconfdZ xyz_normedZpts3dZconf_outrE/data/cameron/da3_repo/src/depth_anything_3/model/utils/head_utils.pyactivate_head_gss@"F      r!csZeZdZUdZeedfed<deedfddffdd Zdej dej fd d Z Z S) PermutezHnn.Module wrapper around Tensor.permute for cleaner nn.Sequential usage..dimsreturnNcst||_dSN)super__init__r#)selfr# __class__rr r'Xs  zPermute.__init__xcCs |j|jSr%)rr#)r(r+rrr forward\s zPermute.forward) __name__ __module__ __qualname____doc__rint__annotations__r'rTensorr, __classcell__rrr)r r"Ss r"dpos_grid embed_dimomega_0r$c Cs|j\}}}|dks J|d|}t|d|dddf|d}t|d|dddf|d}tj||gdd} | |||S)a* Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC) Args: pos_grid: Tensor of shape (H, W, 2) containing 2D coordinates embed_dim: Output channel dimension for embeddings Returns: Tensor of shape (H, W, embed_dim) with positional embeddings rr Nr)r8rr )shapereshapemake_sincos_pos_embedrcatview) r6r7r8HWZgrid_dimZpos_flatZemb_xZemb_yembrrr position_grid_to_embed`s  rBposcCs|ddksJtj|dtj|jd}||d}d||}|d}td||}t|}t|}tj||gdd }| S) a" This function generates a 1D positional embedding from a given grid using sine and cosine functions. # noqa Args: - embed_dim: The embedding dimension. - pos: The position to generate the embedding from. Returns: - emb: The generated 1D positional embedding. rr)dtypedeviceg@?r zm,d->mdrr9) rarangefloat32rEr;einsumsincosr=float)r7rCr8omegaremb_sinemb_cosrArrr r<{s      r<widthheight aspect_ratiorDrEcCs|dur t|t|}|ddd}||}d|}| |d|}||d|} | |d|} ||d|} tj|| |||d} tj| | |||d} tj| | dd\}}tj||fd d }|S) a Create a normalized UV grid of shape (width, height, 2). The grid spans horizontally and vertically according to an aspect ratio, ensuring the top-left corner is at (-x_span, -y_span) and the bottom-right corner is at (x_span, y_span), normalized by the diagonal of the plane. Args: width (int): Number of points horizontally. height (int): Number of points vertically. aspect_ratio (float, optional): Width-to-height ratio. Defaults to width/height. dtype (torch.dtype, optional): Data type of the resulting tensor. device (torch.device, optional): Device on which the tensor is created. Returns: torch.Tensor: A (width, height, 2) tensor of UV coordinates. NrrFg?r)stepsrDrExy)indexingr r9)rLrlinspacemeshgridstack)rPrQrRrDrEZ diag_factorZspan_xZspan_yZleft_xZright_xZtop_yZbottom_yx_coordsy_coordsuuvvZuv_gridrrr create_uv_gridsr]bilinearTr+size scale_factormode align_cornersc sdur|dus Jdt|jd|t|jd|fd}dd|jd|jd}||krVtj|||ddd}fd d |D}tj|dd Stjj|d S) zi Safe interpolation implementation to avoid INT_MAX overflow in torch.nn.functional.interpolate. Nz-Either size or scale_factor must be provided.r i`rr)chunksr cs g|] }tjj|dqS)r_rarb)nn functional interpolate).0crbrar_rr sz&custom_interpolate..r9re) r1r:rchunkr= contiguousrfrgrh) r+r_r`rarbZINT_MAXtotalrdoutsrrkr custom_interpolates $$rq)rrN)r5)NNN)NNr^T)typingrrrtorch.nnrfZtorch.nn.functionalrgrr!Moduler"r3r1rLrBr<rDrEr]strboolrqrrrr sd  :   7