o oiM? @s^dZddlZddlmZddlZddlmmZddl m Z ddl m Z ddl m Z  d+d ejd eeefd ed eejejffd dZdejdeeefd ejfddZdejded ejfddZ d,dejdededed ejf ddZdejd ejd ejfd!d"Zdejd ejd ejfd#d$Ze  % d-d&ejd'ed eeefd(ed ejf d)d*ZdS).zHIn this module several equalization methods are exposed: he, ahe, clahe.N)Tuple_torch_histc_cast)perform_keep_shape_image) histogramFimgs grid_sizeeven_tile_sizereturnc Cs<|}|jdd\}}t||d}t||d}|r3||dr&dnd7}||dr0dnd7}||d|}||d|} ||jdksQ| |jdkrUtd|dks]| dkritj|d| d|gdd }|jd } |d| | d||d ||d} | jd |dkrt | jd |dkrt | |fS)aCompute tiles on an image according to a grid size. Note that padding can be added to the image in order to crop properly the image. So, the grid_size (GH, GW) x tile_size (TH, TW) >= image_size (H, W) Args: imgs: batch of 2D images with shape (B, C, H, W) or (C, H, W). grid_size: number of tiles to be cropped in each direction (GH, GW) even_tile_size: Determine if the width and height of the tiles must be even. Returns: tensor with tiles (B, GH, GW, C, TH, TW). B = 1 in case of a single image is provided. tensor with the padded batch of 2D imageswith shape (B, C, H', W'). NrrzBCannot compute tiles on the image according to the given grid sizereflect)mode) shapemathceil ValueErrorFpadunfoldsqueeze contiguousAssertionError) rr r batchhw kernel_vert kernel_horzpad_vertpad_horzctilesr(O/home/ubuntu/.local/lib/python3.10/site-packages/kornia/enhance/equalization.py_compute_tiles s2    r* padded_imgs tile_sizecCs|dkr td|jd|ddkrtd|jd|ddkr(td|dd}|dd}|jd }|d||d||d ||d}|jd |krXt|jd|ddkret|jd|ddkrrt|S) aCompute interpolation tiles on a properly padded set of images. Note that images must be padded. So, the tile_size (TH, TW) * grid_size (GH, GW) = image_size (H, W) Args: padded_imgs: batch of 2D images with shape (B, C, H, W) already padded to extract tiles of size (TH, TW). tile_size: shape of the current tiles (TH, TW). Returns: tensor with the interpolation tiles (B, 2GH, 2GW, C, TH/2, TW/2). zImages Tensor must be 4D.r rz Images are not correctly padded.rrr rr)dimrrrrr)r+r,interp_kernel_vertinterp_kernel_horzr& interp_tilesr(r(r)_compute_interpolation_tilesZs,       r2r'binscCst||dddS)Nrrr3minmaxr)r'r3r(r(r) _my_histcsr7D@ tiles_x_imnum_binsclipdiffc s|dkr td|j\}}}}}} || } |d| } |sBtjr1tfdd| D} n*ttt t | gt | } ntj dd| j d} t| | td } | | 9} |d krt|| d}| j|d | | d}t|}||}| |d dd7} tj| j d}|| jdd}| ||d ddk7} d| }t| d|}|dd}|s|}|||||f}|S) aCompute luts for a batched set of tiles. Same approach as in OpenCV (https://github.com/opencv/opencv/blob/master/modules/imgproc/src/clahe.cpp) Args: tiles_x_im: set of tiles per image to apply the lut. (B, GH, GW, C, TH, TW) num_bins: number of bins. default: 256 clip: threshold value for contrast limiting. If it is 0 then the clipping is disabled. diff: denote if the differentiable histagram will be used. Default: False Returns: Lut for each tile (B, GH, GW, C, 256). zTensor must be 6D.rcsg|] }t|dddqS)rrr4r).0tiler;r(r) sz!_compute_luts..rr)devicegMbP?g)r6N)r.rrviewtorchjit is_scriptingstacklistmapr7lenlinspacerCrtensorrr6clamp_sum remainderdiv transposearangerepeatcumsumclampfloor)r:r;r<r=bghgwr&thtwpixelsr'histosr3max_valclippedresidualredistv_range mat_range lut_scalelutsr(rAr) _compute_lutss:   "   rgr1rfc Cs8|dkr td|dkrtd|j\}}}}}}tjddtjd}|dkrRtd |d |dd jdd d }t gd gd g|dd}||7}tjddtjd} |dkrtd |d |dd jdd d } t gdgdg|dd} | | 7} tj |||d||jdfd|j |j d} |ddddt |dd d ddt |dd d f| dddd|d dd|d df<|dd|dddfddt |dd d f| ddd ddd|d df<|dd|dddfddt |dd d f| ddd ddd|d d f<|ddddt |dd d | dddff| dddd|d d ddf<|ddddt |dd d | ddd ff| dddd|d d dd f<|dd|t |dd d d d dd| t |dd d d f| ddd dd dddf<| S)a Assign the required luts to each tile. Args: interp_tiles: set of interpolation tiles. (B, 2GH, 2GW, C, TH/2, TW/2) luts: luts for each one of the original tiles. (B, GH, GW, C, 256) Returns: mapped luts (B, 2GH, 2GW, 4, C, 256) r>interp_tiles tensor must be 6D.luts tensor must be 5D.rr-)dtyper rtrunc) rounding_mode)rrrr)rrrr)rrrr)rrrrrrkrCN)r.rrrEemptylongrSrDrQrMfullrkrCr6rTpermute) r1rfnum_imgsrYrZr&_j_idxsj_floori_idxsi_floorluts_x_interp_tilesr(r(r) _map_lutss6  &"&"^PPPP> rzc Cs|dkr td|dkrtdt||}|j\}}}}}}|ddd} | d|||d |||} t |d|  | |||d |||} t |} tj d |d dd|j|jd d |d dddd ||} | d||d ||} tj d |d dd|j|jd d |d |d |} | d||d ||} | |d d d || |d d ||d}| |d d d d d d }| ddd dd dfd\}}}}t||t||}t||t||}t||t||| ddd dd df<| dddd|d dd|d df| dddd|d dd|d f<| ddd ddfd \}}}}t||d t||| ddd ddf<| ddd d|d fd \}}}}t||d t||| ddd d|d f<| dddd dfd \}}}}t||t||| dddd df<| dd|d d dfd \}}}}t||t||| dd|d d df<| dS)a/Equalize the tiles. Args: interp_tiles: set of interpolation tiles, values must be in the range [0, 1]. (B, 2GH, 2GW, C, TH/2, TW/2) luts: luts for each one of the original tiles. (B, GH, GW, C, 256) Returns: equalized tiles (B, 2GH, 2GW, C, TH/2, TW/2) r>rhrirjr rrr-r rrng@Nrrgo@)r.rrzrrpflatten unsqueezeexpandrEgathertoreshape zeros_likerSrkrCrQrRrrTunbindaddcmulsubr)r1rf mapped_lutsrsrYrZr&r[r\flatten_interp_tilespreinterp_tiles_equalizedtiles_equalizedihiwtiwtihtltrblbrtrXrtleftrightr(r(r)_compute_equalized_tilessR       .(,J$.(2$((, rrinput clip_limitslow_and_differentiablecCsJt|tstdt|t|tstdt|t|dkr+tdt|t|dts9t|dtr=td|ddksI|ddkrPtd||}t||d \}}|jd |jd f}t ||}t |||d } t || } | dd dddd |} |jd d\} } | dd| d| f} || kr| d} | S)aApply clahe equalization on the input tensor. .. image:: _static/img/equalize_clahe.png NOTE: Lut computation uses the same approach as in OpenCV, in next versions this can change. Args: input: images tensor to equalize with values in the range [0, 1] and shape :math:`(*, C, H, W)`. clip_limit: threshold value for contrast limiting. If 0 clipping is disabled. grid_size: number of tiles to be cropped in each direction (GH, GW). slow_and_differentiable: flag to select implementation Returns: Equalized image or images with shape as the input. Examples: >>> img = torch.rand(1, 10, 20) >>> res = equalize_clahe(img) >>> res.shape torch.Size([1, 10, 20]) >>> img = torch.rand(2, 3, 10, 20) >>> res = equalize_clahe(img) >>> res.shape torch.Size([2, 3, 10, 20]) z(Input clip_limit type is not float. Got z'Input grid_size type is not Tuple. Got r z4Input grid_size is not a Tuple with 2 elements. Got rrz=Input grid_size type is not valid, must be a Tuple[int, int].z/Input grid_size elements must be positive. Got Tr r)r<r=rr-riN.) isinstancefloat TypeErrortypetuplerKrr*rr2rgrrr reshape_asr.r)rrr rr hist_tiles img_paddedr,r1rfequalized_tileseq_imgsr r!r(r(r)equalize_claheKs0 "     r)F)r8r9F)r9rF)__doc__rtypingrrEtorch.nn.functionalnn functionalrkornia.utils.helpersrkornia.utils.imagerrTensorintboolr*r2r7rrgrzrrr(r(r(r)sb      ":/ 68P