a d+ @sddlZddlmZddlZddlmZddlmmZddl m Z m Z ddl m Z ddlmZeeeejdddZdeejd d d Zeeed ddZGdddejZdejeeeeeejdddZGdddejZdS)N)Tuple)get_gaussian_kernel2dspatial_gradient)pi)KORNIA_CHECK_SHAPE)kdkykxreturncCs&|||}t|}|||||S)z;Utility function, which returns neigh2channels conv kernel.)torchZeyeview)rrr Znumelweightr`/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/kornia/feature/siftdesc.py_get_reshape_kernel s  r)ksizer cCsBt|d}|t|d|}t|||d}|S)zReturn a weighted pooling kernel for SIFT descriptor. Args: ksize: kernel_size. Returns: the pooling kernel with shape :math:`(ksize, ksize)`. @g?)floatr ZarangeabsZger)rZks_2Zxc2kernelrrrget_sift_pooling_kernels r) patch_sizenum_spatial_binsr cCsldt||d}||}|d}|d||dd|d}||krbtd|d|d|||fS)zReturn a tuple with SIFT parameters. Args: patch_size: the given patch size. num_spatial_bins: the ggiven number of spatial bins. Returns: ksize, stride, pad. rz Patch size zC is incompatible with requested number of spatial bins z{ for SIFT descriptor. Usually it happens when patch size is too small for num_spatial_bins specified)int ValueError)rrrstridepadZout_sizerrrget_sift_bin_ksize_stride_pad#s  r!csheZdZdZedddZdeeeeed d fd d Z e j dddZ e j dddZ ddZZS)SIFTDescriptoruModule which computes SIFT descriptors of given patches. Args: patch_size: Input patch size in pixels. num_ang_bins: Number of angular bins. num_spatial_bins: Number of spatial bins. clipval: clipping value to reduce single-bin dominance rootsift: if ``True``, RootSIFT (Arandjelović et. al, 2012) is computed. Returns: SIFT descriptor of the patches with shape. Shape: - Input: :math:`(B, 1, \text{num_spatial_bins}, \text{num_spatial_bins})` - Output: :math:`(B, \text{num_ang_bins * num_spatial_bins ** 2})` Example: >>> input = torch.rand(23, 1, 32, 32) >>> SIFT = SIFTDescriptor(32, 8, 4) >>> descs = SIFT(input) # 23x128 r cCsf|jjddt|jddt|jddt|jddt|jddt|jdS) N( num_ang_bins=, num_spatial_bins=z patch_size= rootsift=clipval=)) __class____name__str num_ang_binsrrrootsiftclipvalselfrrr__repr__RsD     zSIFTDescriptor.__repr__)rT皙?N)rr.rr/r0r c std|_||_||_||_||_||_|j}t|t d}t ||f||fd|_ t ||\|_|_|_t|jd}tjdd|d|df|j|jf|j|jfdd|_|jjj|dd|d|ddS) N绽|=rTrrrF) kernel_sizerpaddingbias)super__init__epsr.rr0r/rrmathsqrtrgkr!Z bin_ksizeZ bin_strider rnnConv2dsizepkr datacopy_reshape) r2rr.rr/r0kssigmanwr+rrr=gs,   (zSIFTDescriptor.__init__cCs |jjSN)rEr detachr1rrrget_pooling_kernelsz!SIFTDescriptor.get_pooling_kernelcCs |jSrM)rArNr1rrrget_weighting_kernelsz#SIFTDescriptor.get_weighting_kernelcCst|dd|j|jg|jd}|j|j|j|_t|d}|dddddf}|dddddf}t |||||j }t |||j dt }||j |||j}t|j|dt }t|} || } | |j} | d|j} d| |} | |}g}td|jD]>}|| |k|j| | |k|j|}||q*tj|dd}||d }tj|d d }t|d t|j}tj|d d }|jrt tj|dd |j }|S) NB1rdiffrr?dimrpg)rrshaperEtodtypedevicerr r@r>atan2rrAZ expand_asZtype_asrr.floorrangeappendcatr F normalizeclampr0r/)r2inputrQgradsgxgymagorio_bigbo0_big_wo1_big_bo0_bigbo1_bigwo0_bigwo1_bigang_binsioutrrrforwards8     . zSIFTDescriptor.forward)r4r5rTr6)r, __module__ __qualname____doc__r-r3rboolrr=r TensorrOrPrv __classcell__rrrLrr";s$"r"r4r5rTr6)rfrr.rr/r0r cCst||||||S)z`Computes the sift descriptor. See :class:`~kornia.feature.SIFTDescriptor` for details. )r")rfrr.rr/r0rrr sift_describes r}c s\eZdZdZedddZdeeeeeeed d fd d Z e j dddZ ddZ ZS)DenseSIFTDescriptoruModule, which computes SIFT descriptor densely over the image. Args: num_ang_bins: Number of angular bins. (8 is default) num_spatial_bins: Number of spatial bins per descriptor (4 is default). You might want to set odd number and relevant padding to keep feature map size spatial_bin_size: Size of a spatial bin in pixels (4 is default) clipval: clipping value to reduce single-bin dominance rootsift: (bool) if True, RootSIFT (Arandjelović et. al, 2012) is computed stride: default 1 padding: default 0 Returns: torch.Tensor: DenseSIFT descriptor of the image Shape: - Input: (B, 1, H, W) - Output: (B, num_ang_bins * num_spatial_bins ** 2, (H+padding)/stride, (W+padding)/stride) Examples:: >>> input = torch.rand(2, 1, 200, 300) >>> SIFT = DenseSIFTDescriptor() >>> descs = SIFT(input) # 2x128x194x294 r#cCsx|jjddt|jddt|jddt|jddt|jddt|jddt|jd S) Nr$r%r&r'zspatial_bin_size=r(zstride=r)r*) r+r,r-r.rspatial_bin_sizer/rr0r1rrrr3sP     zDenseSIFTDescriptor.__repr__r5rTr6rN)r.rrr/r0rr:r c std|_||_||_||_||_||_||_||_ t |jd }t j dd|d|dfdd|dd|ddfd|_|jjj|dd|d|dt j |||d||f|j|jfd|j |j fd|_|jjjt||| dS) Nr7r8rr)rrFr)r9rr;r:)r<r=r>r.rrr0r/rr rrrBrCrDbin_pooling_kernelr rFrGrH PoolingConvr) r2r.rrr/r0rr:rKrLrrr=s> (    zDenseSIFTDescriptor.__init__cCs |jjSrM)rr rNr1rrrrOsz&DenseSIFTDescriptor.get_pooling_kernelcCst|gd|\}}}}|j|j|j|_|j|j|j|_t|d}|dddddf}|dddddf}t |||||j } t |||j dt } t |j| dt } t| } | | } | |j}|d|j}d| | }| | }g}td|jD]>}|||k|j|||k|j|}||qtj|dd}||}tj|ddd dt |j}tj|ddd }|jrt tj|dd |j }|S) N)rQrRHWrSrrrrTrUr)rVrYrX)rrDrr[r\r]rrr r@r>r^rrr.r_r`rarbrcrdZclamp_r0r/)r2rfrQZCHrrrgrhrirjrkrlrmrnrorprqrrrsrtruZ out_no_normrrrrv!s:    & zDenseSIFTDescriptor.forward)r5rrTr6rr)r,rwrxryr-r3rrzrr=r r{rOrvr|rrrLrr~s**r~)r)r4r5rTr6)r?typingrr Ztorch.nnrBZtorch.nn.functionalZ functionalrcZkornia.filtersrrZkornia.geometry.conversionsrZkornia.testingrrr{rrr!Moduler"rzrr}r~rrrrs4    z