o oiE @sddlmZddlmZmZmZmZmZmZm Z ddl Z ddl m Z m Z ddlmZmZddlmZddlmZddlmZd ed efd ed effd d Zd ed efd ed effddZde eeeeffd e jfddZde eeeeffd e jfddZdVdededed efddZdVdededed efddZded efddZded efddZ ded ee j!d dfd!d"Z"dd#d$ed%ee#d fdeed efd&d'Z$dWd%e ee#d fe jfd)ee%d fd dfd*d+Z&ded,e#d-e#d efd.d/Z' 0dXd%e ee#d fe jfd1e j(j)d2eed efd3d4Z* 0dXd%e ee#d fe jfd1e j(j)d2eed efd5d6Z+ 0dXd%e ee#d fe jfd7e e,efd8e e,efd2ed ef d9d:Z- 0dXd%e ee#d fe jfd;e e,efdZ.d?ed%e ee#d fee#fd@e%d dfdAdBZ/dCee%efd ee%effdDdEZ0  F 0dYdCee%efdGeee%efdHe%dIed ee%eff dJdKZ1dZde eefdMe%d efdNdOZ2de eefd efdPdQZ3ded efdRdSZ4GdTdUdUZ5dS)[)wraps)AnyCallableDictListOptionalTupleUnionN)BetaUniform)Tensor as_tensor)Boxes) Keypoints)_extract_device_dtypef.returnc *tdtdtdtdtffdd }|S)zValidate the 2D input of the wrapped function. Args: f: a function that takes the first argument as tensor. Returns: the wrapped function after input is validated. inputargskwargsrcsXt|stdt|t|jddt|tjtjtj gd|g|Ri|S)N Input type is not a Tensor. Got BCHW)required_shapesaccepted_dtypes) torch is_tensor TypeErrortype_validate_shapeshape_validate_input_dtypefloat16float32float64)rrrrU/home/ubuntu/.local/lib/python3.10/site-packages/kornia/augmentation/utils/helpers.pywrapper)s z _validate_input..wrapperrr rrr*r(r'r)_validate_inputs  r-c r)zValidate the 3D input of the wrapped function. Args: f: a function that takes the first argument as tensor. Returns: the wrapped function after input is validated. rrrrcslt|stdt|t|j}|dkrtd|dt|tjtj tj gd|g|Ri|S)Nrz"Expect input of 5 dimensions, got z insteadr) rrrr lenr"AssertionErrorr#r$r%r&)rrr input_shaper'r(r)r*As  z"_validate_input3d..wrapperr+r,r(r'r)_validate_input3d6s  r2rcC*t|trt|d}|jSt|}|jSzXInfer input shape. Input may be either (tensor,) or (tensor, transform_matrix) r) isinstancetuple_transform_inputr"rtensorr(r(r)_infer_batch_shapeP  r:cCr3r4)r5r6_transform_input3dr"r8r(r(r)_infer_batch_shape3d\r;r=Treference_shape match_channelcCst|dkr |dnd}t|dkr|dnd}t|jdkr$|d}t|jdkr<||jdkr7|dn|d}|rM|rM|jd|ksKtd |S|rW|durWtd |S) Reshape an input tensor to have the same dimensions as the reference_shape. Arguments: input: tensor to be transformed reference_shape: shape used as reference match_channel: if True, C_{src} == C_{ref}. otherwise, no constrain. C =1 by default NrBThe C dimension of tensor did not match with the reference tensor./The reference tensor do not have a C dimension!r/r" unsqueeze ValueErrorrr>r?BCr(r(r)_transform_input_by_shapehs  " rOcCst|dkr |dnd}t|dkr|dnd}t|jdkr$|d}t|jdkr7||jdkr7|d}|rH|rH|jd|ksFtd |S|rR|durRtd |S) r@r.NrArBrCrrErGrHrIrLr(r(r)_transform_input3d_by_shapes    rQcCnt|stdt|t|jdvrtd|jt|jdkr)|d}t|jdkr5|d}|S)zReshape an input tensor to be (*, C, H, W). Accept either (H, W), (C, H, W) or (*, C, H, W). Args: input: Tensor Returns: Tensor r)rErCrAzNInput size must have a shape of either (H, W), (C, H, W) or (*, C, H, W). Got rErrCrrrr r/r"rKrJrr(r(r)r7s   r7cCrR)zReshape an input tensor to be (*, C, D, H, W). Accept either (D, H, W), (C, D, H, W) or (*, C, D, H, W). Args: input: Tensor Returns: Tensor r)rCrAr.zWInput size must have a shape of either (D, H, W), (C, D, H, W) or (*, C, D, H, W). Got rCrrArSrTr(r(r)r<s    r<rcCs$|j|vrtd|d|jdS)zCheck if the dtype of the input tensor is in the range of accepted_dtypes. Args: input: Tensor accepted_dtypes: List. e.g. [torch.float32, torch.float64] zExpected input of . Got N)dtyper)rrr(r(r)r#s r#)r>outputr"cCs|}tt|jt|D]:}d}|dur0|jd|dkr&dkr0nnt|dkr0d}|j|dkrDtd|d|j|||}q|S)aCollapse the broadcasted batch dimensions an input tensor to be the specified shape. Args: output: Tensor shape: List/tuple of int reference_shape: Tensor representation of shape to control which dimensions are collapsed. Returns: Tensor rNrFrEz Dimension z# of input is expected to be 1, got )cloneranger/r"r0squeeze)rWr"r> out_tensordimidxr(r(r)_transform_output_shapes6 r^rrcCsDd}|D]}t|t|krd}nq|s td|d|ddS)zCheck if the dtype of the input tensor is in the range of accepted_dtypes. Args: shape: tensor shape required_shapes: List. e.g. ["BCHW", "BCDHW"] FTzExpected input shape in rU.N)r/r)r"rpassedrequired_shaper(r(r)r!sr! channel_indexnumbercCs|j||kS)aTValidate if an input has the right shape. e.g. to check if an input is channel first. If channel first, the second channel of an RGB input shall be fixed to 3. To verify using: _validate_input_shape(input, 1, 3) Args: input: Tensor channel_index: int number: int Returns: bool )r")rrbrcr(r(r)_validate_input_shape srdFdist same_on_batchcCslt|tr t|}|r1tdg|ddR}||}|j|dgdgt|jdRS||S)zSample from a uniform reparameterized sampling function that accepts 'same_on_batch'. If same_on_batch is True, all values generated will be exactly same given a batch_size (shape[0]). By default, same_on_batch is set to False. rFNr)r5r6rSizersamplerepeatr/r")r"rerf rsample_sizerhr(r(r)_adapted_rsamplings  & rkcCsbt|tr t|}|r,|tdg|ddRj|dgdgt|dRS||S)zSample from a uniform sampling function that accepts 'same_on_batch'. If same_on_batch is True, all values generated will be exactly same given a batch_size (shape[0]). By default, same_on_batch is set to False. rFNr)r5r6rrgsamplerir/)r"rerfr(r(r)_adapted_sampling2s @ rmlowhighcCbtt|tr|ndt|tr|ndg\}}t|||d}t|||d}t||dd}t|||S)aSample from a uniform sampling function that accepts 'same_on_batch'. If same_on_batch is True, all values generated will be exactly same given a batch_size (shape[0]). By default, same_on_batch is set to False. By default, sampling happens on the default device and dtype. If low/high is a tensor, sampling will happen in the same device/dtype as low/high tensor. NdevicerVF validate_args)rr5r r r rk)r"rnrorfrrrVrer(r(r)_adapted_uniformDs" ruabcCrp)avSample from a beta sampling function that accepts 'same_on_batch'. If same_on_batch is True, all values generated will be exactly same given a batch_size (shape[0]). By default, same_on_batch is set to False. By default, sampling happens on the default device and dtype. If a/b is a tensor, sampling will happen in the same device/dtype as a/b tensor. NrqFrs)rr5r r r rk)r"rvrwrfrrrVrer(r(r) _adapted_beta]s , rxparamnamecCs0|jt|krtd|d|d|jdS)NzInvalid shape for z . Expected rU)r"rrgr0)ryr"rzr(r(r)_shape_validationrsr{paramscCsFi}|D]\}}t|tr|||iq|||iq|S)zEPerform deep copy on any dict. Support tensor copying here. )itemsr5r updaterX)r|outkvr(r(r) deepcopy_dictws  rignoreparams_override if_none_existin_placecCsz|dur|S|r |nt|}|D](\}}||vr|||<q|dkr$q|dkr3td|d|dtd|d|S) aUOverride params dict w.r.t params_override. Args: params: source parameters. params_override: key-values to override the source parameters. if_none_exist: behaviour if the key in `params_override` does not exist in `params`. 'raise' | 'ignore'. in_place: if to override in-place or not. NrraisezParam `z` not existed in `z`.`z` is not a valid option.)rr} RuntimeErrorrK)r|rrrrrrr(r(r)override_parameterss r vertices_plusmodecCsvt|tr*t|jdkr|jddtddgks#td|jdtj||d}t|ts9tdt |d|S)a=Preprocess input boxes. Args: input: 2D boxes, shape of :math:`(N, 4, 2)`, :math:`(B, N, 4, 2)` or a list of :math:`(N, 4, 2)`. See below for more details. mode: The format in which the boxes are provided. * 'xyxy': boxes are assumed to be in the format ``xmin, ymin, xmax, ymax`` where ``width = xmax - xmin`` and ``height = ymax - ymin``. With shape :math:`(N, 4)`, :math:`(B, N, 4)`. * 'xyxy_plus': similar to 'xyxy' mode but where box width and length are defined as ``width = xmax - xmin + 1`` and ``height = ymax - ymin + 1``. With shape :math:`(N, 4)`, :math:`(B, N, 4)`. * 'xywh': boxes are assumed to be in the format ``xmin, ymin, width, height`` where ``width = xmax - xmin`` and ``height = ymax - ymin``. With shape :math:`(N, 4)`, :math:`(B, N, 4)`. * 'vertices': boxes are defined by their vertices points in the following ``clockwise`` order: *top-left, top-right, bottom-right, bottom-left*. Vertices coordinates are in (x,y) order. Finally, box width and height are defined as ``width = xmax - xmin`` and ``height = ymax - ymin``. With shape :math:`(N, 4, 2)` or :math:`(B, N, 4, 2)`. * 'vertices_plus': similar to 'vertices' mode but where box width and length are defined as ``width = xmax - xmin + 1`` and ``height = ymax - ymin + 1``. ymin + 1``. With shape :math:`(N, 4, 2)` or :math:`(B, N, 4, 2)`. Note: **2D boxes format** is defined as a floating data type tensor of shape ``Nx4x2`` or ``BxNx4x2`` where each box is a `quadrilateral `_ defined by it's 4 vertices coordinates (A, B, C, D). Coordinates must be in ``x, y`` order. The height and width of a box is defined as ``width = xmax - xmin + 1`` and ``height = ymax - ymin + 1``. Examples of `quadrilaterals `_ are rectangles, rhombus and trapezoids. rArENz&Only BxNx4x2 tensor is supported. Got r_)rzExpect `Boxes` type. Got ) r5r r/r"rrgrr from_tensorr )rrr(r(r)preprocess_boxess !* rcCspt|tr't|jdkr|jddtdgks"td|jdt|d}t|tr6tdt|d|S) zPreprocess input keypoints.rCrFNrEz$Only BxNx2 tensor is supported. Got r_FzExpect `Keypoints` type. Got ) r5r r/r"rrgrrr rTr(r(r)preprocess_keypointss (  rcCs|S)zPreprocess input class tags.r(rTr(r(r)preprocess_classessrcs.eZdZdZdededdffdd ZZS)MultiprocessWrapperzYWhen used as a base class, makes the class work with the 'spawn' multiprocessing context.rrrNcs:tdd|D}dd|D}tj|i|dS)Ncss(|]}t|tjr|n|VqdS)Nr5rr rX).0argr(r(r) s&z/MultiprocessWrapper.__init__..cSs*i|]\}}|t|tjr|n|qSr(r)rkeyvalr(r(r) s*z0MultiprocessWrapper.__init__..)r6r}super__init__)selfrr __class__r(r)rszMultiprocessWrapper.__init__)__name__ __module__ __qualname____doc__rr __classcell__r(r(rr)rs"r)T)r)F)NrF)r)6 functoolsrtypingrrrrrrr rtorch.distributionsr r kornia.corer r kornia.geometry.boxesrkornia.geometry.keypointsr kornia.utilsrr-r2rgr:r=boolrOrQr7r<rVr#intr^strr!rd distributions Distributionrkrmfloatrurxr{rrrrrrr(r(r(r)s $   ""$$     2        ."   *