o oiV@sddlmZddlmZmZmZmZmZmZddl Z ddl m Z m Z m Z ddlmZddlmZmZmZmZddlmZddlmZmZmZdd lmZdd lmZdd lm Z eeeeeffZ!d ed edefddZ"GdddeZ#Gddde#Z$dS))Enum)AnyCallableDictOptionalTupleUnionN) Bernoulli DistributionRelaxedBernoulli)RandomGeneratorBase)_adapted_rsampling_adapted_sampling_transform_output_shapeoverride_parameters) ImageModule)Tensortensorzeros)Boxes) Keypoints)is_autocast_enabledselfinputreturncGstdt|jd)zaDefine the computation performed at every call. Should be overridden by all subclasses. zModule [z3] is missing the required "apply_tranform" function)NotImplementedErrortype__name__rrrL/home/ubuntu/.local/lib/python3.10/site-packages/kornia/augmentation/base.py_apply_transform_unimplemented,sr!c seZdZUdZdZ    d7dedededed d f fd d ZeZ e d e fe d<de de d dffdd Zd efddZde d e fddZd ddde dee ded e fddZde d d fddZd e d!eed fd e fd"d#Zd$eed fd eee ffd%d&Zd'ejd(ejd d fd)d*Zd$eed fdededed e f d+d,Z d8d-eeee fd.eeee fde d eeee feee fffd/d0Zd$eed fd eee ffd1d2Z de d-eee fd.eee fd e fd3d4Z!d9de d-eeee fde d e fd5d6Z"Z#S):_BasicAugmentationBasea_BasicAugmentationBase base class for customized augmentation implementations. Plain augmentation base class without the functionality of transformation matrix calculations. By default, the random computations will be happened on CPU with ``torch.get_default_dtype()``. To change this behaviour, please use ``set_rng_device_and_dtype``. For automatically generating the corresponding ``__repr__`` with full customized parameters, you may need to implement ``_param_generator`` by inheriting ``RandomGeneratorBase`` for generating random parameters and put all static parameters inside ``self.flags``. You may take the advantage of ``PlainUniformGenerator`` to generate simple uniform parameters with less boilerplate code. Args: p: probability for applying an augmentation. This param controls the augmentation probabilities element-wise. p_batch: probability for applying an augmentation to a batch. This param controls the augmentation probabilities batch-wise. same_on_batch: apply the same transformation across the batch. keepdim: whether to keep the output shape the same as input ``True`` or broadcast it to the batch form ``False``. F??pp_batch same_on_batchkeepdimrNcst||_||_||_||_i|_|||dks |dkr&t|j|_|dks.|dkr4t|j|_ d|_ i|_ | t dt dS)Ngr$cpu)super__init__r%r&r'r(_paramsr _p_gen _p_batch_gen_param_generatorflagsset_rng_device_and_dtypetorchdeviceget_default_dtype)rr%r&r'r( __class__rr r+Ns   z_BasicAugmentationBase.__init__.apply_transformargskwargscs:tjjj|i|\}}}}|||tj|i|S)z9Set the device and dtype for the random number generator.)r2_C_nn _parse_tor1r*to)rr8r9r3dtype_r5rr r=gs z_BasicAugmentationBase.tocCsd|jd|jd|j}t|jtr|jd|}|jD]!\}}t|tr8|d|d|j 7}q!|d|d|7}q!|j j d|dS)Nzp=z , p_batch=z, same_on_batch=z, =()) r%r&r' isinstancer/r r0itemsrnamelowerr6r)rtxtkvrrr __repr__ms  z_BasicAugmentationBase.__repr__rcC|SNrrrrr __unpack_input__xsz'_BasicAugmentationBase.__unpack_input__Tshape match_channelrOrPcCt)zStandardize input tensors.r)rrrOrPrrr transform_tensor{z'_BasicAugmentationBase.transform_tensorcCrQ)z3Check if the input tensor is formatted as expected.rRrrrr validate_tensorsz&_BasicAugmentationBase.validate_tensoroutput output_shapecCs|jrt||S|S)zStandardize output tensors.)r(r)rrVrWrrr transform_output_tensorsz._BasicAugmentationBase.transform_output_tensor batch_shapecCs|jdur |||jSiSrL)r/r')rrYrrr generate_parameterss z*_BasicAugmentationBase.generate_parametersr3r>cCs,||_||_|jdur|j||dSdS)zChange the random generation device and dtype. Note: The generated random numbers are not reproducible across different devices and dtypes. N)r3r>r/r1)rr3r>rrr r1s  z/_BasicAugmentationBase.set_rng_device_and_dtypecCs|dkr tdd}n|dkrtd}nt|jtfr#td|j|}ntd|j|}|dkri|dkr=t|dd}n'|dkrHt|d}nt|jtfrZt|df|j|}n t|df|j|}||}n||d}t |j dkr{|dS|S)Nr)r[).r) rrCr.r r rsumr-repeatlenrO)rrYr%r&r' batch_prob elem_probrrr __batch_prob_generator__s(   z/_BasicAugmentationBase.__batch_prob_generator__paramsr0cKs|d|vr|dnd}|dur|jn|}|dur|jn|}|r)t||dd}||_n ||_t||dd}t||dd}||fS)N save_kwargsFT)in_place)r,r0r)rrcr0r9rdrrr #_process_kwargs_to_params_and_flagssz:_BasicAugmentationBase._process_kwargs_to_params_and_flagscCs~|||j|j|j}|dk}|tt| g|ddR}|dur+i}||d<t |tj d}| d|i|S)Nr#r[r`)r>forward_input_shape) rbr%r&r'rZr2Sizeintr]itemrlongupdate)rrYr`to_applyr, input_sizerrr forward_parameterss,z)_BasicAugmentationBase.forward_parameterscCs||||SrL)r7)rrrcr0rrr apply_funcsz!_BasicAugmentationBase.apply_funcc Ks||}|j}||}|j}|dur||}d|vr(tdg|d|d<|j||jfi|\}}||||}|jrE| ||S|S)aPerform forward operations. Args: input: the input tensor. params: the corresponding parameters for an operation. If None, a new parameter suite will be generated. **kwargs: key-value pairs to override the parameters and flags. Note: By default, all the overwriting parameters in kwargs will not be recorded as in ``self._params``. If you wish it to be recorded, you may pass ``save_kwargs=True`` additionally. Nr`Tr) rMrOrSrorrfr0rpr(rX) rrrcr9 in_tensor input_shaperYr0rVrrr forwards   z_BasicAugmentationBase.forward)r#r$FF)NNrL)$r __module__ __qualname____doc__ONNX_EXPORTABLEfloatboolr+r!r7rr__annotations__rr=strrJrMrrSrUrrirXrrZr2r3r>r1rbrfrorprs __classcell__rrr5r r"4s|    "  % "* 0r"c @seZdZdZ d*dedeeefdeeefdeedef dd Z d*dedeeefdeeefdeedef d d Z d*dedeeefdeeefdeed edef d dZ d*dedeeefdeeefdeed edef ddZ d*de deeefdeeefdeed ede f ddZ d*dedeeefdeeefdeed edef ddZ d*dedeeefdeeefdeed edef ddZ d*dedeeefdeeefdeedef ddZ d*dedeeefdeeefdeedef ddZ d*de deeefdeeefdeede f ddZ d*de deeefdeeefdeede f ddZ d*dedeeefdeeefdeedef dd Z d*dedeeefdeeefdeedef d!d"Z d*dedeeefdeeefdeedef d#d$Z d*dedeeefdeeefdeedef d%d&Z d*d'edeeefdeeeefdefd(d)ZdS)+_AugmentationBasea_AugmentationBase base class for customized augmentation implementations. Advanced augmentation base class with the functionality of transformation matrix calculations. Args: p: probability for applying an augmentation. This param controls the augmentation probabilities element-wise for a batch. p_batch: probability for applying an augmentation to a batch. This param controls the augmentation probabilities batch-wise. same_on_batch: apply the same transformation across the batch. keepdim: whether to keep the output shape the same as input ``True`` or broadcast it to the batch form ``False``. Nrrcr0 transformrcCrQrLrRrrrcr0r~rrr r7rTz!_AugmentationBase.apply_transformcCrKrLrrrrr apply_non_transform s z%_AugmentationBase.apply_non_transformr9c Ks|j|dur |jn||fi|\}}|d}|dk}|j}||} || |r6|j| |||d} n@|sD|j| |||d} n2|j| |||d} |j| ||||durZ|n||d} t ro| |j } | |j } | |f| } |j r~t| |n| } t r| |j } | SNr`r#r~)rfr,rOrSrUallr7anyrrrr> index_putr(r) rrrcr0r~r9r`rm ori_shaperqrVappliedrrr transform_inputs+s:      z"_AugmentationBase.transform_inputsc Ks|j|dur |jn||fi|\}}|d}|dk}|j}|d} |j|| dd} || |r=|j| |||d} n1|sK|j| |||d} n#|j| |||d} |j| ||||dura|n||d} | |f| } |j rzt | || d} | S| } | S)Nr`r#rgFrNr)reference_shape) rfr,rOrSrUrapply_transform_maskrapply_non_transform_maskrr(r) rrrcr0r~r9r`rmrrOrqrVrrrr transform_masksUs6  z!_AugmentationBase.transform_masksc Kst|tstdt|d|j|dur|jn||fi|\}}|d}|dk}|r;|j||||d}|S| sJ|j ||||d}|S|j ||||d}|j|||||dur`|n||d} t ru||j }| |j } | |f| }|S)NzOnly `Boxes` is supported. Got .r`r#r)rCr RuntimeErrorrrfr,ryrapply_transform_boxrapply_non_transform_boxrr>r rrrcr0r~r9r`rmrVrrrr transform_boxesys6     z!_AugmentationBase.transform_boxesc Kst|tstdt|d|j|dur|jn||fi|\}}|d}|dk}|r9|j||||d}|S|sH|j ||||d}|S|j ||||d}|j|||||dur^|n||d} t rs||j }| |j } | |f| }|S)Nz#Only `Keypoints` is supported. Got rr`r#r) rCrrrrfr,rapply_transform_keypointrapply_non_transform_keypointrr>rrrrr transform_keypointss6     z%_AugmentationBase.transform_keypointsc Ks|j|dur |jn||fi|\}}|d}|dk}|r*|j||||d}|S|s9|j||||d}|S|j||||d}|j|||||durO|n||d} ||f| }|Sr)rfr,rapply_transform_classrapply_non_transform_classrrrrr transform_classess,   z#_AugmentationBase.transform_classescCrQ)zMProcess masks corresponding to the inputs that are no transformation applied.rRrrrr rrTz*_AugmentationBase.apply_non_transform_maskcCrQ)z?Process masks corresponding to the inputs that are transformed.rRrrrr rrTz&_AugmentationBase.apply_transform_maskcCrK)zMProcess boxes corresponding to the inputs that are no transformation applied.rrrrr rrTz)_AugmentationBase.apply_non_transform_boxcCrQ)z?Process boxes corresponding to the inputs that are transformed.rRrrrr rrTz%_AugmentationBase.apply_transform_boxcCrK)zQProcess keypoints corresponding to the inputs that are no transformation applied.rrrrr rrTz._AugmentationBase.apply_non_transform_keypointcCrQ)zCProcess keypoints corresponding to the inputs that are transformed.rRrrrr rrTz*_AugmentationBase.apply_transform_keypointcCrK)zRProcess class tags corresponding to the inputs that are no transformation applied.rrrrr rrTz+_AugmentationBase.apply_non_transform_classcCrQ)zDProcess class tags corresponding to the inputs that are transformed.rRrrrr r$rTz'_AugmentationBase.apply_transform_classrqcCs |dur|j}||||}|SrL)r0r)rrqrcr0rVrrr rp.sz_AugmentationBase.apply_funcrL)rrtrurvrrr{rrr7rrrrrrrrrrrrrrrrrprrrr r}s         /   )   *   (   "                         r})%enumrtypingrrrrrrr2torch.distributionsr r r $kornia.augmentation.random_generatorr kornia.augmentation.utilsr rrr kornia.corerModulerrrkornia.geometry.boxesrkornia.geometry.keypointsr kornia.utilsrTensorWithTransformMatr!r"r}rrrr s       S