U h8?@s ddlZddlZddlZddlmZmZmZmZmZm Z m Z m Z ddl Z ddlZddlmZddlmZmZddlmZmZddlmZddlmZmZdd lmZm Z m!Z!m"Z"m#Z#m$Z$Gd d d eZ%Gd d d eZ&Gddde&Z'Gddde&Z(GdddeZ)dS)N)AnyCallableDictListOptionalSequenceTupleUnion)one_hot) tree_flattentree_unflatten) transforms tv_tensors) functional)_RandomApplyTransform Transform)_check_sequence_input_parse_labels_getterhas_anyis_pure_tensor query_chw query_sizecseZdZdZejZeee fdfdd Z de e e e fe e e fe e d fd d Zee e e e d fdd Zee eee fdddZe eee fe dddZZS) RandomErasingaNRandomly select a rectangle region in the input image or video and erase its pixels. This transform does not support PIL Image. 'Random Erasing Data Augmentation' by Zhong et al. See https://arxiv.org/abs/1708.04896 Args: p (float, optional): probability that the random erasing operation will be performed. scale (tuple of float, optional): range of proportion of erased area against input image. ratio (tuple of float, optional): range of aspect ratio of erased area. value (number or tuple of numbers): erasing value. Default is 0. If a single int, it is used to erase all pixels. If a tuple of length 3, it is used to erase R, G, B channels respectively. If a str of 'random', erasing each pixel with random values. inplace (bool, optional): boolean to make this transform inplace. Default set to False. Returns: Erased input. Example: >>> from torchvision.transforms import v2 as transforms >>> >>> transform = transforms.Compose([ >>> transforms.RandomHorizontalFlip(), >>> transforms.PILToTensor(), >>> transforms.ConvertImageDtype(torch.float), >>> transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), >>> transforms.RandomErasing(), >>> ]) )returncs"tt|jdkrdn|jdS)Nrandom)value)dictsuper _extract_params_for_v1_transformr)self __class__T/var/www/html/venv/lib/python3.8/site-packages/torchvision/transforms/v2/_augment.pyr2sz.RandomErasing._extract_params_for_v1_transform?g{Gz?gQ?g333333?gffffff @F)pscaleratiorinplacecs:tj|dt|tjtttfs*tdt|trD|dkrDt dt|ttfsZtdt|ttfsptd|d|dks|d|dkrt d |ddks|ddkrt d ||_ ||_ t|ttfrt|g|_n:t|trd|_n(t|ttfrd d |D|_n||_||_tt|j |_dS) N)r)z=Argument value should be either a number or str or a sequencerz&If value is str, it should be 'random'zScale should be a sequencezRatio should be a sequencerrz,Scale and ratio should be of kind (min, max)zScale should be between 0 and 1cSsg|] }t|qSr#)float).0vr#r#r$ Tsz*RandomErasing.__init__..)r__init__ isinstancenumbersNumberstrtuplelist TypeError ValueErrorwarningswarnr*r+intr-rr,torchlogtensor _log_ratio)r r)r*r+rr,r!r#r$r18s0   zRandomErasing.__init__)rinptargskwargsrcsJt|tjtjfr4tt|jdt|jdtj ||f||S)Nz:() is currently passing through inputs of type tv_tensors.z(. This will likely change in the future.) r2r BoundingBoxesMaskr:r;type__name__r _call_kernel)r rrArBrCr!r#r$rH[s zRandomErasing._call_kernel flat_inputsrcCst|\}}}|jdk r:t|jd|fkr:td|d||}|j}tdD] }|td|j d|j d }t td|d|d } t t t|| } t t t|| } | |krP| |ksqP|jdkrtj|| | gtjd} nt|jddddf} tjd|| ddd } tjd|| ddd }qxqPdd||df\} }} } } t| || | | d S) Nrz@If value is a sequence, it should have either a single value or z (number of inpt channels) r)dtypersize)ijhwr/)rrlenr9r@ranger=emptyZuniform_r*itemexpr<roundmathsqrtZfloat32Znormal_r?randintr)r rJZimg_cZimg_hZimg_wZareaZ log_ratio_Z erase_areaZ aspect_ratiorRrSr/rPrQr#r#r$ _get_paramscs6 &  zRandomErasing._get_paramsrAparamsrcCs,|ddk r(|jtj|f|d|ji}|S)Nr/r,)rHFZeraser,r rAr`r#r#r$ _transforms zRandomErasing._transform)r%r&r'r(F)rG __module__ __qualname____doc__ _transformsrZ_v1_transform_clsrr5rrr-rboolr1rrHrr^rc __classcell__r#r#r!r$rs$  #&rcsdeZdZddddeeeddfddZdd Zej ed d d Z ej eej d ddZ Z S)_BaseMixUpCutMix?Ndefault)alpha num_classes labels_getter)rmrnrcsHtt||_tjt|gt|g|_||_ t ||_ dSN) rr1r-rmr= distributionsBetar?_distrnr_labels_getter)r rmrnror!r#r$r1s    z_BaseMixUpCutMix.__init__csht|dkr|n|d}t|\}}|}t|tjjtjtjrXt t j d |t tjst dt djdkrt djdjdkrjdk rjd jkrt d jd d jd jdkrjdkrt d jddddt||Dd|tfddt|D<fddt||D}t||S)Nrrz9() does not support PIL images, bounding boxes and masks.z%The labels must be a tensor, but got z instead.)rzlabels should be index based with shape (batch_size,) or probability based with shape (batch_size, num_classes), but got a tensor of shape ruzYWhen passing 2D labels, the number of elements in last dimension must match num_classes:  != z$. You can Leave num_classes to None.z=num_classes must be passed if the labels are index-based (1D))labels batch_sizecSsg|]\}}|r|qSr#r#r.rAZneeds_transformr#r#r$r0sz,_BaseMixUpCutMix.forward..Tc3s|]\}}|kr|VqdSrpr#)r.idxrA)rxr#r$ sz+_BaseMixUpCutMix.forward..cs$g|]\}}|r|n|qSr#)rcrz)r`r r#r$r0s)rTr Z_needs_transform_listrPILImagerrDrEr9rFrGrtr2r=Tensorndimshapernr^zipnext enumerater )r inputsrJspecZneeds_transform_listZ flat_outputsr#)rxr`r r$forwards<      $  z_BaseMixUpCutMix.forward)rArycCsdt|tjrdnd}|j|kr6td|d|jd|jd|kr`td|jdd|d dS) NzExpected a batched input with z dims, but got z dimensions instead.rzRThe batch size of the image or video does not match the batch size of the labels: rw.)r2rVideorr9r)r rAryZexpected_num_dimsr#r#r$_check_image_or_videos z&_BaseMixUpCutMix._check_image_or_video)labellamrcCsJ|jdkrt||jd}|jjs(|}|ddd|| |S)Nr)rnrrk) rr rnrLZis_floating_pointr-rollmul_add_mul)r rrr#r#r$ _mixup_labels  z_BaseMixUpCutMix._mixup_label) rGrdrer-rr<r1rr=rrrrir#r#r!r$rjs$ - rjc@sFeZdZdZeeeeefdddZeeeefedddZ dS) MixUpaApply MixUp to the provided batch of images and labels. Paper: `mixup: Beyond Empirical Risk Minimization `_. .. note:: This transform is meant to be used on **batches** of samples, not individual images. See :ref:`sphx_glr_auto_examples_transforms_plot_cutmix_mixup.py` for detailed usage examples. The sample pairing is deterministic and done by matching consecutive samples in the batch, so the batch needs to be shuffled (this is an implementation detail, not a guaranteed convention.) In the input, the labels are expected to be a tensor of shape ``(batch_size,)``. They will be transformed into a tensor of shape ``(batch_size, num_classes)``. Args: alpha (float, optional): hyperparameter of the Beta distribution used for mixup. Default is 1. num_classes (int, optional): number of classes in the batch. Used for one-hot-encoding. Can be None only if the labels are already one-hot-encoded. labels_getter (callable or "default", optional): indicates how to identify the labels in the input. By default, this will pick the second parameter as the labels if it's a tensor. This covers the most common scenario where this transform is called as ``MixUp()(imgs_batch, labels_batch)``. It can also be a callable that takes the same input as the transform, and returns the labels. rIcCstt|jddS)Nr#r)rr-rssample)r rJr#r#r$r^szMixUp._get_paramsr_cCs|d}||dkr"|j||dSt|tjtjfs`_. .. note:: This transform is meant to be used on **batches** of samples, not individual images. See :ref:`sphx_glr_auto_examples_transforms_plot_cutmix_mixup.py` for detailed usage examples. The sample pairing is deterministic and done by matching consecutive samples in the batch, so the batch needs to be shuffled (this is an implementation detail, not a guaranteed convention.) In the input, the labels are expected to be a tensor of shape ``(batch_size,)``. They will be transformed into a tensor of shape ``(batch_size, num_classes)``. Args: alpha (float, optional): hyperparameter of the Beta distribution used for mixup. Default is 1. num_classes (int, optional): number of classes in the batch. Used for one-hot-encoding. Can be None only if the labels are already one-hot-encoded. labels_getter (callable or "default", optional): indicates how to identify the labels in the input. By default, this will pick the second parameter as the labels if it's a tensor. This covers the most common scenario where this transform is called as ``CutMix()(imgs_batch, labels_batch)``. It can also be a callable that takes the same input as the transform, and returns the labels. rIcCst|jd}t|\}}tj|dd}tj|dd}dtd|}t||}t||} ttj ||dd} ttj || dd} ttj |||d} ttj || |d} | | | | f}td| | | | ||}t ||d S) Nr#rMrNr%rkr)min)max)box lam_adjusted) r-rsrrr=r\rZr[r<clampr)r rJrHWZr_xZr_yrZr_w_halfZr_h_halfx1y1x2y2rrr#r#r$r^'s     zCutMix._get_paramsr_c Cs||dkr|j||ddSt|tjtjfs8t|r|j||dd|d\}}}}|dd}|}|d ||||f|d ||||f<t|tjtjfrtj ||d }|S|SdS) Nrxrrryrrrr.r) rr2rr~rrrrcloner) r rAr`rrrrZrolledrr#r#r$rc=s  (zCutMix._transformNrr#r#r#r$r srcsheZdZdZeeeefdfdd Zee e e e fdddZ e e e e fe dd d Z ZS) JPEGa>Apply JPEG compression and decompression to the given images. If the input is a :class:`torch.Tensor`, it is expected to be of dtype uint8, on CPU, and have [..., 3 or 1, H, W] shape, where ... means an arbitrary number of leading dimensions. Args: quality (sequence or number): JPEG quality, from 1 to 100. Lower means more compression. If quality is a sequence like (min, max), it specifies the range of JPEG quality to randomly select from (inclusive of both ends). Returns: image with JPEG compression. qualitycstt|tr||g}nt|dddd|dkrR|dkrRdkrrnnt|dtrrt|dtstd|||_dS)Nr)ru)Z req_sizesrrdz7quality must be an integer from 1 to 100, got quality =)rr1r2r<rr9r)r rr!r#r$r1`s   Fz JPEG.__init__rIcCs,t|jd|jddd}t|dS)Nrrr#r)r=r\rrWr)r rJrr#r#r$r^ls"zJPEG._get_paramsr_cCs|jtj||ddS)Nrr)rHraZjpegrbr#r#r$rcpszJPEG._transform)rGrdrerfr r<rr1rrrr5r^rcrir#r#r!r$rPs r)*rZr3r:typingrrrrrrrr Z PIL.Imager}r=Ztorch.nn.functionalr Ztorch.utils._pytreer r Z torchvisionr rgrZtorchvision.transforms.v2rrarcrrZ_utilsrrrrrrrrjrrrr#r#r#r$s (   K0E