U diCz@sRddlZddlZddlmZddlmZmZmZmZm Z m Z m Z ddl Z ddlZddlmZddlmZmZmZmZmZmZmZmZmZmZmZddlmZmZm Z m!Z!m"Z"m#Z#er>ddl$Z%ddl&Z%e'e'e%j(j)e'dkre%j*j+Z,ne%j*Z,er>dd l-m.Z.e,j/e.j/e,j0e.j0e,j1e.j1e,j2e.j2e,j3e.j3e,j4e.j4iZ5erTerTddl6Z6e7e8Z9e d e j:d ed ee j:ed fZ;e ed d d ed ed eed eed eed fZGdddeZ?ee@e eAe@eeffZBddZCGdddeZDddZEddZFddZGddZHe j:eId d!d"ZJdMeAee;d$d%d&ZKe j:d'd(d)ZLdNe j:e e eAe eAd*ffe=d+d,d-ZMdOe j:e e e=e@feAd.d/d0ZNdPe j:e=e eAeAfd1d2d3ZOee@e ee ffeId4d5d6ZPee@e ee ffeId4d7d8ZQeee@e ee ffeId9d:d;ZReee@e ee ffeId9dd?d@ZUdRe eIe eTe eIe e eTeeTfe e eTeeTfe eIe eAe eIe ee@eAfe eIe ee@eAfe dAdB dCdDZVGdEdFdFZWe>e e>d*feeddGdHdIZXee@ee@dJdKdLZYdS)SN)BytesIO) TYPE_CHECKINGDictIterableListOptionalTupleUnion)version) ExplicitEnum is_jax_tensoris_numpy_array is_tf_tensoris_torch_availableis_torch_tensoris_torchvision_availableis_vision_availableloggingrequires_backendsto_numpy)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STDIMAGENET_STANDARD_MEANIMAGENET_STANDARD_STDOPENAI_CLIP_MEANOPENAI_CLIP_STDz9.1.0)InterpolationModezPIL.Image.Imagez torch.Tensorz np.ndarrayz np.ndarrrayc@seZdZdZdZdS)ChannelDimensionZchannels_firstZ channels_lastN)__name__ __module__ __qualname__FIRSTLASTr$r$cCs8t|ttfr(|D]}t|sdSqn t|s4dSdS)NFT)r,listtuple valid_imagesr>)Zimgsr0r$r$r%rAs rAcCst|ttfrt|dSdS)NrF)r,r?r@r>r/r$r$r% is_batcheds rB)r<returncCs,|jtjkrdSt|dko*t|dkS)zV Checks to see whether the pixel values have already been rescaled to [0, 1]. Frr )dtypenpuint8minmaxr;r$r$r%is_scaled_images rI)expected_ndimsrCcCst|r |St|tjjr |gSt|rx|j|dkr@t|}n4|j|krR|g}n"td|dd|d|jd|Stdt|ddS) a Ensure that the input is a list of images. If the input is a single image, it is converted to a list of length 1. If the input is a batch of images, it is converted to a list of images. Args: images (`ImageInput`): Image of images to turn into a list of images. expected_ndims (`int`, *optional*, defaults to 3): Expected number of dimensions for a single input image. If the input image has a different number of dimensions, an error is raised. r z%Invalid image shape. Expected either z or z dimensions, but got z dimensions.ztInvalid image type. Expected either PIL.Image.Image, numpy.ndarray, torch.Tensor, tf.Tensor or jax.ndarray, but got .N) rBr,r-r.r>ndimr?r9r:)ZimagesrKr$r$r%make_list_of_imagess   rN)rCcCs@t|stdt|tr8t|tjjr8t|St |S)NzInvalid image type: ) r>r9r:rr,r-r.rEarrayrr/r$r$r%to_numpy_arrays  rP.)r< num_channelsrCcCs|dk r |nd}t|tr |fn|}|jdkr8d\}}n$|jdkrLd\}}ntd|j|j||kr|j||krtd|jdtjS|j||krtjS|j||krtj Std dS) a[ Infers the channel dimension format of `image`. Args: image (`np.ndarray`): The image to infer the channel dimension of. num_channels (`int` or `Tuple[int, ...]`, *optional*, defaults to `(1, 3)`): The number of channels of the image. Returns: The channel dimension of the image. Nr rJrJ)rz(Unsupported number of image dimensions: z4The channel dimension is ambiguous. Got image shape z,. Assuming channels are the first dimension.z(Unable to infer channel dimension format) r,intrMr9shapeloggerwarningrr"r#)r<rQZ first_dimZlast_dimr$r$r%infer_channel_dimension_formats"     rY)r<input_data_formatrCcCsJ|dkrt|}|tjkr$|jdS|tjkr8|jdStd|dS)a Returns the channel dimension axis of the image. Args: image (`np.ndarray`): The image to get the channel dimension axis of. input_data_format (`ChannelDimension` or `str`, *optional*): The channel dimension format of the image. If `None`, will infer the channel dimension from the image. Returns: The channel dimension axis of the image. NrJr Unsupported data format: )rYrr"rMr#r9)r<rZr$r$r%get_channel_dimension_axiss    r\)r< channel_dimrCcCs^|dkrt|}|tjkr.|jd|jdfS|tjkrL|jd|jdfStd|dS)a Returns the (height, width) dimensions of the image. Args: image (`np.ndarray`): The image to get the dimensions of. channel_dim (`ChannelDimension`, *optional*): Which dimension the channel dimension is in. If `None`, will infer the channel dimension from the image. Returns: A tuple of the image's height and width. Nr[)rYrr"rVr#r9)r<r]r$r$r%get_image_sizes   ra) annotationrCcCsVt|trRd|krRd|krRt|dttfrRt|ddksNt|ddtrRdSdS)Nimage_id annotationsrTFr,dictr?r@lenrbr$r$r%"is_valid_annotation_coco_detection1s ricCs^t|trZd|krZd|krZd|krZt|dttfrZt|ddksVt|ddtrZdSdS)NrcZ segments_info file_namerTFrerhr$r$r%!is_valid_annotation_coco_panoptic@s  rk)rdrCcCstdd|DS)Ncss|]}t|VqdSr+)ri.0annr$r$r% Qsz3valid_coco_detection_annotations..allrdr$r$r% valid_coco_detection_annotationsPsrscCstdd|DS)Ncss|]}t|VqdSr+)rkrlr$r$r%roUsz2valid_coco_panoptic_annotations..rprrr$r$r%valid_coco_panoptic_annotationsTsrt)r<timeoutrCc Cs ttdgt|tr|ds*|drHtjtt j ||dj }qt j |rbtj|}q|drz|dd}z"t|}tjt|}Wqtk r}ztd|d |W5d }~XYqXnt|tjjr|}ntd tj|}|d }|S) a3 Loads `image` to a PIL Image. Args: image (`str` or `PIL.Image.Image`): The image to convert to the PIL Image format. timeout (`float`, *optional*): The timeout value in seconds for the URL request. Returns: `PIL.Image.Image`: A PIL Image. Zvisionzhttp://zhttps://)ruz data:image/,r zIncorrect image source. Must be a valid URL starting with `http://` or `https://`, a valid path to an image file, or a base64 encoded string. Got z. Failed with NzuIncorrect format used for image. Should be an url linking to an image, a base64 string, a local path, or a PIL image.RGB)r load_imager,str startswithr-r.openrrequestsgetcontentospathisfilesplitbase64 decodebytesencode Exceptionr9 TypeErrorZImageOpsZexif_transposeconvert)r<ruZb64er$r$r%rxXs.      rxPILImageResampling Z do_rescaleZrescale_factorZ do_normalizeZ image_meanZ image_stdZdo_padZsize_divisibilityZdo_center_cropZ crop_sizeZ do_resizesizeresamplec Csx|r|dkrtd|r(|dkr(td|rD|dks<|dkrDtd|rX|dkrXtd| rt| dksl| dkrttddS)a Checks validity of typically used arguments in an `ImageProcessor` `preprocess` method. Raises `ValueError` if arguments incompatibility is caught. Many incompatibilities are model-specific. `do_pad` sometimes needs `size_divisor`, sometimes `size_divisibility`, and sometimes `size`. New models and processors added should follow existing arguments when possible. Nz=`rescale_factor` must be specified if `do_rescale` is `True`.zzDepending on the model, `size_divisibility`, `size_divisor`, `pad_size` or `size` must be specified if `do_pad` is `True`.zP`image_mean` and `image_std` must both be specified if `do_normalize` is `True`.z<`crop_size` must be specified if `do_center_crop` is `True`.zA`size` and `resample` must be specified if `do_resize` is `True`.)r9rr$r$r%validate_preprocess_argumentss   rc@seZdZdZddZdddZddZeje e e fejd d d Z dd dZ ddZdddZd ddZddZddZd!ddZdS)"ImageFeatureExtractionMixinzD Mixin that contain utilities for preparing image features. cCs4t|tjjtjfs0t|s0tdt|ddS)Nz Got type zS which is not supported, only `PIL.Image.Image`, `np.array` and `torch.Tensor` are.)r,r-r.rEndarrayrr9r:selfr<r$r$r%_ensure_format_supportedsz4ImageFeatureExtractionMixin._ensure_format_supportedNcCs||t|r|}t|tjr|dkr@t|jdtj}|jdkrf|j ddkrf| ddd}|rr|d}| tj }t j|S|S)a" Converts `image` to a PIL Image. Optionally rescales it and puts the channel dimension back as the last axis if needed. Args: image (`PIL.Image.Image` or `numpy.ndarray` or `torch.Tensor`): The image to convert to the PIL Image format. rescale (`bool`, *optional*): Whether or not to apply the scaling factor (to make pixel values integers between 0 and 255). Will default to `True` if the image type is a floating type, `False` otherwise. NrrJrRr rS)rrr4r,rErflatZfloatingrMrV transposeastyperFr-r.Z fromarray)rr<rescaler$r$r% to_pil_images    z(ImageFeatureExtractionMixin.to_pil_imagecCs&||t|tjjs|S|dS)z Converts `PIL.Image.Image` to RGB format. Args: image (`PIL.Image.Image`): The image to convert. rw)rr,r-r.rrr$r$r% convert_rgbs z'ImageFeatureExtractionMixin.convert_rgb)r<scalerCcCs||||S)z7 Rescale a numpy image by scale amount )r)rr<rr$r$r%rs z#ImageFeatureExtractionMixin.rescaleTcCs||t|tjjr"t|}t|r2|}|dkrLt|jdtj n|}|rh| | tj d}|r|j dkr|ddd}|S)a Converts `image` to a numpy array. Optionally rescales it and puts the channel dimension as the first dimension. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image to convert to a NumPy array. rescale (`bool`, *optional*): Whether or not to apply the scaling factor (to make pixel values floats between 0. and 1.). Will default to `True` if the image is a PIL Image or an array/tensor of integers, `False` otherwise. channel_first (`bool`, *optional*, defaults to `True`): Whether or not to permute the dimensions of the image to put the channel dimension first. Nrp?rJrSr )rr,r-r.rErOrr4rintegerrrfloat32rMr)rr<r channel_firstr$r$r%rPs  z*ImageFeatureExtractionMixin.to_numpy_arraycCsB||t|tjjr|St|r0|d}ntj|dd}|S)z Expands 2-dimensional `image` to 3 dimensions. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image to expand. r)Zaxis)rr,r-r.rZ unsqueezerE expand_dimsrr$r$r%r s  z'ImageFeatureExtractionMixin.expand_dimsFcCsv||t|tjjr(|j|dd}n>|rft|tjrN||tj d}nt |rf|| d}t|tjrt|tjst ||j }t|tjst ||j }nrt |r"ddl}t||jst|tjr||}n ||}t||js"t|tjr||}n ||}|jdkrf|jddkrf||ddddf|ddddfS|||SdS)a  Normalizes `image` with `mean` and `std`. Note that this will trigger a conversion of `image` to a NumPy array if it's a PIL Image. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image to normalize. mean (`List[float]` or `np.ndarray` or `torch.Tensor`): The mean (per channel) to use for normalization. std (`List[float]` or `np.ndarray` or `torch.Tensor`): The standard deviation (per channel) to use for normalization. rescale (`bool`, *optional*, defaults to `False`): Whether or not to rescale the image to be between 0 and 1. If a PIL image is provided, scaling will happen automatically. T)rrrNrJrR)rr,r-r.rPrErrrrrfloatrOrDr3ZTensorZ from_numpyZtensorrMrV)rr<ZmeanZstdrr3r$r$r% normalize!s4            (z%ImageFeatureExtractionMixin.normalizec CsT|dk r |ntj}||t|tjjs4||}t|trFt|}t|t s^t |dkrF|rt|t rt||fn|d|df}n|j \}}||kr||fn||f\}} t|t r|n|d} || kr|S| t | | |} } |dk r,|| kr t d|d|| |kr,t || | |} } ||kr>| | fn| | f}|j ||dS)a Resizes `image`. Enforces conversion of input to PIL.Image. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image to resize. size (`int` or `Tuple[int, int]`): The size to use for resizing the image. If `size` is a sequence like (h, w), output size will be matched to this. If `size` is an int and `default_to_square` is `True`, then image will be resized to (size, size). If `size` is an int and `default_to_square` is `False`, then smaller edge of the image will be matched to this number. i.e, if height > width, then image will be rescaled to (size * height / width, size). resample (`int`, *optional*, defaults to `PILImageResampling.BILINEAR`): The filter to user for resampling. default_to_square (`bool`, *optional*, defaults to `True`): How to convert `size` when it is a single int. If set to `True`, the `size` will be converted to a square (`size`,`size`). If set to `False`, will replicate [`torchvision.transforms.Resize`](https://pytorch.org/vision/stable/transforms.html#torchvision.transforms.Resize) with support for resizing only the smallest edge and providing an optional `max_size`. max_size (`int`, *optional*, defaults to `None`): The maximum allowed for the longer edge of the resized image: if the longer edge of the image is greater than `max_size` after being resized according to `size`, then the image is resized again so that the longer edge is equal to `max_size`. As a result, `size` might be overruled, i.e the smaller edge may be shorter than `size`. Only used if `default_to_square` is `False`. Returns: image: A resized `PIL.Image.Image`. Nr rz max_size = zN must be strictly greater than the requested size for the smaller edge size = )r)rBILINEARrr,r-r.rr?r@rUrgrr9resize) rr<rrZdefault_to_squaremax_sizewidthheightshortlongZrequested_new_shortZ new_shortZnew_longr$r$r%rUs0   $    z"ImageFeatureExtractionMixin.resizecCs||t|ts||f}t|s0t|tjrp|jdkrD||}|jddkr`|jddn |jdd}n|j d|j df}|d|dd}||d}|d|dd}||d}t|t j j r| ||||fS|jddkrdnd}|s4t|tjr| ddd}t|r4|ddd}|dkrz||dkrz|dkrz||dkrz|d||||fS|jdd t|d|dt|d|df} t|tjrtj|| d } nt|r|| } | d |dd} | |d} | d |dd} | |d}|| d| | | |f<|| 7}|| 7}|| 7}|| 7}| dtd|t| jd |td|t| jd |f} | S) a Crops `image` to the given size using a center crop. Note that if the image is too small to be cropped to the size given, it will be padded (so the returned result has the size asked). Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor` of shape (n_channels, height, width) or (height, width, n_channels)): The image to resize. size (`int` or `Tuple[int, int]`): The size to which crop the image. Returns: new_image: A center cropped `PIL.Image.Image` or `np.ndarray` or `torch.Tensor` of shape: (n_channels, height, width). rSrrRr NTF.r^)rVr_)rr,r@rrErrMrrVrr-r.ZcroprZpermuterHZ zeros_likeZ new_zerosrG)rr<rZ image_shapetopZbottomleftrightrZ new_shapeZ new_imageZtop_padZ bottom_padZleft_padZ right_padr$r$r% center_cropsZ    ,   02    z'ImageFeatureExtractionMixin.center_cropcCs>||t|tjjr"||}|dddddddfS)a Flips the channel order of `image` from RGB to BGR, or vice versa. Note that this will trigger a conversion of `image` to a NumPy array if it's a PIL Image. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image whose color channels to flip. If `np.ndarray` or `torch.Tensor`, the channel dimension should be first. Nr_)rr,r-r.rPrr$r$r%flip_channel_orders  z.ImageFeatureExtractionMixin.flip_channel_orderrcCsL|dk r |ntjj}||t|tjjs6||}|j||||||dS)a Returns a rotated copy of `image`. This method returns a copy of `image`, rotated the given number of degrees counter clockwise around its centre. Args: image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): The image to rotate. If `np.ndarray` or `torch.Tensor`, will be converted to `PIL.Image.Image` before rotating. Returns: image: A rotated `PIL.Image.Image`. N)rexpandcenter translate fillcolor)r-r.NEARESTrr,rrotate)rr<Zanglerrrrrr$r$r%rs   z"ImageFeatureExtractionMixin.rotate)N)NT)F)NTN)NrNNN)rr r!__doc__rrrrErr rrUrrPrrrrrrr$r$r$r%rs   4 CKr)annotation_formatsupported_annotation_formatsrdrCcCsT||krtdtd||tjkr6t|s6td|tjkrPt|sPtddS)NzUnsupported annotation format: z must be one of zInvalid COCO detection annotations. Annotations must a dict (single image) or list of dicts (batch of images) with the following keys: `image_id` and `annotations`, with the latter being a list of annotations in the COCO format.zInvalid COCO panoptic annotations. Annotations must a dict (single image) or list of dicts (batch of images) with the following keys: `image_id`, `file_name` and `segments_info`, with the latter being a list of annotations in the COCO format.)r9formatr&r'rsr(rt)rrrdr$r$r%validate_annotations s  r)valid_processor_keyscaptured_kwargscCs6t|t|}|r2d|}td|ddS)Nz, zUnused or unrecognized kwargs: rL)set differencejoinrWrX)rrZ unused_keysZunused_key_strr$r$r%validate_kwargs&s r)rJ)N)N)N)N) NNNNNNNNNNNN)Zrriortypingrrrrrrr r4rEr| packagingr utilsr r rrrrrrrrrZutils.constantsrrrrrrZ PIL.Imager-Z PIL.ImageOpsparse __version__ base_versionr.Z ResamplingrZtorchvision.transformsrrZBOXrZHAMMINGZBICUBICZLANCZOSZpil_torch_interpolation_mappingr3Z get_loggerrrWrZ ImageInputZ VideoInputrr&r)ryrUZAnnotationTyper1r2r=r>rArBboolrIrNrPrYr\rarirkrsrtrrxrrrrr$r$r$r%s $ 4           '  &   $$ - *b