o 0 i@s>ddlZddlZddlmZddlmZddlmZddlm Z m Z ddl Z ddl Z ddlmZmZmZmZmZmZmZmZmZmZmZddlmZmZmZmZmZm Z erddl!Z"ddl#Z"e"j$j%Z&erdd l'm(Z(e&j)e(j*e&j+e(j+e&j,e(j,e&j-e(j-e&j.e(j.e&j/e(j/iZ0niZ0erddl1Z1e2e3Z4e d e j5d e6d e6e j5e6d fZ7Gd d d eZ8GdddeZ9GdddeZ:e;eddZ?GdddeZ@ddZAddZBde6fddZCddZDdd ZEd!d"ZFd#e j5d$eGfd%d&ZHdnd(e=d$e6e7fd)d*ZI 'dnde e6e7e7fd(e=d$e7fd+d,ZJ 'dnde e6e7e7fd(e=d$e6e7fd-d.ZKd$e j5fd/d0ZL dod#e j5d1e e e=eMe=d2ffd$e8fd3d4ZN dod#e j5d5e e e8ed?ZQd@e;ez*is_valid_list_of_images..all)r=r!r!r"is_valid_list_of_imagessrEcCs\t|dtrdd|DSt|dtjrtj|ddSt|dtjr,tj|ddSdS)NrcSg|] }|D]}|qqSr!r!)r@Zsublistitemr!r!r" z$concatenate_list..Zaxis)dim)r*listnpndarrayZ concatenater1Tensorcat)Z input_listr!r!r"concatenate_listsrQcCs:t|ttfr|D] }t|sdSq dSt|sdSdS)NFT)r*rLtuple valid_imagesr<)Zimgsr.r!r!r"rSsrScCst|ttfr t|dSdS)NrF)r*rLrRr<r-r!r!r" is_batcheds rTr:returncCs,|jtjkrdSt|dkot|dkS)zV Checks to see whether the pixel values have already been rescaled to [0, 1]. Frr)dtyperMuint8minmaxr9r!r!r"is_scaled_images rZexpected_ndimscCst|r|St|r |gSt|r9|j|dkrt|}|S|j|kr(|g}|Std|dd|d|jdtdt|d)a Ensure that the output 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. rz%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 .)rTr/r<ndimrLr7r8r=r\r!r!r"make_list_of_imagess*  r`cCst|ttfr tdd|Dr tdd|Dr dd|DSt|ttfrJt|rJt|ds8|dj|kr:|S|dj|dkrJdd|DSt|ret|sW|j|krZ|gS|j|dkret|Std |) a Ensure that the output is a flat list of images. If the input is a single image, it is converted to a list of length 1. If the input is a nested list of images, it is converted to a flat list of images. Args: images (`Union[list[ImageInput], ImageInput]`): The input image. expected_ndims (`int`, *optional*, defaults to 3): The expected number of dimensions for a single input image. Returns: list: A list of images or a 4d array of images. cs|] }t|ttfVqdSr)r*rLrRr@Zimages_ir!r!r"rAz+make_flat_list_of_images..cs|] }t|p | VqdSr)rErcr!r!r"rArdcSrFr!r!r@Zimg_listr.r!r!r"rHrIz,make_flat_list_of_images..rrcSrFr!r!rgr!r!r"rHrIz*Could not make a flat list of images from r*rLrRrDrEr/r^r<r7r_r!r!r"make_flat_list_of_imagess$ ricCst|ttfrtdd|Drtdd|Dr|St|ttfrFt|rFt|ds3|dj|kr6|gS|dj|dkrFdd|DSt|rct|sS|j|krW|ggS|j|dkrct|gStd) as Ensure that the output is a nested list of images. Args: images (`Union[list[ImageInput], ImageInput]`): The input image. expected_ndims (`int`, *optional*, defaults to 3): The expected number of dimensions for a single input image. Returns: list: A list of list of images or a list of 4d array of images. csrar)rbrcr!r!r"rA rdz-make_nested_list_of_images..csrer)rfrcr!r!r"rA rdrrcSsg|]}t|qSr!)rLr?r!r!r"rHsz.make_nested_list_of_images..z]Invalid input type. Must be a single image, a list of images, or a list of batches of images.rhr_r!r!r"make_nested_list_of_imagess$  rjcCs@t|s tdt|trt|tjjrt|St |S)NzInvalid image type: ) r<r7r8rr*r+r,rMarrayrr-r!r!r"to_numpy_arrays  rl num_channels.cCs|dur|nd}t|tr|fn|}|jdkrd\}}n|jdkr&d\}}n|jdkr0d\}}ntd|j|j||vrS|j||vrStd |jd tjS|j||vr]tjS|j||vrgtj Std ) 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. Nrr[r[)r)rorpz(Unsupported number of image dimensions: z4The channel dimension is ambiguous. Got image shape z. Assuming channels are the first dimension. Use the [input_data_format](https://huggingface.co/docs/transformers/main/internal/image_processing_utils#transformers.image_transforms.rescale.input_data_format) parameter to assign the channel dimension.z(Unable to infer channel dimension format) r*intr^r7shapeloggerwarningrrr )r:rmZ first_dimZlast_dimr!r!r"infer_channel_dimension_format(s&       rvinput_data_formatcCsF|durt|}|tjkr|jdS|tjkr|jdStd|)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. Nr[rUnsupported data format: )rvrrr^r r7)r:rwr!r!r"get_channel_dimension_axisOs    ry channel_dimcCsZ|durt|}|tjkr|jd|jdfS|tjkr&|jd|jdfStd|)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. Nrx)rvrrrsr r7)r:rzr!r!r"get_image_sizegs   r~ image_size max_height max_widthc CsB|\}}||}||}t||}t||}t||} || fS)a Computes the output image size given the input image and the maximum allowed height and width. Keep aspect ratio. Important, even if image_height < max_height and image_width < max_width, the image will be resized to at least one of the edges be equal to max_height or max_width. For example: - input_size: (100, 200), max_height: 50, max_width: 50 -> output_size: (25, 50) - input_size: (100, 200), max_height: 200, max_width: 500 -> output_size: (200, 400) Args: image_size (`tuple[int, int]`): The image to resize. max_height (`int`): The maximum allowed height. max_width (`int`): The maximum allowed width. )rXrr) rrrheightwidthZ height_scaleZ width_scaleZ min_scaleZ new_heightZ new_widthr!r!r"#get_image_size_for_max_height_widths   r annotationcCsVt|tr)d|vr)d|vr)t|dttfr)t|ddks't|ddtr)dSdS)Nimage_id annotationsrTFr*dictrLrRlenrr!r!r""is_valid_annotation_coco_detections"rcCs^t|tr-d|vr-d|vr-d|vr-t|dttfr-t|ddks+t|ddtr-dSdS)NrZ segments_info file_namerTFrrr!r!r"!is_valid_annotation_coco_panoptics"rrcCtdd|DS)Ncsr>r))rr@annr!r!r"rArBz3valid_coco_detection_annotations..rCrr!r!r" valid_coco_detection_annotationsrcCr)Ncsr>r))rrr!r!r"rArBz2valid_coco_panoptic_annotations..rCrr!r!r"valid_coco_panoptic_annotationsrrtimeoutc Csttdgt|tre|ds|dr$tjtt j ||dj }nLt j |r1tj|}n?|dr=|dd}zt|}tjt|}Wn!tyd}z td|d |d }~wwt|tjjsptd 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://rz data:image/,rzIncorrect 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 Exceptionr7 TypeErrorZImageOpsZexif_transposeconvert)r:rZb64er!r!r"rs0     rcsXt|ttfr&t|rt|dttfrfdd|DSfdd|DSt|dS)aLoads images, handling different levels of nesting. Args: images: A single image, a list of images, or a list of lists of images to load. timeout: Timeout for loading images. Returns: A single image, a list of images, a list of lists of images. rcsg|] }fdd|DqS)cg|]}t|dqSrrr?rr!r"rHz*load_images...r!)r@Z image_grouprr!r"rHszload_images..crrrr?rr!r"rHrr)r*rLrRrr)r=rr!rr" load_imagess   r do_rescalerescale_factor do_normalize image_mean image_stddo_padpad_sizedo_center_crop crop_size do_resizesizeresamplePILImageResampling interpolationrc Cs|r |dur td|r|durtd|r"|dus|dur"td|r,|dur,td| dur8| dur8td| rJ| durF| dusL| dusNtddSdSdS)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`.zgDepending on the model, `size_divisor` or `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`.zbOnly one of `interpolation` and `resample` should be specified, depending on image processor type.zO`size` and `resample/interpolation` must be specified if `do_resize` is `True`.)r7) rrrrrrrrrrrrrr!r!r"validate_preprocess_argumentss"    rc@seZdZdZddZdddZddZd ejd e e e fd ejfd d Z d ddZ ddZd!ddZd"ddZddZddZd#ddZdS)$ImageFeatureExtractionMixinzD Mixin that contain utilities for preparing image features. cCs8t|tjjtjfst|stdt|ddSdS)Nz Got type zU which is not supported, only `PIL.Image.Image`, `np.ndarray` and `torch.Tensor` are.)r*r+r,rMrNr r7r8selfr:r!r!r"_ensure_format_supported>s z4ImageFeatureExtractionMixin._ensure_format_supportedNcCs||t|r |}t|tjrE|dur t|jdtj}|jdkr3|j ddvr3| ddd}|r9|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. Nrr[rnrro)rr r2r*rMrNflatZfloatingr^rs transposeastyperWr+r,Z fromarray)rr:rescaler!r!r" to_pil_imageEs    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. r)rr*r+r,rrr!r!r" convert_rgbcs  z'ImageFeatureExtractionMixin.convert_rgbr:scalerUcCs||||S)z7 Rescale a numpy image by scale amount )r)rr:rr!r!r"rqs z#ImageFeatureExtractionMixin.rescaleTcCs||t|tjjrt|}t|r|}|dur&t|jdtj n|}|r4| | tj d}|rB|j dkrB|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?r[ror)rr*r+r,rMrkr r2rintegerrrfloat32r^r)rr:r channel_firstr!r!r"rlxs  z*ImageFeatureExtractionMixin.to_numpy_arraycCsD||t|tjjr|St|r|d}|Stj|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. rrJ)rr*r+r,r Z unsqueezerM expand_dimsrr!r!r"rs  z'ImageFeatureExtractionMixin.expand_dimsFcCsh||t|tjjr|j|dd}n|r3t|tjr'||tj d}n t |r3|| d}t|tjrXt|tjsHt ||j }t|tjsWt ||j }n6t |rddl}t||jswt|tjrr||}n||}t||jst|tjr||}n||}|jdkr|jddvr||ddddf|ddddfS|||S)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)rrrNr[rn)rr*r+r,rlrMrNrrrr floatrkrVr1rOZ from_numpyZtensorr^rs)rr:meanZstdrr1r!r!r" normalizes6             ( z%ImageFeatureExtractionMixin.normalizec CsJ|dur|ntj}||t|tjjs||}t|tr#t|}t|t s.t |dkr|rBt|t r9||fn|d|df}n\|j \}}||krO||fn||f\}} t|t r\|n|d} || krf|S| t | | |} } |dur|| krt d|d|| |krt || | |} } ||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`. Nrrz max_size = zN must be strictly greater than the requested size for the smaller edge size = )r)rBILINEARrr*r+r,rrLrRrrrrr7resize) rr:rrZdefault_to_squaremax_sizerrZshortlongZrequested_new_shortZ new_shortZnew_longr!r!r"rs4   $ z"ImageFeatureExtractionMixin.resizecCsx||t|ts||f}t|st|tjr8|jdkr"||}|jddvr0|jddn|jdd}n |j d|j df}|d|dd}||d}|d|dd}||d}t|t j j rr| ||||fS|jddv}|st|tjr| ddd}t|r|ddd}|dkr||dkr|dkr||dkr|d||||fS|jddt|d|dt|d|df} t|tjrtj|| d} n t|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). rorrnrN.r{)rsr|)rr*rRr rMrNr^rrsrr+r,croprZpermuterYZ zeros_likeZ new_zerosrX)rr:rZ image_shapetopbottomleftrightrZ new_shapeZ new_imageZtop_padZ bottom_padZleft_padZ right_padr!r!r" center_crop#sP    ,   (2    4z'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,rlrr!r!r"flip_channel_orderns  z.ImageFeatureExtractionMixin.flip_channel_orderrcCsL|dur|ntjj}||t|tjjs||}|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.rotater))NT)F)NTN)NrNNN)rrr__doc__rrrrMrNrrrrrrlrrrrrrr!r!r!r"r9s "   4CKrannotation_formatsupported_annotation_formatscCsX||vrtdtd||tjurt|std|tjur(t|s*tddSdS)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.)r7formatr$r%rr&r)rrrr!r!r"validate_annotationss  rvalid_processor_keyscaptured_kwargscCs:t|t|}|rd|}td|ddSdS)Nz, zUnused or unrecognized kwargs: r])set differencejoinrtru)rrZ unused_keysZunused_key_strr!r!r"validate_kwargss  rT)frozenc@szeZdZUdZdZeeed<dZeeed<dZ eeed<dZ eeed<dZ eeed<dZ eeed<d d Z dS) SizeDictz> Hashable dictionary to store image size information. Nrr longest_edge shortest_edgerrcCs$t||r t||Std|d)NzKey z not found in SizeDict.)hasattrgetattrKeyError)rkeyr!r!r" __getitem__s  zSizeDict.__getitem__)rrrrrrrr__annotations__rrrrrrr!r!r!r"rs  r)r[r)) NNNNNNNNNNNNN)^rrcollections.abcr dataclassesriortypingrrr2rMrutilsrr r r r r rrrrrZutils.constantsrrrrrrZ PIL.Imager+Z PIL.ImageOpsr,Z ResamplingrZtorchvision.transformsrrZ NEAREST_EXACTZBOXrZHAMMINGZBICUBICZLANCZOSZpil_torch_interpolation_mappingr1Z get_loggerrrtrNrLZ ImageInputrr$r'rrrrZAnnotationTyper/r0r;r<rErQrSrTboolrZr`rirjrlrRrvryr~rrrrrrrrrrrrrr!r!r!r"s6   4        ) ( '  ( &   ""&&$+      5a