o Ďi&@sdZdgZddlZddlmZmZddlmZddlm Z ddl Z ddl Z ddlZddlmmZddlmZddlmZdd lmZdd lmZGd ddZdS) zClassifier functionality of SpeciesNet. Defines the SpeciesNetClassifier class, responsible for image classification for SpeciesNet. It handles loading of classification models, preprocessing of input images, and generating species predictions. SpeciesNetClassifierN)AnyOptional)logging)format_timespan)Failure)BBox) ModelInfo)PreprocessedImagec @seZdZdZdZdZdZ  ddedeedeed dfd d Z  dd ee j j dee e ded eefddZded eed eeeffddZde ede eed e eeeffddZdS)rz#Classifier component of SpeciesNet.ig333333?iN model_nametarget_species_txtdevicereturncs|t}t|_|durtd||_ntjr d_n tj j r*d_nd_tj jj jdd_ j j D]}d|_qCtjjdd d }d d t|D_Wdn1siwY|durt|dd d }fd d|D_Wdn1swYdd jDfddjD_t}tdt||jdS)aLoads the classifier resources. Args: model_name: String value identifying the model to be loaded. It can be a Kaggle identifier (starting with `kaggle:`), a HuggingFace identifier (starting with `hf:`) or a local folder to load the model from. device: Specific device identifier, e.g. "cpu" or "cuda". If None, "cuda" and "mps" will be used if available. Nz Using caller-supplied device %s.cudampscpuF) map_location weights_onlyrzutf-8)modeencodingcSsi|] \}}||qS)strip).0idxlinerrQ/home/jeff/fluffinator/venv/lib/python3.10/site-packages/speciesnet/classifier.py _sz1SpeciesNetClassifier.__init__..cs&g|]}|jvr|qSr)rlabelsvalues)rrselfrr ds z1SpeciesNetClassifier.__init__..cSsi|]\}}||qSrr)rrlabelrrrricg|]}|qSrr)rr#) labels_to_idxrrr"jz(Loaded SpeciesNetClassifier in %s on %s.)timer model_inforinfor torchr is_availablebackendsrload classifiermodeleval parameters requires_gradopenclassifier_labels enumerate readlinesr target_labelsitems target_idxrupper)r!r r r start_timeparamfpend_timer)r&r!r__init__1sB         zSpeciesNetClassifier.__init__Timgbboxesresizec Cs|durdSt|}t|tj}|jjdkrD|rCt|t|dj |j t|dj |j t|dj |j t|dj |j }n |jjdkrdt t|j dtj|j tj}t|||j g}|rrtj|tjtjgdd}t|tj}|gd}t||j |j S) a Preprocesses an image according to this classifier's needs. This method prepares an input image for classification. It handles image loading, cropping, and resizing to the expected input size for the classifier model. In `always_crop` mode images are cropped according to the bounding boxes provided. In `full_image` mode the top and bottom of the image are cropped to prevent learning correlations between camera brand and species priors. See the paper for more details. Args: img: PIL image to preprocess. If `None`, no preprocessing is performed. bboxes: Optional list of bounding boxes. Needed for some types of classifiers to crop the image to specific bounding boxes during preprocessing. resize: Whether to resize the image to some expected dimensions. Returns: A preprocessed image, or `None` if no PIL image was provided initially. N always_cropr full_imageg?F) antialias)r)F pil_to_tensorconvert_image_dtyper+float32r)type_cropintyminheightxminwidthmaxrMAX_CROP_RATIO MAX_CROP_SIZE center_croprCIMG_SIZEuint8permuter numpy)r!rArBrC img_tensor target_heightrrr preprocessss:     zSpeciesNetClassifier.preprocessfilepathcCs||g|gdS)aPRuns inference on a given preprocessed image. Args: filepath: Location of image to run inference on. Used for reporting purposes only, and not for loading the image. img: Preprocessed image to run inference on. If `None`, a failure message is reported back. Returns: A dict containing either the top-5 classifications for the given image (in decreasing order of confidence scores), or a failure message if no preprocessed image was provided. r) batch_predict)r!r_rArrrpredictszSpeciesNetClassifier.predict filepathsimgsc sRig}g}t||D] \}}|dur|tjjgd|<q ||||jdq |s4tStj |dtj d}t | j}|t jdd}t j|ddd \}} tt||| D]5\\}} } |fd d | D| d d |<tdr|djfdd jDdqjfdd |DS)aRuns inference on a batch of preprocessed images. Args: filepaths: List of image locations to run inference on. Used for reporting purposes only, and not for loading the images. imgs: List of preprocessed images to run inference on. If an image is `None`, a corresponding failure message is reported back. Returns: A list of dict results. Each dict result contains either the top-5 classifications for the corresponding image (in decreasing order of confidence scores), or a failure message if no preprocessed image was provided. N)r_failuresr)axisdtype)dim)kricsg|]}j|qSr)rrrr rrr"r$z6SpeciesNetClassifier.batch_predict..)classesscores)r_classificationsr:rocsg|] }t|qSr)floatrl)file_idxlogitsrrr"s)target_classes target_logitscr%rr)rr_) predictionsrrr" r')zipr CLASSIFIERnameappendarrlistrnpstackrLr+ from_numpytor r0rsoftmaxtopkr6r[tolisthasattrupdater8r:) r!rbrcinference_filepaths batch_arrr_rA batch_tensorrnindices scores_arr indices_arrr)rqrrrur!rr`sF      z"SpeciesNetClassifier.batch_predict)NN)NT)__name__ __module__ __qualname____doc__rXrUrVstrrr@PILImager{rboolr r^dictrrar`rrrrr*sR E   A   )r__all__r(typingrrabslr humanfriendlyrr[r| PIL.Imagerr+!torchvision.transforms.functional transforms functionalrIspeciesnet.constantsrspeciesnet.utilsrr r rrrrrs