rhddlmZddlmZmZddlZddlZddlmZddl m Z ddl m Z ddl mZmZd d lmZd d lmZej(eZeGd de ZeGdde ZdgZy)) dataclass)OptionalUnionN)Cache)$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)AutoModelForImageTextToText)ShieldGemma2Configc:eZdZUdZdZeejed<y)0ShieldGemma2ImageClassifierOutputWithNoAttentionz^ShieldGemma2 classifies imags as violative or not relative to a specific policy Args: N probabilities) __name__ __module__ __qualname____doc__rrtorchTensor__annotations__/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/models/shieldgemma2/modeling_shieldgemma2.pyrr$s-1M8ELL)0rrc eZdZUeed<dddddZdeffd ZdZd Zd Z d Z d Z d Z dZ e ddeej"deej$deej&deej"deeeej$efdeej"deej"deej$deej"deedeedeedeedeeej&fdefdZxZS) "ShieldGemma2ForImageClassificationconfigzmodel.model.language_modelzmodel.model.vision_towerz!model.model.multi_modal_projectorz model.lm_head)zmodel.language_model.modelzmodel.vision_towerzmodel.multi_modal_projectorzmodel.language_model.lm_headct||t|dd|_t|dd|_t j ||_y)N)ryes_token_indexi *no_token_indexi)super__init__getattrr!r"r from_configmodel)selfr __class__s rr$z+ShieldGemma2ForImageClassification.__init__7sI '&v/@&I%f.>E0<>@@rcN|jjj|yr+)r'r,set_output_embeddings)r(new_embeddingss rr5z8ShieldGemma2ForImageClassification.set_output_embeddingsFs !!77GrcN|jjj|yr+)r'r, set_decoder)r(decoders rr8z.ShieldGemma2ForImageClassification.set_decoderIs !!--g6rcJ|jjjSr+)r'r, get_decoderr.s rr;z.ShieldGemma2ForImageClassification.get_decoderLzz((4466rcJ|jjjSr+)r'r, tie_weightsr.s rr>z.ShieldGemma2ForImageClassification.tie_weightsOr<r input_ids pixel_valuesattention_mask position_idspast_key_valuestoken_type_idscache_position inputs_embedslabels use_cacheoutput_attentionsoutput_hidden_states return_dictlogits_to_keepreturnc |jd||||||||| | | | | |d|}|j}|ddd|j|jgf}t j |d}t ||S)aY Returns: A `ShieldGemma2ImageClassifierOutputWithNoAttention` instance containing the logits and probabilities associated with the model predicting the `Yes` or `No` token as the response to that prompt, captured in the following properties. * `logits` (`torch.Tensor` of shape `(batch_size, 2)`): The first position along dim=1 is the logits for the `Yes` token and the second position along dim=1 is the logits for the `No` token. * `probabilities` (`torch.Tensor` of shape `(batch_size, 2)`): The first position along dim=1 is the probability of predicting the `Yes` token and the second position along dim=1 is the probability of predicting the `No` token. ShieldGemma prompts are constructed such that predicting the `Yes` token means the content *does violate* the policy as described. If you are only interested in the violative condition, use `violated = outputs.probabilities[:, 1]` to extract that slice from the output tensors. When used with the `ShieldGemma2Processor`, the `batch_size` will be equal to `len(images) * len(policies)`, and the order within the batch will be img1_policy1, ... img1_policyN, ... imgM_policyN. )r?r@rArBrCrDrErFrGrHrIrJrKrLN)dim)logitsrr)r'rQr!r"rsoftmaxr)r(r?r@rArBrCrDrErFrGrHrIrJrKrL lm_kwargsoutputsrQselected_logitsrs rforwardz*ShieldGemma2ForImageClassification.forwardRsN$** %)%+))'/!5#)  " B)=)=t?R?R(S!ST o2> ?"'  r)NNNNNNNNNNNNNr)rrrrr_checkpoint_conversion_mappingr$r-r0r3r5r8r;r>r rr LongTensor FloatTensorrrlistrboolintrrV __classcell__)r)s@rrr-s &B8'J(7 &"L1L @>AH77715481537KO595959-1$(,0/3&*34= E,,-= u001= !. = u//0 = "%U->->(?(F"GH = !!1!12= !!1!12=   1 12= ))*= D>= $D>= 'tn= d^= c5<</0= " :#= = rr) dataclassesrtypingrrrtorch.utils.checkpoint cache_utilsrmodeling_outputsrmodeling_utilsr utilsr r autor configuration_shieldgemma2r get_loggerrloggerrr__all__rrrrjs ""  D-/:   H % 17[1 1b b b L) r