§ `ƒi—ãóô—ddlmZddlmZmZddlZddlmZddlm Z ddl m Z ddl m Z mZd d lmZd d lmZeje¦«ZeGd „de ¦«¦«Ze Gd„de ¦«¦«ZdgZdS)é)Ú dataclass)ÚOptionalÚUnionNé)ÚCache)Ú$ImageClassifierOutputWithNoAttention)ÚPreTrainedModel)Úauto_docstringÚloggingé)ÚAutoModelForImageTextToTexté)ÚShieldGemma2Configcó8—eZdZUdZdZeejed<dS)Ú0ShieldGemma2ImageClassifierOutputWithNoAttentionz^ShieldGemma2 classifies imags as violative or not relative to a specific policy Args: NÚ probabilities) Ú__name__Ú __module__Ú __qualname__Ú__doc__rrÚtorchÚTensorÚ__annotations__©óúŠ/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/shieldgemma2/modeling_shieldgemma2.pyrr#s5€€€€€€ððð-1€M8˜EœLÔ)Ð0Ð0Ñ0Ð0Ð0rrc óà‡—eZdZUeed<dddddœZdefˆfd„ Zd„Zd „Zd „Z d „Z d „Z d „Z d„Z e d!deejdeejdeejdeejdeedeejdeejdeejdeejdeedeedeedeedeeejfdefd „¦«ZˆxZS)"Ú"ShieldGemma2ForImageClassificationÚconfigzmodel.model.language_modelzmodel.model.vision_towerz!model.model.multi_modal_projectorz model.lm_head)zmodel.language_model.modelzmodel.vision_towerzmodel.multi_modal_projectorzmodel.language_model.lm_headcóØ•—t¦« |¬¦«t|dd¦«|_t|dd¦«|_t j|¬¦«|_dS)N)rÚyes_token_indexi *Úno_token_indexi»)ÚsuperÚ__init__Úgetattrr!r"r Ú from_configÚmodel)ÚselfrÚ __class__s €rr$z+ShieldGemma2ForImageClassification.__init__6saø€Ý ‰Œ×Ò ÐÑ'Ô'Ð'Ý& vÐ/@À&ÑIÔIˆÔÝ% fÐ.>ÀÑEÔEˆÔÝ0Ô<ÀFÐKÑKÔKˆŒ ˆ ˆ rcó>—|jj ¦«S©N)r'Úlanguage_modelÚget_input_embeddings©r(s rr-z7ShieldGemma2ForImageClassification.get_input_embeddings<s€ØŒzÔ(×=Ò=Ñ?Ô?Ð?rcóD—|jj |¦«dSr+)r'r,Úset_input_embeddings)r(Úvalues rr0z7ShieldGemma2ForImageClassification.set_input_embeddings?s!€Ø Œ Ô!×6Ò6°uÑ=Ô=Ð=Ð=Ð=rcó>—|jj ¦«Sr+)r'r,Úget_output_embeddingsr.s rr3z8ShieldGemma2ForImageClassification.get_output_embeddingsBs€ØŒzÔ(×>Ò>Ñ@Ô@Ð@rcóD—|jj |¦«dSr+)r'r,Úset_output_embeddings)r(Únew_embeddingss rr5z8ShieldGemma2ForImageClassification.set_output_embeddingsEs!€Ø Œ Ô!×7Ò7¸ÑGÔGÐGÐGÐGrcóD—|jj |¦«dSr+)r'r,Ú set_decoder)r(Údecoders rr8z.ShieldGemma2ForImageClassification.set_decoderHs!€Ø Œ Ô!×-Ò-¨gÑ6Ô6Ð6Ð6Ð6rcó>—|jj ¦«Sr+)r'r,Ú get_decoderr.s rr;z.ShieldGemma2ForImageClassification.get_decoderKó€ØŒzÔ(×4Ò4Ñ6Ô6Ð6rcó>—|jj ¦«Sr+)r'r,Ú tie_weightsr.s rr>z.ShieldGemma2ForImageClassification.tie_weightsNr<rNrÚ input_idsÚ pixel_valuesÚattention_maskÚ position_idsÚpast_key_valuesÚtoken_type_idsÚcache_positionÚ inputs_embedsÚlabelsÚ use_cacheÚoutput_attentionsÚoutput_hidden_statesÚ return_dictÚlogits_to_keepÚreturnc óÈ—|jd||||||||| | | | | |dœ|¤Ž}|j}|dd…d|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@rArBrCrDrErFrGrHrIrJrKrLNéÿÿÿÿ)Údim)Úlogitsrr)r'rQr!r"rÚsoftmaxr)r(r?r@rArBrCrDrErFrGrHrIrJrKrLÚ lm_kwargsÚoutputsrQÚselected_logitsrs rÚforwardz*ShieldGemma2ForImageClassification.forwardQs°€ðN$”*ð ØØ%Ø)Ø%Ø+Ø)Ø)Ø'ØØØ/Ø!5Ø#Ø)ð ð ðð ð ˆð"”ˆØ     B¨Ô)=¸tÔ?RÐ(SÐ!SÔTˆÝœ  o¸2Ð>Ñ>Ô>ˆ Ý?Ø"Ø'ð ñ ô ð r)NNNNNNNNNNNNNr)rrrrrÚ_checkpoint_conversion_mappingr$r-r0r3r5r8r;r>r rrÚ LongTensorÚ FloatTensorrrÚboolrÚintrrVÚ __classcell__)r)s@rrr,s)ø€€€€€€à ÐÐÑà&BØ8Ø'JØ(7ð &ð&Ð"ðLÐ1ðLðLðLðLðLðLð @ð@ð@ð>ð>ð>ðAðAðAðHðHðHð7ð7ð7ð7ð7ð7ð7ð7ð7ðð15Ø48Ø15Ø37Ø+/Ø59Ø59Ø59Ø-1Ø$(Ø,0Ø/3Ø&*Ø34ð= ð= à˜EÔ,Ô-ð= ð˜uÔ0Ô1ð= ð! ¤Ô.ð = ð ˜uÔ/Ô0ð = ð " %œð = ð! Ô!1Ô2ð= ð! Ô!1Ô2ð= ð  Ô 1Ô2ð= ð˜Ô)Ô*ð= ð˜D”>ð= ð$ Dœ>ð= ð' tœnð= ð˜d”^ð= ð˜c 5¤<Ð/Ô0ð= ð" :ð#= ð= ð= ñ„^ð= ð= ð= ð= ð= rr)Ú dataclassesrÚtypingrrrÚ cache_utilsrÚmodeling_outputsrÚmodeling_utilsr Úutilsr r Úautor Úconfiguration_shieldgemma2rÚ get_loggerrÚloggerrrÚ__all__rrrúrhsXðð "Ð!Ð!Ð!Ð!Ð!Ø"Ð"Ð"Ð"Ð"Ð"Ð"Ð"à € € € à Ð Ð Ð Ð Ð ØDÐDÐDÐDÐDÐDØ-Ð-Ð-Ð-Ð-Ð-ððððððððð/Ð.Ð.Ð.Ð.Ð.Ø:Ð:Ð:Ð:Ð:Ð:ð ˆÔ ˜HÑ %Ô %€ð ð1ð1ð1ð1ð1Ð7[ñ1ô1ñ „ð1ððb ðb ðb ðb ðb ¨ñb ôb ñ„ðb ðL)ð €€€r