o wiL@sdZddlZddlZddlmZddlmZmZmZm Z ddl Z ddl Z ddl Z ddl mZddlmZmZmZddlmZdd lmZdd lmZdd lmZdd lmZmZmZdd lm Z m!Z!ddl"m#Z#m$Z$m%Z%m&Z&m'Z'ddl(m)Z)m*Z*m+Z+e&,e-Z.ddZ/ dYde j0de1de1de1de1de j0f ddZ2dZd!d"Z3d#d$Z4d%d&Z5ee$d'd(Gd)d*d*e#Z6ee$d+d(Gd,d-d-e#Z7ee$Gd.d/d/e#Z8Gd0d1d1ej9Z:Gd2d3d3ej9Z; d[d4ej9d5e j0d6e j0d7e j0d8ee j0d9e1d:e1fd;d<Zd>ej9Z=Gd?d@d@ej9Z>GdAdBdBeZ?e$GdCdDdDe!Z@GdEdFdFej9ZAGdGdHdHej9ZBe$dId(GdJdKdKe@ZCGdLdMdMej9ZDGdNdOdOej9ZEe$dPd(GdQdRdRe@ZFe$GdSdTdTe@ZGe$dUd(GdVdWdWe@ZHgdXZIdS)\zPyTorch Siglip model.N) dataclass)AnyCallableOptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)_calculate_fan_in_and_fan_out)ACT2FN)_prepare_4d_attention_mask)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPoolingImageClassifierOutput)ALL_ATTENTION_FUNCTIONSPreTrainedModel) ModelOutputauto_docstringcan_return_tuplelogging torch_int) SiglipConfigSiglipTextConfigSiglipVisionConfigcCsdd}||d|ks||d|krtjddd||||}||||}|d|dd|d|||td|||j||ddS) NcSsdt|tddS)N?@)matherfsqrt)xr$g/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/siglip/modeling_siglip.pynorm_cdf,sz _trunc_normal_..norm_cdfzjmean is more than 2 std from [a, b] in nn.init.trunc_normal_. The distribution of values may be incorrect.) stacklevelrr)minmax) warningswarnuniform_erfinv_mul_r r"add_clamp_)tensormeanstdabr&lur$r$r%_trunc_normal_)s   r9rrr2r3r4r5r6returncCsNtt|dd|||||WddS1s wYdS)anFills the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}( ext{mean}, ext{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for generating the random values works best when :math:`a \leq ext{mean} \leq b`. NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0 and the result is subsequently scaled and shifted by the mean and std args. Args: tensor: an n-dimensional `torch.Tensor` mean: the mean of the normal distribution std: the standard deviation of the normal distribution a: the minimum cutoff value b: the maximum cutoff value rrN)torchno_gradr9r/r0)r2r3r4r5r6r$r$r%trunc_normal_tf_Ms "r?fan_innormalc Cst|\}}|dkr |}n|dkr|}n |dkr||d}||}|dkr3t|t|dddS|dkrWt|jt|dWddS1sPwYdS|d krtd |}t|| |WddS1sywYdStd |) Nr@fan_outfan_avgr'truncated_normalg۶%?r4rAuniformr zinvalid distribution ) r r?r r"r=r>normal_r- ValueError) r2scalemode distributionr@rBdenomvarianceboundr$r$r%variance_scaling_gs(   " "rOcCt|ddddS)Nr@rDrJrKrOr2r$r$r% lecun_normal_rTcCrP)Nr@rArQrRrSr$r$r%default_flax_embed_initrUrVz} Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states. ) custom_introc@jeZdZUdZdZeejed<dZ eejed<dZ ee ejdfed<dZ ee ejdfed<dS)SiglipVisionModelOutputz image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`): The image embeddings obtained by applying the projection layer to the pooler_output. N image_embedslast_hidden_state. hidden_states attentions) __name__ __module__ __qualname____doc__rZrr= FloatTensor__annotations__r[r\tupler]r$r$r$r%rY rYze Base class for text model's outputs that also contains a pooling of the last hidden states. c@rX)SiglipTextModelOutputz text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`): The text embeddings obtained by applying the projection layer to the pooler_output. N text_embedsr[.r\r]) r^r_r`rargrr=rbrcr[r\rdr]r$r$r$r%rfrerfc@seZdZUdZdZeejed<dZ eejed<dZ eejed<dZ eejed<dZ eejed<dZ eed<dZeed <d eefd d ZdS) SiglipOutputa loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`): Contrastive loss for image-text similarity. logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`): The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text similarity scores. logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`): The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image similarity scores. text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`SiglipTextModel`]. image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`SiglipVisionModel`]. text_model_output (`BaseModelOutputWithPooling`): The output of the [`SiglipTextModel`]. vision_model_output (`BaseModelOutputWithPooling`): The output of the [`SiglipVisionModel`]. Nlosslogits_per_imagelogits_per_textrgrZtext_model_outputvision_model_outputr<cstfddDS)Nc3s.|]}|dvr |nt|VqdS))rlrmN)getattrto_tuple).0kselfr$r% s  z(SiglipOutput.to_tuple..)rdkeysrrr$rrr%ros zSiglipOutput.to_tuple)r^r_r`rarirr=rbrcrjrkrgrZrlrrmrdrror$r$r$r%rhs   rhcsXeZdZdeffdd Zdejdededejfdd Zdd ej dejfd d Z Z S)SiglipVisionEmbeddingsconfigcst||_|j|_|j|_|j|_tj|j |j|j|jdd|_ |j|jd|_ |j |_ t |j |j|_|jdt|j ddddS)Nvalid) in_channels out_channels kernel_sizestridepaddingr' position_idsrF persistent)super__init__rw hidden_size embed_dim image_size patch_sizerConv2d num_channelspatch_embedding num_patches num_positions Embeddingposition_embeddingregister_bufferr=arangeexpandrsrw __class__r$r%rs "zSiglipVisionEmbeddings.__init__ embeddingsheightwidthr<c Cs|jd}|jjjd}tjs||kr||kr||jS|jjd}|jd}||j}||j} t |d} | d| | |}| dddd}t j j||| fddd }| dddddd|}|S) a This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution images. This method is also adapted to support torch.jit tracing and no class embeddings. Adapted from: - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211 rrrg?r r'bicubicF)sizerJ align_corners)shaperweightr=jit is_tracingr~ unsqueezerrreshapepermuter functional interpolateview) rsrrrrrpatch_pos_embeddim new_height new_widthsqrt_num_positionsr$r$r%interpolate_pos_encodings&      z/SiglipVisionEmbeddings.interpolate_pos_encodingF pixel_valuesc Csj|j\}}}}|jjj}||j|d}|ddd}|r+|||||}|S|||j }|S)N)dtyper'r) rrrrtoflatten transposerrr~) rsrr_rr target_dtype patch_embedsrr$r$r%forwards zSiglipVisionEmbeddings.forwardF) r^r_r`rrr=Tensorintrrbr __classcell__r$r$rr%rvs &rvc sXeZdZdeffdd Z   d deejdeejdeejdej fd d Z Z S) SiglipTextEmbeddingsrwcsRt|j}t|j||_t|j||_|j dt |j ddddS)Nr~rFr) rrrrr vocab_sizetoken_embeddingmax_position_embeddingsrrr=rrrsrwrrr$r%rs  zSiglipTextEmbeddings.__init__N input_idsr~ inputs_embedsr<cCs|dur |jdn|jd}|jjjd}||kr#td|d||dur2|jddd|f}|dur;||}||}||}|S)NrrzRSequence length must be less than max_position_embeddings (got `sequence length`: z and max_position_embeddings: )rrrrHr~r)rsrr~r seq_lengthmax_position_embeddingposition_embeddingsrr$r$r%r*s"  zSiglipTextEmbeddings.forwardNNN) r^r_r`rrrr= LongTensorrbrrrr$r$rr%rsrmodulequerykeyvalueattention_maskscalingdropoutc Ks|t||dd|}|dur||}tjj|dtjd|j}tjj |||j d}t||} | dd } | |fS)Nrr)rr)ptrainingrr') r=matmulrrrsoftmaxfloat32rrrr contiguous) rrrrrrrkwargs attn_weights attn_outputr$r$r%eager_attention_forwardEs  rc s\eZdZdZfddZ  d dejdeejdeed e ejeejffd d Z Z S) SiglipAttentionz=Multi-headed attention from 'Attention Is All You Need' papercst||_|j|_|j|_|j|j|_|j|j|jkr-td|jd|jd|jd|_ |j |_ d|_ t |j|j|_t |j|j|_t |j|j|_t |j|j|_dS)Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).F)rrrwrrnum_attention_heads num_headshead_dimrHrIattention_dropoutr is_causalrLineark_projv_projq_projout_projrrr$r%r_s$   zSiglipAttention.__init__NFr\routput_attentionsr<c Cs|j\}}}||}||}||} ||||j|jdd}||||j|jdd}| |||j|jdd} t} |j j dkr[|j j dkrU|rUt dnt |j j } | |||| ||j|j|jsjdn|jd\} } | |||} || } |sd} | | fS) z#Input shape: Batch x Time x Channelrr'eagersdpaz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.r:)rrrN)rrrrrrrrrrw_attn_implementationlogger warning_oncerrrIrrrrr) rsr\rr batch_sizerrqueriesruvaluesattention_interfacerrr$r$r%rss:        zSiglipAttention.forward)NF) r^r_r`rarr=rrboolrdrrr$r$rr%r\s rcs2eZdZfddZdejdejfddZZS) SiglipMLPcsDt||_t|j|_t|j|j |_ t|j |j|_ dSN) rrrwr hidden_act activation_fnrrrintermediate_sizefc1fc2rrr$r%rs  zSiglipMLP.__init__r\r<cCs"||}||}||}|Sr)rrr)rsr\r$r$r%rs   zSiglipMLP.forward)r^r_r`rr=rrrr$r$rr%rs rc sVeZdZdeeefffdd Z d dejdejde e de ej fd d Z ZS) SiglipEncoderLayerrwcsRt|j|_tj|j|jd|_t||_ tj|j|jd|_ t ||_ dS)Neps) rrrrr LayerNormlayer_norm_eps layer_norm1r self_attn layer_norm2rmlprrr$r%rs  zSiglipEncoderLayer.__init__Fr\rrr<cCsb|}||}|j|||d\}}||}|}||}||}||}|f}|r/||f7}|S)a= Args: hidden_states (`torch.FloatTensor`): Input to the layer of shape `(batch, seq_len, embed_dim)`. attention_mask (`torch.FloatTensor`): Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*, defaults to `False`): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. )r\rr)rrrr)rsr\rrresidualroutputsr$r$r%rs      zSiglipEncoderLayer.forwardr)r^r_r`rrrrr=rrrrdrbrrr$r$rr%rs rc@s8eZdZeZdZdZgdZdZdZ dZ dZ ddZ dS)SiglipPreTrainedModelsiglipT)rrrvr#SiglipMultiheadAttentionPoolingHeadcCsft|tr%t|jtr|jjjn|jj}tjj|j j dt |ddSt|tj r2t|j dSt|trytj|jj tj|jj tj|jj tj|jj tj|jjtj|jjtj|jjtj|jjdSt|trtj|jj tj|jj tjj|jjddtjj|jjdddSt|trtj|jjtj|jjjtj|jjjdSt|t rt!"t!#d}|j$j%||j&j'dSt|t(rtjj|j)j |jjjd|jj*ddSt|tj+tj,frt-|j |jdurtj|jdSdSt|tj.r1|jj'|j j%ddSdS)zInitialize the weightsrrEgư>rrN)/ isinstancervrwr vision_configrrinitrGrrnpr"rrVrxavier_uniform_rrrrzeros_biasrrrrprobedata attentionin_proj_weight in_proj_bias SiglipModelr=logr2 logit_scalefill_ logit_biaszero_SiglipForImageClassification classifierinitializer_factorrrrTr)rsrrlogit_scale_initr$r$r% _init_weightssX   "          z#SiglipPreTrainedModel._init_weightsN) r^r_r`r config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_supports_flash_attn_2_supports_sdpa_supports_flex_attn_supports_attention_backendrr$r$r$r%rs rc sZeZdZdZdeffdd Ze   d deej dee dee d e fd d Z Z S) SiglipEncoderz Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`SiglipEncoderLayer`]. Args: config: SiglipConfig rwcs:t|_tfddtjD|_d|_dS)Ncsg|]}tqSr$)r)rprrwr$r% 1sz*SiglipEncoder.__init__..F) rrrwr ModuleListrangenum_hidden_layerslayersgradient_checkpointingrrr#r%r.s   zSiglipEncoder.__init__Nrroutput_hidden_statesr<c Cs|dur|n|jj}|dur|n|jj}|rdnd}|rdnd}|}|jD]}|r.||f}||||d} | d}|rB|| df}q%|rJ||f}t|||dS)ad Args: inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. Nr$)rrr)r[r\r])rwrr*r(r) rsrrrr*encoder_statesall_attentionsr\ encoder_layer layer_outputsr$r$r%r5s2     zSiglipEncoder.forwardr)r^r_r`rarrrrr=rrrrrr$r$rr%r"%s r"csreZdZdeffdd Zee     d deej deej deej dee d ee d e f d d Z Z S)SiglipTextTransformerrwcs\t||_|j}t||_t||_tj ||j d|_ t ||j |_|jdk|_dS)Nrflash_attention_2)rrrwrrrr"encoderrrrfinal_layer_normrprojection_sizeheadr_use_flash_attention_2rrr$r%rvs   zSiglipTextTransformer.__init__Nrrr~rr*r<c Cs|dur|n|jj}|dur|n|jj}|durtd|}|d|d}|j||d}|dur<|jsrrrrr$r$r%get_input_embeddings z$SiglipTextModel.get_input_embeddingscCs||jj_dSrrA)rsrr$r$r%set_input_embeddingssz$SiglipTextModel.set_input_embeddingsNrrr~rr*cCs|j|||||dS)a Examples: ```python >>> from transformers import AutoTokenizer, SiglipTextModel >>> model = SiglipTextModel.from_pretrained("google/siglip-base-patch16-224") >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224") >>> # important: make sure to set padding="max_length" as that's how the model was trained >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled (EOS token) states ```rrr~rr*)r>)rsrrr~rr*r$r$r%rszSiglipTextModel.forwardr;)r^r_r`rrrrModulerBrDrrrr=rrrrrr$r$rr%r<s2r<c sXeZdZdeffdd Zee   d deedeedeed e fd d Z Z S) SiglipVisionTransformerrwcsjt||_|j}t||_t||_tj ||j d|_ t |ds%dn|j |_|jr3t||_dSdS)Nrvision_use_headT)rrrwrrvrr"r1rrrpost_layernormhasattrrHuse_headrr4rrr$r%rs   z SiglipVisionTransformer.__init__NFrr*rr<c Cs~|dur|n|jj}|dur|n|jj}|j||d}|j|||d}|j}||}|jr3||nd}t |||j |j dS)N)r)rrr*r6) rwrr*rr1r[rIrKr4rr\r]) rsrrr*rr\r9r[r7r$r$r%rs$  zSiglipVisionTransformer.forwardNNF) r^r_r`rrrrrrrrrr$r$rr%rGs  rGcs.eZdZdZdeffdd ZddZZS)rzMultihead Attention Pooling.rwcs\tttdd|j|_tjj|j|j dd|_ tj |j|j d|_ t||_dS)NrT) batch_firstr)rrr Parameterr=randnrr MultiheadAttentionrr rr layernormrrrrr$r%rs z,SiglipMultiheadAttentionPoolingHead.__init__cCsX|jd}|j|dd}||||d}|}||}|||}|dddfS)Nrr)rr repeatr rQr)rs hidden_staterr rr$r$r%r's  z+SiglipMultiheadAttentionPoolingHead.forward)r^r_r`rarrrrr$r$rr%rsrzM The vision model from SigLIP without any head or projection on top. c sleZdZeZdZdeffdd ZdejfddZ e e   dd e e d e e d e defd dZZS)SiglipVisionModelrrwcr=r)rrrG vision_modelr?rrr$r%r=s   zSiglipVisionModel.__init__r<cCr@r)rUrrrrr$r$r%rBErCz&SiglipVisionModel.get_input_embeddingsNFrr*rcCs|j||||dS)a Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, SiglipVisionModel >>> model = SiglipVisionModel.from_pretrained("google/siglip-base-patch16-224") >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled features ```rrr*r)rU)rsrrr*rr$r$r%rHs zSiglipVisionModel.forwardrL)r^r_r`rrmain_input_namerrrFrBrrrrrrrr$r$rr%rT4s&rTcseZdZeZdeffdd Ze     ddeej deej deej dee d ee d ej f d d Z e    ddeej dee d ee de d ej f ddZ ee        ddeejdeej deej deejdee dee d ee de d efddZZS)rrwcst|t|jtstdt|jdt|jts(tdt|jd|j}|j}t |}t |}|j |_ |j |_ ttd|_ttd|_|dS)NzMconfig.text_config is expected to be of type SiglipTextConfig but is of type .zQconfig.vision_config is expected to be of type SiglipVisionConfig but is of type r)rrr text_configr TypeErrortyperrr< _from_configrTr>rUrrNr=rOrrr?)rsrwrYrr>rUrr$r%rrs,      zSiglipModel.__init__Nrrr~rr*r<cCsF|dur|n|jj}|dur|n|jj}|j|||||d}|j}|S)aJ Returns: text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`SiglipTextModel`]. Examples: ```python >>> from transformers import AutoTokenizer, AutoModel >>> import torch >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224") >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224") >>> # important: make sure to set padding="max_length" as that's how the model was trained >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt") >>> with torch.no_grad(): ... text_features = model.get_text_features(**inputs) ```NrE)rwrr*r>r7)rsrrr~rr* text_outputsr:r$r$r%get_text_featuresszSiglipModel.get_text_featuresFrrcCsD|dur|n|jj}|dur|n|jj}|j||||d}|j}|S)a Returns: image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`SiglipVisionModel`]. Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, AutoModel >>> import torch >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224") >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.no_grad(): ... image_features = model.get_image_features(**inputs) ```NrV)rwrr*rUr7)rsrrr*rvision_outputsr:r$r$r%get_image_featuress!zSiglipModel.get_image_features return_lossc  CsB|dur|n|jj}|dur|n|jj}|j||||d} |j|||||d} | j} | j} | | jdddd} | | jdddd} t| | | j } |j | j |j | j }}| ||} | }d}|rtj| d| j d }t|  d|}tjj|| }tj|dd  }|}t||| | | | | d S) a return_loss (`bool`, *optional*): Whether or not to return the contrastive loss. Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, AutoModel >>> import torch >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224") >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> texts = ["a photo of 2 cats", "a photo of 2 dogs"] >>> # important: we pass `padding=max_length` since the model was trained with this >>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> logits_per_image = outputs.logits_per_image >>> probs = torch.sigmoid(logits_per_image) # these are the probabilities >>> print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'") 31.9% that image 0 is 'a photo of 2 cats' ```NrVrEr'rT)rrkeepdimr)devicer)rirjrkrgrZrlrm)rwrr*rUr>r7normr=rtrrcrrexpeyer ones_likerr logsigmoidsumr3rh)rsrrrr~rarr*rr_r]rZrgrkrrrjrirhm1_diag1logliknllr$r$r%rsP,zSiglipModel.forwardr;)NNNF)NNNNNNNF)r^r_r`rrrrrr=rrrbr^r`rrrhrrr$r$rr%rns  - 0  rz SigLIP vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of the patch tokens) e.g. for ImageNet. csteZdZdZdeddffdd Zee     ddee j dee j d ee d ee d e de f d d Z ZS)rrrwr<NcsZt||j|_t|j}|j|_|jdkr"t|jj |jnt |_ | dS)Nr) rr num_labelsrTr\rrUrrrIdentityrr?)rsrwrUrr$r%r]s  " z%SiglipForImageClassification.__init__Flabelsrr*rc Cs^|dur|n|jj}|dur|n|jj}|j||||d}|j}tj|dd}||}d} |dur||j }|jj durb|j dkrHd|j_ n|j dkr^|j tj ksY|j tjkr^d|j_ nd|j_ |jj dkrt} |j dkrz| ||} n+| ||} n%|jj dkrt} | |d|j |d} n|jj dkrt} | ||} t| ||j|jd S) a$ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Examples: ```python >>> from transformers import AutoImageProcessor, SiglipForImageClassification >>> import torch >>> from PIL import Image >>> import requests >>> torch.manual_seed(3) # doctest: +IGNORE_RESULT >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> # note: we are loading a `SiglipModel` from the hub here, >>> # so the head will be randomly initialized, hence the predictions will be random if seed is not set above. >>> image_processor = AutoImageProcessor.from_pretrained("google/siglip-base-patch16-224") >>> model = SiglipForImageClassification.from_pretrained("google/siglip-base-patch16-224") >>> inputs = image_processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits >>> # model predicts one of the two classes >>> predicted_class_idx = logits.argmax(-1).item() >>> print("Predicted class:", model.config.id2label[predicted_class_idx]) Predicted class: LABEL_1 ```N)rr*rrrd regressionsingle_label_classificationmulti_label_classificationr)rilogitsr\r])rwrr*rUr[r=r3rrrc problem_typerorlongrr squeezer rrrr\r]) rsrrqrr*rrsequence_outputruriloss_fctr$r$r%rosL)     "       z$SiglipForImageClassification.forward)NNNNF)r^r_r`rWrrrrrr=rrrrrr$r$rr%rTs.r)rrr<rTr)r:rr;r)rr@rA)r:)Jrar r+ dataclassesrtypingrrrrnumpyrr=torch.utils.checkpointrtorch.nnrr r torch.nn.initr activationsr modeling_attn_mask_utilsrmodeling_layersrmodeling_outputsrrrmodeling_utilsrrutilsrrrrrconfiguration_sigliprrr get_loggerr^rr9rfloatr?rOrTrVrYrfrhrFrvrrrrrrr"r/r<rGrrTrr__all__r$r$r$r%s       %   #I/ H0AP?305f r