S^(hidZddlmZddlZddlmZddlmZmZm Z m Z ddl Z ddl ZddlmZddlmZmZdd lmZmZmZmZmZmZdd lmZmZmZdd lm Z m!Z!m"Z"m#Z#m$Z$d d l%m&Z&m'Z'm(Z(e#jRe*Z+dZ,dZ-d8d9dZ.d:dZ/d;dZ0eGdde Z1GddejdjfZ4GddejdjfZ5GddejdjfZ6GddejdjfZ7GddejdjfZ8Gdd ejdjfZ9Gd!d"ejdjfZ:eGd#d$ejdjfZ;Gd%d&ejdjfZGd+d,eZ?d-Z@d.ZAd/ZBd0ZCGd1d2e?ZDGd3d4e?ZEe!e@Gd5d6e?ZFgd7ZGy)z(TFCLIPOutput.to_tuple..ws=  LLDGRYZ^`aRbRkRkRm m s-0)tuplekeysrPs`r,rMzTFCLIPOutput.to_tuplevs# YY[   r.)returnz Tuple[Any]) __name__ __module__ __qualname____doc__rB__annotations__rDrErFrGrHrIrMr.r,rArAWs^*"D !,0)0+/O(/'+K$+(,L%,6:3:8<5< r.rAc2eZdZdfd ZdddZddZxZS)TFCLIPVisionEmbeddingsc t|di||j|_|j|_|j |_|j|j zdz|_|j dz|_||_tjj|j|j |j dddt|jj|jjzd|_y) Nrvalid channels_lastFpatch_embedding)filters kernel_sizestridespadding data_formatuse_biaskernel_initializernamer[)super__init__ hidden_size embed_dim image_size patch_size num_patches num_positionsconfigrlayersConv2Drinitializer_rangeinitializer_factorrbrPrskwargs __class__s r,rlzTFCLIPVisionEmbeddings.__init__~s "6"++ ++ ++ OOt>1D!--1 $||22NNOO'.t{{/L/Lt{{OmOm/mn" 3 r.c|jj}|j|jft |jdz|zdd|_t jd5|j|j|jft |jj|zdd|_ ddd|jryd|_ t|ddft j|jj5|jjddd|jj gdddyy#1swYxYw#1swYyxYw)NTclass_embeddingshape initializer trainablerjposition_embedding embeddingsrb)rsrw add_weightrnrr}r" name_scoperrrvrbuiltrLrbrjbuild num_channels)rP input_shapefactors r,rzTFCLIPVisionEmbeddings.builds<//#>>#'(>:+DKK,I,IF,RS! '6'D #  ::  4*D 1 =t33889 Y$$**D$dkk>V>V+WX Y Y >   Y Ys*AE  4E EE cjt|\}}}}tj|d}|j|}tj|||j df}tj |j|d|jf}tj||fd}||jz}|S)z0`pixel_values` is expected to be of NCHW format.)rr_r rpermtensorrr)raxis) rr"r;rbreshaperq broadcast_tor}rnconcatr) rP pixel_values batch_sizerheightwidth patch_embeds class_embedsrs r,callzTFCLIPVisionEmbeddings.calls3=\2J/ L&% ||L|D ++L9 zzj$JZJZ\^=_` t';';JPQSWSaSaCbc YY l;!D $"9"99 r.rsrNrztf.TensorShape)r tf.TensorrUrrVrWrXrlrr __classcell__rzs@r,r]r]}s .Y2r.r]cLeZdZdfd Zddfd Z d ddZxZS) TFCLIPTextEmbeddingsc Tt|di||j|_||_y)Nr[)rkrlrmrnrsrxs r,rlzTFCLIPTextEmbeddings.__init__s' "6"++ r.ctjd5|j|jj|j ft |jj|jjzdd|_ dddtjd5|j|jj|j ft |jj|jjzdd|_ dddt|5|y#1swYxYw#1swY%xYw)Ntoken_embeddingTweightr~rr)r"rrrs vocab_sizernrrwrvrmax_position_embeddingsrrkr)rPrrzs r,rzTFCLIPTextEmbeddings.builds ]], - //{{--t~~>+DKK,J,JT[[MjMj,jk *DK ]]/ 0 &*oo{{::DNNK+DKK,J,JT[[MjMj,jk! '6'D #   k"!    sA/D*#A/D6*D36D?c| | td|At||jjt j |j |}t|dd}|/t jt jd|dd}t j |j|}t j||dddf }||z}|S) z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. Nz5You have to specify either input_ids or inputs_embedsparamsindicesrr)startlimitrr)input multiples) ValueErrorrrsrr"gatherrr expand_dimsr7rr%)rP input_ids position_ids inputs_embedsrposition_embedsfinal_embeddingss r,rzTFCLIPTextEmbeddings.calls  !6TU U  *9dkk6L6L MIIT[[)LM /4  >>"((+b/*RYZ))4+B+BLY''KPQNTUWXCYZ(?:r.rsrrr)NNN)rrCrrCrrCrUrrrs@r,rrsC#**.,0-1  & * +    r.rcVeZdZdZdfd ZddZ d d dZd dZxZS) TFCLIPAttentionz=Multi-headed attention from 'Attention Is All You Need' paperc 8t|d i||j|_|j|_|j|jz|_|j |jz|jk7r&t d|jd|jd|j}|jdzd|jzdzz|z}|jdz|z}tj|j |_ tjj|jt|d|_tjj|jt|d|_tjj|jt|d |_tjj'|j( |_tjj|jt|d |_y) Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).r|r_q_projunitsrirjk_projv_proj)rateout_projr[)rkrlrmrnnum_attention_headsattention_head_sizerrwnum_hidden_layersr3sqrtsqrt_att_head_sizerrtDenserrrrDropoutattention_dropoutdropoutr)rPrsryr in_proj_std out_proj_stdrzs r,rlzTFCLIPAttention.__init__s "6"++#)#=#= #'>>T5M5M#M  # #d&>&> >$.. PMdnnM]^,,-R1  **~~t+V5M5M1MRV0VWZ`` ,6 "&))D,D,D"Ell((.._[5QX`)  ll((.._[5QX`)  ll((.._[5QX`)  ||++1I1I+J  **.._\5RYc+  r.ctj||d|j|jf}tj|gdS)Nrrrr_rr r)r"rrrr;)rPrrs r,transpose_for_scoresz$TFCLIPAttention.transpose_for_scores's;6*b$BZBZ\`\t\t1uv||F66r.c.t|d}|j|}|j|}|j|} |j ||} |j ||} |j | |} t j | | d} t j|j| j}t j| |} |t j| |} |t j| |} t| d}|j||}t j || }t j|gd  }t j||d|j f }|j#|| }|r||f}|S|f}|S) z#Input shape: Batch x Time x ChannelrinputsT transpose_br r)r8r)rtrainingrrr)r)rrrrrr"matmulr$rr!divideaddrrr;rrnr)rP hidden_statesattention_maskcausal_attention_maskoutput_attentionsrrmixed_query_layermixed_key_layermixed_value_layer query_layer key_layer value_layerattention_scoresdk_attention_probsattention_probsattention_outputoutputss r,rzTFCLIPAttention.call.s .q1  KK}K=++]+; KK}K=//0A:N --ozJ //0A:N 99[)N WWT,,4D4J4J K99%5r: ! ,!vv&68MN   %!vv&6G *1AK,,.>,R99_kB<<(8|L::-=jRTVZVdVdEef==)9H=M;L#%56ScQdr.c|jryd|_t|ddZtj|jj 5|jj dd|jgdddt|ddZtj|jj 5|jj dd|jgdddt|ddZtj|jj 5|jj dd|jgdddt|dd[tj|jj 5|jj dd|jgdddyy#1swYAxYw#1swYxYw#1swYxYw#1swYyxYw)NTrrrr) rrLr"rrrjrrnrrrrPrs r,rzTFCLIPAttention.buildcs ::  44 ( 4t{{//0 @ !!4t~~">? @ 44 ( 4t{{//0 @ !!4t~~">? @ 44 ( 4t{{//0 @ !!4t~~">? @ 4T * 6t}}112 B ##T4$@A B B 7 @ @ @ @ @ @ B Bs0 )F32)G)G )G3F=G  GG!rsr)rrrintrUrF rrrrrrrboolrrrUzTuple[tf.Tensor]r) rVrWrXrYrlrrrrrs@r,rrs\G  F7 3 3"3 ) 3  3  3 3jBr.rc0eZdZdfd ZddZddZxZS) TFCLIPMLPc t|di|t|j|_|j }|j dzd|jzdzz|z}d|j zdz|z}tjj|jt|d|_ tjj|j t|d|_||_y)Nr|r_fc1rfc2r[)rkrlr hidden_act activation_fnrwrmrrrtrintermediate_sizerrrrs)rPrsryrrfc_stdrzs r,rlzTFCLIPMLP.__init__vs "6".v/@/@A**))4/Q9Q9Q5QVZ4Z[^dd f(((T1F:<<%%**v?V]b& <<%%$$9U\a&  r.cp|j|}|j|}|j|}|S)Nr)rrr)rPrs r,rzTFCLIPMLP.calls8 6 **=9  6 r.c"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr) rrLr"rrrjrrsrmrrrs r,rzTFCLIPMLP.builds ::  4 % 1txx}}- FdDKK,C,CDE F 4 % 1txx}}- LdDKK,I,IJK L L 2 F F L Ls 3C9<3D9DDr)rrrUrrrVrWrXrlrrrrs@r,rrus" Lr.rcJeZdZdfd Z d ddZddZxZS)TFCLIPEncoderLayerc Nt|di||j|_t |d|_t jj|jd|_ t|d|_ t jj|jd|_ y)N self_attnrj layer_norm1epsilonrjmlp layer_norm2r[)rkrlrmrnrrrrtLayerNormalizationlayer_norm_epsrrr r rxs r,rlzTFCLIPEncoderLayer.__init__s "6"++(kB <<::6CXCX_l:mV%0 <<::6CXCX_l:mr.c|}|j|}|j|||||}|d}||z}|}|j|}|j|}||z}|f|ddz}|S)a Args: hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`tf.Tensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. causal_attention_mask (`tf.Tensor`): causal attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`): Whether or not to return the attentions tensors of all attention layers. See `outputs` under returned tensors for more detail. rrrrrrr)rrN)rrr r ) rPrrrrrresidualattention_outputsrs r,rzTFCLIPEncoderLayer.calls&!(( (>  NN')"7/ + *!,  =0  (( (> }=  =0  "%6qr%::r.c|jryd|_t|ddMtj|jj 5|jj ddddt|ddZtj|jj 5|jj dd|jgdddt|ddMtj|jj 5|jj ddddt|dd[tj|jj 5|jj dd|jgdddyy#1swY4xYw#1swYxYw#1swYxYw#1swYyxYw)NTrrr r ) rrLr"rrrjrrrnr r rs r,rzTFCLIPEncoderLayer.buildsp ::  4d + 7t~~223 +$$T* + 4 - 9t//445 E  &&dDNN'CD E 4 % 1txx}}- %t$ % 4 - 9t//445 E  &&dDNN'CD E E : + + E E % % E Es0 F%)F& F2&)F>F#&F/2F;>Grrrrrrs@r,rrsTn ' '"' ) '  '  ' 'REr.rcVeZdZdZdfd Z d ddZddZxZS) TFCLIPEncoderz Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`TFCLIPEncoderLayer`]. Args: config: CLIPConfig c t|di|t|jDcgc]}t |d|c}|_ycc}w)Nz layers_._rr[)rkrlr7rrrt)rPrsryirzs r,rlzTFCLIPEncoder.__init__sB "6"QVW]WoWoQpqA)&1#Gq qsAc|rdnd}|rdnd} t|jD]+\} } |r||fz}| |||||} | d}|s#| | dfz} -|r||fz}|std||| fDSt||| S)Nr[rrrc3&K|] }|| ywrr[)rNvs r,rQz%TFCLIPEncoder.call..shqZ[Zghs)last_hidden_stater attentions) enumeratertrRr ) rPrrrroutput_hidden_states return_dictrall_hidden_statesall_attentionsr layer_module layer_outputss r,rzTFCLIPEncoder.calls#7BD0d(5 FOA|#$58H$H!(+-&;"3! M*!,M !/=3C2E!E F" 1]4D D h]4E~$Vhh h +;LYg  r.c|jryd|_t|ddK|jD];}tj|j 5|j dddd=yy#1swYIxYw)NTrt)rrLrtr"rrjr)rPrlayers r,rzTFCLIPEncoder.buildsp ::  44 ( 4 &]]5::.&KK%&& & 5&&s A..A7 rr)rrrrrrrrrrrrrrrUz*Union[TFBaseModelOutput, Tuple[tf.Tensor]]r)rVrWrXrYrlrrrrs@r,rrslr&  & "&  ) &  & # & & &  4& P&r.rcpeZdZdfd Z d ddZej fdZddZxZ S) TFCLIPTextTransformerc t|di|t|d|_t |d|_t jj|jd|_ |j|_ |j|_ y)Nrrencoderfinal_layer_normrr[)rkrlrrrr'rrtr r r( eos_token_idrmrnrxs r,rlzTFCLIPTextTransformer.__init__so "6".vLI$V)< % ? ?H]H]dv ? w#//++r.c t|}|j||} |\} } |j| | | j} t |}|j | || ||||} | d}|j |}|jdk(rttj|tjtj|dtjtjj|dfd  }ntj|tjtj|dtjtjjtj||jk(tj dfd  }|s ||f| d dzSt#||| j$| j& S) N)rrr rrrrrrrrrr_rrr)valuesrrr pooler_outputrr)rr_build_causal_attention_maskr!r-r'r(r)r" gather_ndstackr7int64r3argmaxr$int8r rr)rPrrrrrrrrembedding_outputr seq_lengthrencoder_outputssequence_output pooled_outputs r,rzTFCLIPTextTransformer.call*s!+ ??Y\?Z!, J!% A A*j`p`v`v A w&n5,,*)"7/!5#' *!,///G    ! LL&HH[^288DbggnnU^egnFhipqMLL&Qrxx@rwwyDt4499: J%%++T4,HI J J ?  , , ) ) J Js$ D2%D>?)E 2D;>E Err)rr rrrrrrrrrrrrrU5Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]r) rVrWrXrlrr"float32r/rrrs@r,r%r%s ,&E #E "E  E  E # E E E  ?E NJL]$ Jr.r%c|eZdZeZdfd ZddZddZe d d dZ d dZ xZ S) TFCLIPTextMainLayerc Vt|di|||_t|d|_y)N text_modelrr[)rkrlrsr%rHrxs r,rlzTFCLIPTextMainLayer.__init__s( "6" /\Jr.c.|jjSr)rHrrTs r,get_input_embeddingsz(TFCLIPTextMainLayer.get_input_embeddingss)))r.c||jj_t|d|jj_y)Nr)rHrrrr)rPvalues r,set_input_embeddingsz(TFCLIPTextMainLayer.set_input_embeddingss0,1"")0:50A!0D""-r.c | tdt|}|tj|d}|j |||||||} | S)NzYou have to specify input_idsrdimsrLrrrrrrr)rrr"r<rH) rPrrrrrrrrtext_model_outputss r,rzTFCLIPTextMainLayer.callsg  <= = +  !WW+Q?N!__)%/!5#- "!r.c|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTrH)rrLr"rrHrjrrs r,rzTFCLIPTextMainLayer.buildsi ::  4t , 8t334 ,%%d+ , , 9 , , A11A:rrUzkeras.layers.Layer)rLz tf.VariableNNNNNNF)rTFModelInputType | Nonernp.ndarray | tf.Tensor | NonerrXrOptional[bool]rrYrrYrrrUrCr) rVrWrXr config_classrlrJrMrrrrrs@r,rFrFs!LK *E.28<6:,0/3&*"*"6"4 " * " - "$"" ?""<,r.rFcJeZdZdfd Z d ddZddZxZS)TFCLIPVisionTransformerc Nt|di|t|d|_tj j |jd|_t|d|_ tj j |jd|_ |j|_ y)Nrr pre_layrnormrr'post_layernormr[)rkrlr]rrrtr r  pre_layernormrr'r_rmrnrxs r,rlz TFCLIPVisionTransformer.__init__s "6"0lK"\\< , , . . R R N Ns0G$0G1 )G=1)H $G.1G:=H Hc | tdt|}|tj|d}|j |||||||} | d} |j | } | S)N$You have to specify either input_idsrrOrQr)rrr"r<rHrs) rPrrrrrrrr text_outputsr9 text_featuress r,get_text_featuresz!TFCLIPMainLayer.get_text_featuresys  CD D +  !WW+Q?N)%/!5#' %Q ,,M,B r.cz| td|j|||||}|d}|j|}|S)Nrirjrr)rrfrr) rPrrrrrvision_outputsr9image_featuress r,get_image_featuresz"TFCLIPMainLayer.get_image_featuress_  ?@ @**%/!5# + 'q) //}/Er.c | td| tdt|} |tj| d}|j ||||| } |j ||||||| } | d} |j | } | d}|j|}| tj| dd d z } |tj|dd d z }tjj|j}tj|| d |z}tj|}d}|r!t|}tj|d }|s|||| | | f}||f|zS|St!||||| | | S)NrrirrOrjrQr euclideanrT)rordrkeepdimsrr})rBrDrErFrGrHrI)rrr"r<rfrHrrrsnormr3expr~rr;r?rrA)rPrrrr return_lossrrrrrrrrGrFr~rErDrBoutputs r,rzTFCLIPMainLayer.calls  CD D  ?@ @ +  !WW+Q?N**%/!5# + )%/!5#' &a( --\-B "1o **+*> $bgg\{Y[fj&kk !BGG;KVXcg$hh ggkk$"2"23 ))K4PS^^<<8 _-D::dD)D&lT`bpqF'+'7D7V# CV C-+#%* .  r.rrrrVrrWrrXrrXrrYrrYrrYrrrUrrm rrWrrYrrYrrYrrrUr NNNNNNNNFrrWrrWrrXrrXrrYrrYrrYrrYrrrUz%Union[TFCLIPOutput, Tuple[tf.Tensor]]) rVrWrXrrZrlrrrrrrrs@r,roro4sL':RN0.28<6:,0/3&**64  *  - $ B15,0/3&* -*-  $    0.2048<6:&*,0/3&*H *H .H 6 H 4 H $ H *H -H $H H  /H H r.roc$eZdZdZeZdZdgZdgZy)TFCLIPPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. cliprN) rVrWrXrYrrZbase_model_prefix_keys_to_ignore_on_load_missing"_keys_to_ignore_on_load_unexpectedr[r.r,rrs& L'6&7#*9):&r.rav This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior. TensorFlow models and layers in `transformers` accept two formats as input: - having all inputs as keyword arguments (like PyTorch models), or - having all inputs as a list, tuple or dict in the first positional argument. The reason the second format is supported is that Keras methods prefer this format when passing inputs to models and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first positional argument: - a single Tensor with `input_ids` only and nothing else: `model(input_ids)` - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])` - a dictionary with one or several input Tensors associated to the input names given in the docstring: `model({"input_ids": input_ids, "token_type_ids": token_type_ids})` Note that when creating models and layers with [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python function! Args: config ([`CLIPConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the model weights. a Args: input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *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) position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. training (`bool`, *optional*, defaults to `False``): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). a Args: pixel_values (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `(batch_size, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. training (`bool`, *optional*, defaults to `False``): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). al Args: input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for details. [What are input IDs?](../glossary#input-ids) pixel_values (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` `Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `(batch_size, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details. attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *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) position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) return_loss (`bool`, *optional*): Whether or not to return the contrastive loss. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. training (`bool`, *optional*, defaults to `False``): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). ceZdZeZdfd Zeeejde e e d ddZ d dZ xZS) TFCLIPTextModelcPt||g|i|t|d|_yNrr)rkrlrFrrPrsrryrzs r,rlzTFCLIPTextModel.__init__s( 3&3F3'V< r.batch_size, sequence_length output_typerZc 6|j|||||||}|S)aO Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFCLIPTextModel >>> model = TFCLIPTextModel.from_pretrained("openai/clip-vit-base-patch32") >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32") >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="tf") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled (EOS token) states ```rQr) rPrrrrrrrrs r,rzTFCLIPTextModel.calls3>)))%/!5# r.c|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYwNTrrrLr"rrrjrrs r,rzTFCLIPTextModel.builde ::  4 & 2tyy~~. & % & & 3 & &rTrrV)rrWrrXrrXrrYrrYrrYrrYrUrCr)rVrWrXrrZrlrrCLIP_TEXT_INPUTS_DOCSTRINGformatrr rrrrs@r,rrs!L= *+E+L+LMj+kl+GVde.28<6:,0/3&*#(&*&6&4 & * & - &$&!& ?&fm&P&r.rceZdZeZdZdfd Zeee e e e d ddZ d dZ xZS) TFCLIPVisionModelrcPt||g|i|t|d|_yr)rkrlrdrrs r,rlzTFCLIPVisionModel.__init__s( 3&3F3)&v> r.rc2|j|||||}|S)a Returns: Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, TFCLIPVisionModel >>> model = TFCLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="tf") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled CLS states ```rjr)rPrrrrrrs r,rzTFCLIPVisionModel.calls.D))%/!5#  r.c|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYwrrrs r,rzTFCLIPVisionModel.build rrTrrm) rrWrrYrrYrrYrrYrUrCr)rVrWrXrrZmain_input_namerlrrCLIP_VISION_INPUTS_DOCSTRINGrr rrrrs@r,rrs#L$O? *+GH+GVfg15,0/3&*#( '-'*'- ' $ ' ! ' ?'hI'R&r.rceZdZeZd fd Zeeejd d d dZ eee  d d dZ eee jdeee d ddZddZddZxZS) TFCLIPModelcPt||g|i|t|d|_yr)rkrlrorrs r,rlzTFCLIPModel.__init__s( 3&3F3#F8 r.rcH|jj||||||}|S)a Returns: text_features (`tf.Tensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`TFCLIPTextModel`]. Examples: ```python >>> from transformers import AutoTokenizer, TFCLIPModel >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32") >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32") >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="tf") >>> text_features = model.get_text_features(**inputs) ```)rrrrrr)rr) rPrrrrrrrrs r,rzTFCLIPModel.get_text_featuress7: 33)%/!5# 4 r.cD|jj||||}|S)aB Returns: image_features (`tf.Tensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`TFCLIPVisionModel`]. Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, TFCLIPModel >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="tf") >>> image_features = model.get_image_features(**inputs) ```)rrrr)rr)rPrrrrrrs r,rzTFCLIPModel.get_image_featuresEs2B55%/!5# 6 r.rc 8|j||||||||} | S)a Returns: Examples: ```python >>> import tensorflow as tf >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, TFCLIPModel >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor( ... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="tf", padding=True ... ) >>> outputs = model(**inputs) >>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score >>> probs = tf.nn.softmax(logits_per_image, axis=1) # we can take the softmax to get the label probabilities ```)rrrrrrrrr) rPrrrrrrrrrrs r,rzTFCLIPModel.callos7R))%)%#/!5#  r.c|Srr[)rPrs r,serving_outputzTFCLIPModel.serving_outputs  r.c|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYwrrrs r,rzTFCLIPModel.buildrrTrrVrrmrrr)rrArUrAr)rVrWrXrrZrlrrrrrrrCLIP_INPUTS_DOCSTRINGrrArrrrrs@r,rrsL9 *+E+L+LMj+kl.28<6:,0/3&*$*$6$4 $ * $ - $$$$ $m$L*+GH15,0/3&* &-&*&- & $ &  & &I&P*+@+G+GHe+fgrsP" !../RSRML   H %4 6-  " ;"  " JGU\\//GT7 5<<--7 tqBell((qBh!L ""!LHBE++BEJ=&ELL&&=&@qJELL..qJh4,%,,,,4,4,n<Bell00<B~'.ELL..'.'.TH ell((H H V ;- ;(T#J ((V9&+9&x;&-;&|*+\&'\&,\&~ [r.