^j[dZddlZddlZddlmZddlmZddlmZddlZddlm Z ddl m Z dd l mZdd lmZdd lmZdd lmZmZdd lmZddlmZddlmZmZmZmZmZm Z ddl!m"Z"m#Z#m$Z$m%Z%m&Z&ejNe(Z)dZ*dZ+dZ,e&e$ze%zZ-edeGddeZ.edeGddeZ/edeGddeZ0Gdde jbZ2Gd d!e jbZ3Gd"d#e jbZ4Gd$d%e jbZ5Gd&d'e jbZ6Gd(d)e jbZ7Gd*d+e jbZ8Gd,d-e jbZ9Gd.d/eZ:Gd0d1e jbZ;Gd2d3e jbZeGd8d9e=Z?eGd:d;e=Z@eGd<d=e=ZAGd>d?e jbZBGd@dAe jbZCGdBdCe jbZDedDGdEdFe=ZEGdGdHe jbZFGdIdJe jbZGGdKdLe jbZHGdMdNe jbZIedOGdPdQe=ZJgdRZKy)SzPyTorch FLAVA model.N) OrderedDict) dataclass)Any)nn)initialization)ACT2FN)create_bidirectional_mask)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPooling)PreTrainedModel)Unpack) ModelOutputTransformersKwargsauto_docstringcan_return_tuplelogging torch_int) FlavaConfigFlavaImageCodebookConfigFlavaImageConfigFlavaMultimodalConfigFlavaTextConfigzfacebook/flava-image-codebookg$(~k@a Output from FlavaModel containing embeddings and outputs from individual encoders. Note that `image_embeddings` and `text_embeddigns` returned are similar to pooled output returned from a transformer. If you want embeddings for contrastive loss or retrieval use a FLAVA model's `image_projection` and `text_projection` layers on `image_embeddings` and `text_embeddings` respectively. ) custom_introceZdZUdZdZej dzed<dZe dzed<dZ ej dzed<dZ e dzed<dZ ej dzed<dZ e dzed<d eefd Zy) FlavaModelOutputa image_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `pixel_values` are present): The image embeddings which are basically the pooled output of [`FlavaImageModel`]. image_output (`BaseModelOutputWithPooling`, *optional*, returned when `pixel_values` are present): The output of the [`FlavaImageModel`]. text_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids` are present): The text embeddings which are basically the pooled output of [`FlavaTextModel`]. text_output (`BaseModelOutputWithPooling`, *optional*, returned when `input_ids` are present): The output of the [`FlavaTextModel`]. multimodal_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids` and `pixel_values` are present and `skip_multimodal_encoder` is `None` or `False`): The multimodal embeddings which are basically the pooled output of [`FlavaTextModel`]. multimodal_output (`BaseModelOutputWithPooling`, returned when `input_ids` and `pixel_values` are present and `skip_multimodal_encoder` is `None` or `False`): The output of the [`FlavaMultimodalModel`]. Nimage_embeddings image_outputtext_embeddings text_outputmultimodal_embeddingsmultimodal_outputreturncHtfdjDS)Nc3dK|]'}|dvr|nt|j)yw))r"r r$Ngetattrto_tuple).0kselfs s/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/flava/modeling_flava.py z,FlavaModelOutput.to_tuple..Vs=  TTDGZabfhiZjZsZsZu u -0tuplekeysr-s`r.r*zFlavaModelOutput.to_tupleUs# YY[   )__name__ __module__ __qualname____doc__rtorch FloatTensor__annotations__r r r!r"r#r$r2rr*r5r.rr4s 26e''$.56:L,t3:04OU&&-459K+d296:5,,t3:;?1D8? %* r5rz@ Class representing pretraining losses from FLAVA model ceZdZUdZdZej dzed<dZej dzed<dZ ej dzed<dZ ej dzed<dZ ej dzed<dZ ej dzed<d e fd Zy) FlavaLossesa mim (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mim_labels` and `pixel_values` are present, `input_ids_masked` is absent and `mim_weight` > 0.): Masked Image Modeling loss as used in BeIT calculated only for unimodal image data. mlm (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mlm_labels` and `input_ids_masked` are present, `pixel_values` is absent and `mlm_weight` > 0.): Masked Language Modeling loss as used in BERT calculated only for unimodal text data. itm (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `itm_labels`, `input_ids_masked`, `pixel_values` are present and `itm_weight` > 0.): Image Text Matching (ITM) loss calculated for paired image-text data. Note that ITM loss is calculated on masked pairs in FLAVA. global_contrastive (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `input_ids` and `pixel_values` are present and `global_contrastive_weight` > 0.): Contrastive loss for image-text similarity similar to CLIP but calculated globally for paired image-text data. This is calculated on unmasked images and texts. mmm_image (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mim_labels`, `pixel_values` and `input_ids_masked` are present and `mmm_image_weight` > 0.): Masked Multimodal Modeling loss's image component calculated on paired image-text data. mmm_text (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mlm_labels`, `pixel_values` and `input_ids_masked` are present and `mmm_text_weight` > 0.): Masked Multimodal Modeling loss's text component calculated on paired image-text data. Nmimmlmitmglobal_contrastive mmm_imagemmm_textr%cBd}|jD] }|d}|S|S)NTF)values)r-all_nonevs r.rHzFlavaLosses.all_none{s5 A}   r5)r6r7r8r9r@r:r;r<rArBrCrDrEboolrHr=r5r.r?r?\s"%)C  T !($(C  T !($(C  T !(37))D07*.Iu  4'.)-He$&-$r5r?a Output from FlavaForPreTraining containing embeddings, and outputs from individual encoders. Note that `image_embeddings` and `text_embeddings` returned are similar to pooled output returned from a transformer. If you want embeddings for contrastive loss or retrieval use a FLAVA model's `image_projection` and `text_projection` layers on `image_embeddings` and `text_embeddings` respectively. ceZdZUdZdZej dzed<dZe ed<dZ ej dzed<dZ e dzed<dZ ej dzed<dZe dzed<dZej dzed <dZe dzed <dZej dzed <dZe dzed <dZej dzed <dZe dzed<dZej dzed<dZe dzed<dZej dzed<dZej dzed<dZej dzed<dZej dzed<dZej dzed<dZej dzed<dZej dzed<deefdZ y)FlavaForPreTrainingOutputay loss (`torch.FloatTensor`, *optional*, returned when `return_loss` is True): Total loss calculated for this model. loss_info (`FlavaLosses`): Detailed info for FLAVA Pretraining losses. Check `FlavaLosses` class description for the information on the keys. image_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `pixel_values` are present): The image embeddings which are basically the pooled output of [`FlavaImageModel`]. image_output (`BaseModelOutputWithPooling`, *optional*, returned when `pixel_values` are present): The output of the [`FlavaImageModel`]. text_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids` are present): The text embeddings which are basically the pooled output of [`FlavaTextModel`]. text_output (`BaseModelOutputWithPooling`, *optional*, returned when `input_ids` are present): The output of the [`FlavaTextModel`]. multimodal_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids` and `pixel_values` are present and `skip_unmasked_multimodal_encoder` is `None` or `False`): The multimodal embeddings which are basically the pooled output of [`FlavaTextModel`]. multimodal_output (`BaseModelOutputWithPooling`, returned when `input_ids` and `pixel_values` are present and `skip_unmasked_multimodal_encoder` is `None` or `False`): The output of the [`FlavaMultimodalModel`]. image_masked_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `pixel_values` are present): The image embeddings which are basically the pooled output of [`FlavaImageModel`]. Uses `bool_masked_pos` to create masked images. image_masked_output (`BaseModelOutputWithPooling`, *optional*, returned when `pixel_values` are present): The output of the [`FlavaImageModel`]. Uses `bool_masked_pos` to create masked images. text_masked_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids_masked` are present): The text embeddings which are basically the pooled output of [`FlavaTextModel`]. text_masked_output (`BaseModelOutputWithPooling`, *optional*, returned when `input_ids_masked` are present): The output of the [`FlavaTextModel`]. multimodal_masked_embeddings (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*, returned when `input_ids` and `pixel_values` are present): The multimodal embeddings which are basically the pooled output of [`FlavaTextModel`]. multimodal_masked_output (`BaseModelOutputWithPooling`, *optional*, returned when `input_ids_masked` and `pixel_values` are present): The output of the [`FlavaMultimodalModel`]. mim_logits (`torch.FloatTensor` of shape `(batch_size, num_image_patches, image_vocab_size)` or of shape `(total_masked_patches, image_vocab_size)` , *optional*, returned when `pixel_values` are present and `input_ids_masked` are not): The logits for MIM unimodal loss. Uses `book_masked_pos` to get masked patches. The flattened output is returned when `bool_masked_pos` has some of the patches masked. mlm_logits (`torch.FloatTensor` of shape `(batch_size, text_seq_length, text_vocab_size)` or of shape `(total_masked_seq_length, text_vocab_size)`, *optional*, returned when `input_ids_masked` are present and `pixel_values` are not): The logits for MLM unimodal loss. The flattened output is returned when `input_ids_masked` has some of the tokens masked. itm_logits (`torch.FloatTensor` of shape `(batch_size, 2)`, *optional*, returned when `input_ids_masked` and `pixel_values` are present): The logits for ITM loss. Note that ITM loss is calculated on masked pairs in FLAVA. contrastive_logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`): The scaled dot product scores between `image_embeddings` and `text_embeddings` but passed through FLAVA's `image_projection` and `text_projection` layers respectively. This represents the image-text similarity scores. This is calculated on unmasked images and texts. contrastive_logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`): The scaled dot product scores between `text_embeddings` and `image_embeddings` but passed through FLAVA's `text_projection` and `image_projection` layers respectively. This is calculated on unmasked images and texts. mmm_image_logits (`torch.FloatTensor` of shape `(batch_size, num_image_patches, image_vocab_size)` or of shape`(total_masked_patches, image_vocab_size)`, *optional*, returned when `pixel_values` and `input_ids_masked` are present): The logits for MMM image multimodal loss. Uses `book_masked_pos` to get masked patches. The flattened output is returned when `bool_masked_pos` has some of the patches masked. mmm_text_logits (`torch.FloatTensor` of shape `(batch_size, text_seq_length, text_vocab_size)` or of shape `(`(total_masked_seq_length, text_vocab_size)`), *optional*, returned when `pixel_values` and `input_ids_masked` are present): The logits for MMM text multimodal loss. The flattened output is returned when `input_ids_masked` has some of the tokens masked. Nloss loss_inforr r!r"r#r$image_masked_embeddingsimage_masked_outputtext_masked_embeddingstext_masked_outputmultimodal_masked_embeddingsmultimodal_masked_output mim_logits mlm_logits itm_logitscontrastive_logits_per_imagecontrastive_logits_per_textmmm_image_logitsmmm_text_logitsr%cTgdtfdjDS)N)r"r r$rRrPrTc3dK|]'}|vr|nt|j)ywNr()r+r,r-transformer_outputss r.r/z5FlavaForPreTrainingOutput.to_tuple..s4sbc)< YY J7Q?  $#d&C&CJPVX]&^^J#d&>&>>J\\*- r5FNF)r6r7r8r9rrJrjr:Tensorintr BoolTensorr __classcell__rws@r.raras/RV&&D5<<&D&DUX&D]b]i]i&DV48). ll))D0#'   r5rac eZdZdZ d deeezeeefzdeeeefzdedeffd Zd dejde d ejfd Z xZ S) rpz# Image to Patch Embedding. rerfrgrhcVt|t|tjj s||f}t|tjj s||f}|d|dz|d|dzz}||_||_||_tj|||||_ y)Nrr) kernel_sizestride) rirj isinstance collectionsabcIterablererfrrrConv2d projection)r-rerfrgrhrrrws r.rjzPatchEmbeddings.__init__Ss *koo&>&>?$j1J*koo&>&>?$j1J!!} 1 5*Q-:VW=:XY $$&))L)\fgr5rrr%c 8|j\}}}}|sV||jdk7s||jdk7r2td|d|d|jdd|jdd |j|j dj dd}|S)NrrzInput image size (*z) doesn't match model (z).r})rre ValueErrorrflatten transpose)r-rrrrgryrzxs r.rzPatchEmbeddings.forwardfs2>2D2D/ L&%'++u8J/J (%9+,Adooa.@-AE OOL ) 1 1! 4 > >q! Dr5)rir) r6r7r8r9rlistr2rjr:rrJrrrs@r.rprpNs 9<,. h$s)OeCHo5h%S/)h h  h& ELL D ]b]i]i r5rpceZdZdZfdZ ddej dzdej dzdej dzfdZxZS) FlavaTextEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.ct|tj|j|j |j |_tj|j|j |_ tj|j|j |_ tj|j |j|_ tj|j|_|j#dt%j&|jj)dd|j#dt%j*|j,j/t$j0dy) N) padding_idxeps position_idsrr|F) persistenttoken_type_ids)dtype)rirjr Embedding vocab_sizern pad_token_idword_embeddingsmax_position_embeddingsrstype_vocab_sizetoken_type_embeddings LayerNormlayer_norm_epsrtrurvregister_bufferr:arangerrmrrlongr-rbrws r.rjzFlavaTextEmbeddings.__init__us !||F,=,=v?Q?Q_e_r_rs#%<<0N0NPVPbPb#c %'\\&2H2H&J\J\%]"f&8&8f>S>STzz&"<"<=   ELL)G)GHOOPWXej    ekk$*;*;*@*@*B%**Ubg  r5N input_idsrrc|j}|d}||jddd|f}|st|dr-|jddd|f}|j |d|}|}n:t j |t j|jj}|j|}|j|} || z} |j|} | | z } |j| } |j| } | S)Nrrr)rdevice)rrhasattrrrr:rmrrrrrsrrv) r-rrr input_shape seq_lengthbuffered_token_type_ids buffered_token_type_ids_expanded inputs_embedsrrxrss r.rzFlavaTextEmbeddings.forwards  nn&  ^  ,,Q  ^ J"%:: "66|D)) ^^J/ \\*- r5NNN) r6r7r8r9rjr:rrrrs@r.rrrsRQ $*..2,0 <<$&  t+ llT) r5rc eZdZdeddffd Z d dej dej dzdedeej ej feej zfdZ xZ S) FlavaSelfAttentionrbr%Nct||j|jzdk7r2t |ds&t d|jd|jd|j|_t |j|jz |_|j|jz|_tj|j|j|j|_ tj|j|j|j|_ tj|j|j|j|_tj|j |_y)Nrembedding_sizezThe hidden size z4 is not a multiple of the number of attention heads .bias)rirjrnnum_attention_headsrrrattention_head_size all_head_sizerLinearqkv_biasquerykeyvaluertattention_probs_dropout_probrvrs r.rjzFlavaSelfAttention.__init__s.     : : :a ?PVXhHi"6#5#5"67334A7  $*#=#= #&v'9'9FHHAN jj/44\BLLQPQR !<< Y5H5HR5PQ+dii8P8P.QQ  %/.@ --//0@b/I,,7 _kB %--aAq9DDF "/"4"4"6s";t?Q?Q>S"S* **,CD 6G=/2O\M]r5r r6r7r8FlavaPossibleConfigsrjr:rrJr2rrrs@r.rrszG3GG*/3"' #||# t+# # u||U\\) *U5<<-@ @ #r5rc|eZdZdZdeddffd ZdejdejdejfdZxZ S) FlavaSelfOutputz The residual connection is defined in FlavaLayer (same as ViTLayer) instead of here (as is the case with other models), due to the layernorm applied before each block. rbr%Nct|tj|j|j|_tj |j|_yr^) rirjrrrndensertrurvrs r.rjzFlavaSelfOutput.__init__sB YYv1163E3EF zz&"<"<= r5r input_tensorcJ|j|}|j|}|Sr^rrvr-rrs r.rzFlavaSelfOutput.forwards$ =1  ]3 r5) r6r7r8r9rrjr:rrrrs@r.rrsE >3>> U\\RWR^R^r5rc eZdZdeddffd Z d dej dej dzdedeej ej feej zfdZ xZ S) FlavaAttentionrbr%Ncbt|t||_t ||_yr^)rirjr attentionroutputrs r.rjzFlavaAttention.__init__s& +F3%f- r5rrrcj|j|||}|j|d|}|f|ddz}|SN)rrrr)r r)r-rrr self_outputsattention_outputrs r.rzFlavaAttention.forwardsO ~~ .L]&  ;;|A F#% QR(88r5rrrs@r.r r sw.3../3"' ||  t+   u||U\\) *U5<<-@ @ r5r c`eZdZdeddffd Zdej dej fdZxZS)FlavaIntermediaterbr%Nct|tj|j|j |_t|jtrt|j|_ y|j|_ yr^) rirjrrrnintermediate_sizerr hidden_actstrr intermediate_act_fnrs r.rjzFlavaIntermediate.__init__ s] YYv1163K3KL f'' -'-f.?.?'@D $'-'8'8D $r5rcJ|j|}|j|}|Sr^)rrr-rs r.rzFlavaIntermediate.forwards& =1 00? r5 r6r7r8rrjr:rrrrs@r.rr s29399U\\ellr5rcxeZdZdeddffd Zdej dej dej fdZxZS) FlavaOutputrbr%Nct|tj|j|j |_tj|j|_ yr^) rirjrrrrnrrtrurvrs r.rjzFlavaOutput.__init__sB YYv779K9KL zz&"<"<= r5rrcT|j|}|j|}||z}|Sr^rr s r.rzFlavaOutput.forward!s. =1  ]3 % 4 r5rrs@r.rrs@>3>> U\\RWR^R^r5rc eZdZdZdeddffd Z d dejdejdzdede ejejfe ejzfd Z xZ S) FlavaLayerz?This corresponds to the Block class in the timm implementation.rbr%Ncrt||j|_d|_t ||_t ||_t||_ tj|j|j|_tj|j|j|_y)Nrr)rirjchunk_size_feed_forward seq_len_dimr r r intermediaterrrrrnrlayernorm_beforelayernorm_afterrs r.rjzFlavaLayer.__init__-s '-'E'E$'/-f5!&) !# V-?-?VEZEZ [!||F,>,>FDYDYZr5rrrc|j|j|||}|d}|dd}||z}|j|}|j|}|j ||}|f|z}|Sr)r r'r(r&r)r-rrrself_attention_outputsrr layer_outputs r.rzFlavaLayer.forward9s "&  ! !- 0)/"0"  2!4(,)=8 ++M: ((6 {{<? /G+r5r) r6r7r8r9rrjr:rrJr2rrrs@r.r"r"*s}I [3 [ [/3"' || t+  u||U\\) *U5<<-@ @ r5r"ceZdZdeddffd Z d dej dej dzdeded edee zf d Z xZ S) FlavaEncoderrbr%Nct|||_tjt |j Dcgc] }t|c}|_d|_ ycc}wr) rirjrbr ModuleListrangenum_hidden_layersr"layergradient_checkpointing)r-rbrrws r.rjzFlavaEncoder.__init__WsN  ]]fF^F^@_#`1Jv$6#`a &+#$asA#rrroutput_hidden_states return_dictc|rdnd}|rdnd}t|jD](\}} |r||fz}| |||} | d}|s || dfz}*|r||fz}|std|||fDSt|||S)Nr=rrc3&K|] }|| ywr^r=)r+rIs r.r/z'FlavaEncoder.forward..wsmq_`_lm)last_hidden_stater attentions) enumerater2r2r ) r-rrrr4r5all_hidden_statesall_self_attentionsi layer_module layer_outputss r.rzFlavaEncoder.forward]s#7BD$5b4(4 POA|#$58H$H!(HYZM)!,M &9]1=MS>ST->k&)D  r5r%c.|jjSr^rxrqr4s r.get_input_embeddingsz$FlavaImageModel.get_input_embeddingss///r5rc&||j_yr^rmr-rs r.set_input_embeddingsz$FlavaImageModel.set_input_embeddingss+0(r5Nrrrrr4r5c ||n|jj}||n|jj}||n|jj}| t d|j |||} |j | ||||} | d} |j| } |j|j| nd} |s | | f| ddzSt| | | j| jS)z bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, image_num_patches)`): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Nz You have to specify pixel_values)rrrrr4r5rrr9 pooler_outputrr:) rbrr4r5rrxrgrhrir rr:) r-rrrrrr4r5kwargsembedding_outputencoder_outputssequence_outputrIs r.rzFlavaImageModel.forwards 2C1N-TXT_T_TqTq$8$D $++JjJj &1%S>ST->k&)D  r5r%c.|jjSr^rxrr4s r.rnz#FlavaTextModel.get_input_embeddingss...r5rc&||j_yr^rrps r.rqz#FlavaTextModel.set_input_embeddingss*/'r5Nrrrrrr4r5c ||n|jj}||n|jj}||n|jj}| t d|j |||} t |j| |}|j| ||||} | d} |j| } |j|j| nd} |s | | f| ddzSt| | | j| jS) a input_ids (`torch.LongTensor` of shape `(batch_size, text_seq_length)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) token_type_ids (`torch.LongTensor` of shape `(batch_size, text_seq_length)`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) NzYou have to specify input_ids)rrrrbrrrsrrrt) rbrr4r5rrxr rgrhrir rr:) r-rrrrrr4r5rvrwrxryrIs r.rzFlavaTextModel.forwards102C1N-TXT_T_TqTq$8$D $++JjJj &1%S>ST->k&)D  r5Nrrr4r5r%c x||n|jj}||n|jj}||n|jj}|j \}}} |j r;|j j|dd} tj| |fd}|dz }t|j||}|j|||||} | d} |j| } |j|j| nd} |s | | f| ddzSt| | | j| j S) z hidden_states (`torch.FloatTensor` of shape `(batch_size, image_num_patches + text_seq_len, hidden_size)`): The concatenated hidden states of unimodal encoders. Nr|rrrrsrrt)rbrr4r5rrWrorr:rr rgrhrir rr:)r-rrrr4r5rvrrrrrxryrIs r.rzFlavaMultimodalModel.forwardssT2C1N-TXT_T_TqTq$8$D $++JjJj &1%++,-Q0  &--/?@,,-.a1  &224IJ_V%=%= >?qAB  (( ** "44$33 + 7 7!-!9!9/;;(5*<8 45F G " $*@*@$BUBU V!yy)>)>@S@ST<< T[[5W5W(XY&(ii0F0FH[H[&\#%'YYt/D/DdFYFY%Z" r5Nrrrrrvr%c z|jd||||dd|}|j}|j||_|S)a  input_ids (`torch.LongTensor` of shape `(batch_size, text_seq_length)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) token_type_ids (`torch.LongTensor` of shape `(batch_size, text_seq_length)`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) Examples: ```python >>> import torch >>> from transformers import AutoProcessor, FlavaModel >>> model = FlavaModel.from_pretrained("{0}") >>> processor = AutoProcessor.from_pretrained("{0}") >>> inputs = processor( ... text=["a photo of a cat", "a photo of a dog"], max_length=77, padding="max_length", return_tensors="pt" ... ) >>> with torch.inference_mode(): ... text_features = model.get_text_features(**inputs) ``` T)rrrrr5r=)rr9rru)r-rrrrrv text_outputsr9s r.get_text_featureszFlavaModel.get_text_featuress\L4C4??4 ))% 4  4 )::%)%9%9:K%L "r5rrrc z|jd||||dd|}|j}|j||_|S)a  bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, image_num_patches)`): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Examples: ```python >>> import torch >>> from transformers import AutoProcessor, FlavaModel >>> from transformers.image_utils import load_image >>> model = FlavaModel.from_pretrained("{0}") >>> processor = AutoProcessor.from_pretrained("{0}") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = load_image(url) >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.inference_mode(): ... image_features = model.get_image_features(**inputs) ``` T)rrrrr5r=)rr9rru)r-rrrrrv image_outputsr9s r.get_image_featureszFlavaModel.get_image_features s^B5ED4D4D5 %+)%= 5  5 *;;&*&;&;>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoProcessor, FlavaModel >>> model = FlavaModel.from_pretrained("facebook/flava-full") >>> processor = AutoProcessor.from_pretrained("facebook/flava-full") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> inputs = processor(text=["a photo of a cat"], images=image, return_tensors="pt", padding=True) >>> outputs = model(**inputs) >>> image_embeddings = outputs.image_embeddings >>> text_embeddings = outputs.text_embeddings >>> multimodal_embeddings = outputs.multimodal_embeddings >>> outputs.image_embeddings.shape torch.Size([1, 197, 768]) >>> text_embeddings.shape torch.Size([1, 7, 768]) >>> multimodal_embeddings.shape torch.Size([1, 205, 768]) ``` NzRFLAVA model requires hidden states to work. Please set `output_hidden_states=True`)rrrrr4r5rr}r|)rrrrrr4r5rrr)rr5)rr r!r"r#r$)rbr5rrrrrrrrWr:onesrrr)r-rrrrrrrrrr4r5rvr image_statesimage_mm_projectionr r! text_statestext_mm_projectionr"r#r$rrrattention_mask_imageattention_multimodalmultimodal_inputs r.rzFlavaModel.forward:s L&1%;q>[O!%!;!;KO!L  $   */A/MVm))<)B)B& GQ((66qLG',zz*gNaNhNh'i$',yy2F1W]^'_$'+$$yy*=?Q)RXYZ  $ 5 5 1ES^!6! %6a$8 ! %!   -%+#"7/   r5r) NNNNNNNNNTN)r6r7r8rr<rjrrr:rrrr2r rrrJr LongTensorr;rrrrs@r.rXrXs? ){)V/3.2,0 /<</ t+/ t+ / llT) / +, / + +//b4804.2 *ll*))D0*#'+ *  t+ * +, * + +**X.215.2.2/30448/3)-%)#'N ##d*N ''$.N  t+ N  t+ N , N &&-N $llT1N "&N  $;N #N D[N  ! !N N r5rXc`eZdZdedeffd Zdej dej fdZxZS)FlavaImageCodebookResPathin_sizeout_sizec t||dz}t}tj|d<tj ||dd|d<tj|d<tj ||dd|d<tj|d <tj ||dd|d <tj|d <tj ||dd |d <tj ||_y)Nrelu_1rrrpaddingconv_1relu_2conv_2relu_3conv_3relu_4rconv_4)rirjrrReLUr Sequentialpath)r-rrrvhid_sizerrws r.rjz"FlavaImageCodebookResPath.__init__s q=}X7H!QOXX8X1aPXX8X1aPXX8X1aPXMM$' r5rr%c$|j|Sr^)rr-rs r.rz!FlavaImageCodebookResPath.forwardsyy|r5 r6r7r8rrjr:rrrrs@r.rrs1((s( %,,r5rcdeZdZdededeffd Zdej dej fdZxZS)FlavaImageCodebookBlockrr num_layersc t|d|dzz |_||k7rtj||dd|_ntj |_t|||_y)Nrr}rr) rirj post_gainrrid_pathIdentityrres_path)r-rrrrvrws r.rjz FlavaImageCodebookBlock.__init__sW j!m, h 99WhAqQDL;;=DL1'8D r5rr%cb|j||j|j|zzSr^)rrrrs r.rzFlavaImageCodebookBlock.forwards'||A$--2B!BBBr5rrs@r.rrs? E Es E ECC%,,Cr5rc neZdZd dededededef fd Zdejdejfd ZxZ S) FlavaImageCodebookLayerGroup num_blocksrrruse_poolc$t|t}t|D]4}|dk(rt ||||d|dz<t ||||d|dz<6|rt j d|d<t j||_y)Nrblock_rr})rpool) rirjrr0rr MaxPool2drgroup) r-rrrrrblocksr>rws r.rjz%FlavaImageCodebookLayerGroup.__init__s z" cAAv+B7HV`+aAw'(+B8XWa+bAw'(  c \\a8F6N]]6* r5rr%c$|j|Sr^)rrs r.rz$FlavaImageCodebookLayerGroup.forwardszz!}r5rz) r6r7r8rrJrjr:rrrrs@r.rrsH +3 +C +# +QT +`d +%,,r5ra" The FLAVA's image codebook model inspired from DALL-E's original encoder. Outputs raw hidden states and can be used to generate image tokens for an image based on DALL-E's vocab. Used to generate labels for MIM. Use `get_codebook_indices` to get image tokens for an image. ceZdZUdZeed<dZdZdZdede ffd Z de jde jfd Z de jde jfd Zde jde jfd ZxZS) FlavaImageCodebookmodelrbrrbFrvc t||||_|j|_|j|_|j |_|j |_|j|_|j|j z}t}tj|d<tjd|j z|jdd|d<t}tj|jd|j zdd|d <t|j |d|j zd|j z|d <t|j |d|j zd |j z|d <t|j |d |j zd |j z|d<t|j |d |j zd|j zd|d<tj||d<tj||_|j|jj r|j#D] }d|_ yy)Nrelurrrconvrinputgroup_1r}group_2rgroup_3F)rgroup_4r)rirjrb num_groupsinput_channelsnum_blocks_per_grouprnrrrrrrrrrjfreeze parameters requires_grad)r-rbrvr output_blocksrparamrws r.rjzFlavaImageCodebook.__init__s    ++$33$*$?$?!!-- ++__t'@'@@ # " f " !d.>.>*>]^hi j f))D$7$7T=M=M9M[\fghw8  % %z1t7G7G3GTM]M]I] y9  % %z1t7G7G3GTM]M]I] y9  % %z1t7G7G3GTM]M]I] y9  % %z1t7G7G3GTM]M]I]hm y==7xmmF+   ;;  * ,&+# , r5r%ctdtdtd|j|}tj|dS)NaI Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Codebook pixel values can be obtained using [`AutoImageProcessor`] by passing `return_codebook_pixels=True`. See [`FlavaImageProcessor.__call__`] for details. Examples: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoImageProcessor, FlavaImageCodebook >>> model = FlavaImageCodebook.from_pretrained("E") >>> image_processor = AutoImageProcessor.from_pretrained("a") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> inputs = image_processor([image], return_codebook_pixels=True, return_tensors="pt") >>> inputs = dict(pixel_values=inputs.codebook_pixel_values) >>> outputs = model.get_codebook_indices(**inputs) ``` r)axis)_CHECKPOINT_FOR_CODEBOOK_DOCrr:argmaxr-rz_logitss r.get_codebook_indicesz'FlavaImageCodebook.get_codebook_indices@sI 9:V8VWCC_B` a  4;;|,||H1--r5c\|j|}tjd|S)Nrr)rrSoftmaxrs r.get_codebook_probsz%FlavaImageCodebook.get_codebook_probs^s&;;|, rzza **r5c 0dtdtdt|jdk7rtd|jd|jd|jk7r(td|jdd |j|j |S) NaJ Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Codebook pixel values can be obtained using [`AutoImageProcessor`] by passing `return_codebook_pixels=True`. See [`FlavaImageProcessor.__call__`] for details. Examples: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoImageProcessor, FlavaImageCodebook >>> model = FlavaImageCodebook.from_pretrained("ra") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> inputs = image_processor([image], return_codebook_pixels=True, return_tensors="pt") >>> inputs = dict(pixel_values=inputs.codebook_pixel_values) >>> outputs = model(**inputs) >>> print(outputs.shape) (1, 196) ``` rz input shape z is not 4drz input has z channels but model built for )rlenrrrr)r-rrvs r.rzFlavaImageCodebook.forwardbs 9:V8VWCC_B` a  : |!! "a '|L,>,>+?zJK K   a D$7$7 7z,*<*S>STr5cl|j|}|j|}|j|}|Sr^)rrrrs r.rz$FlavaPredictionHeadTransform.forwards4 =1 --m< }5 r5r6r7r8rjrrrs@r.rrsUr5rc&eZdZdfd ZdZxZS)rRcHt|||_t||_t j |j|jd|_ t jtj|j|_ |||j_yy)NTr)rirjrbr transformrrrnrdecoderrlr:rmrweight)r-rbrrws r.rjz"FlavaMaskedPredictionHead.__init__sw  5f=yy!3!3V5F5FTR LLV->->!?@  "(DLL  r5cJ|j|}|j|}|Sr^)rrrs r.rz!FlavaMaskedPredictionHead.forwards" NN1  LLOr5r^rrs@r.rRrRs )r5rRc$eZdZfdZdZxZS) FlavaITMHeadct|||_t||_t j |jd|_y)Nr}) rirjrbrBrirrrnseq_relationshiprs r.rjzFlavaITMHead.__init__s:  !&) " &*<*#ii(;< <<(8:M:W:WXY[\:]^all,,8L8V8VWXZ[8\]`kk&88(f&es 9J$J rrs@r.rrsN 9r5rzk The FLAVA model for pretraining which outputs losses, embeddings, logits and transformer outputs. c'`eZdZdddddZddedej dzffd Zd ejfd Z e d d ejdzd ejdzdejdzdejdzdejdzdejdzdejdzdejdzdejdzdedzdejdzdejdzdejdzdedzdededzdedzdeejezf$dZxZS)!FlavaForPreTrainingzmmm_text_head.decoder.biaszmim_head.decoder.biaszmlm_head.decoder.biaszmmm_image_head.decoder.bias)zmmm_text_head.biasz mim_head.biasz mlm_head.biaszmmm_image_head.biasNrbimage_codebookcbt||t||_||_|j&|j rt |j|_t|j|_ t|j|_ t||_t|j|_t|j|_t#||_|jj&|_|jj&|_|j,|_|j.|_|j0|_|j2|_|j4|_|j6|_|j8|_|j:|_|j=y)z image_codebook ([`nn.Module`]): If passed, the image codebook will be set to this. Otherwise, it will be initialized using the image_codebook_config defined in the config first as the first parameter. N)rirjrXrLr$ init_codebookrimage_codebook_configrRrmim_headrmlm_headr itm_headmmm_image_head mmm_text_headrglobal_contrastive_headrimage_vocab_sizetext_vocab_size mlm_weight mim_weightglobal_contrastive_weightce_ignore_index itm_weightmmm_image_weightmmm_text_weight skip_unmasked_multimodal_encoderrj)r-rbr$rws r.rjzFlavaForPreTraining.__init__sQ  ' ,    &6+?+?"4V5Q5Q"RD 2&2E2EF 1&2D2DE $V, 78K8KL6v7I7IJ'A&'I$ & 3 3 > >%11<< ++ ++)/)I)I&%55 ++ & 7 7%55060W0W- r5rcn|jdkDr!|j|jdd}|S)Nr}rr|)rrrrs r. _resize_to_2dz!FlavaForPreTraining._resize_to_2ds, 557Q;qvvay"%Ar5rinput_ids_maskedrcodebook_pixel_valuesrrrrrr7 mlm_labels mim_labels itm_labelsrr4r5 return_lossr%c ||n|jj}||n|jj}| | n |j} ||tj d|}|j |||||| | ||d }|j ||||| |||d }d}|j}|j}|j}|j}|j}dx}x}x}x}x}x} }!dx}"x}#x}$}%dx}&x}'}(||C| A|r?|j td| td|jj|} |jdkDr| ||})| |j| } |j|}|j | |j#d<|)dd| j%d  dddf})| j#|j }*| |*}+|)|*ddf})|j'|)}"|rjt(j*j-|"j/d |j0|+j/d }||jz}n|j'|)}"|j2dkDr|||},| |j| } |,dd| j%d  dddf},| j#|j }*| |*}-|,|*ddf},|j5|,}#|rjt(j*j-|#j/d |j6|-j/d }||j2z}n|j5|,}#|j8dkDr||j;|}&| p| j#d}.|.|.j=z}|r/t(j*j-|&| }!|!|j8z}!|||}| | |} | | |} ||}||j>dkDr|})|j%d d z }/|)ddd d |/zddf})| |j| } |j|}|j | |j#d<| j#|j }*| |*}+|)|*ddf})|jA|)}%|rjt(j*j-|%j/d |j0|+j/d }||j>z}n|jA|)}%||jBdkDr|},|,dd|j%d  dddf},| |j| } | j#|j }*| |*}-|,|*ddf},|jE|,}$|rjt(j*j-|$j/d |j6|-j/d }||jBz}n|jE|,}$|x|u|jFdkDre|j jI|dddddf}0t(j*jK|0d }0|j jM|dddddf}1t(j*jK|1d }1|jNr8|j jPjRjUtVtX|j[|1|0|j jP\}'}(}2||'|}'|(|}(|2|}2|rWt(j*j-|'|2}3t(j*j-|(|2}4|3|4zd z } | |jFz} t]|||!| || }5|r0|5j_s tad|5jcD}|s.||jd|jdjgnd||jh|jhjgnd|j|jj|jjjgnd||jd|jdjgnd||jh|jhjgnd||jj|jjjgnd|"|#|&|'|(|%|$f}6|r|5j_s||5f|6z}6tmd|6DStod%id|d|5d|d|jdd|d|jhd|jd|jjd|d|jdd|d|jhd|d|jjd|"d|#d |&d!|'d"|(d#|%d$|$S)&a input_ids (`torch.LongTensor` of shape `(batch_size, text_seq_len)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) input_ids_masked (`torch.LongTensor` of shape `(batch_size, text_seq_len)`): Indices of input sequence tokens in the vocabulary. These ones are the masked version of the original task to be used with MLM. Indices can be obtained using [`AutoTokenizer`] along with [`DataCollatorForMaskedLanguageModeling`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) codebook_pixel_values (`torch.FloatTensor` of shape `(batch_size, num_image_patches, patch_size, patch_size, 3)`, *optional*): Pixel values for image patches that are used to compute the image codebook labels for masked image modeling. token_type_ids (`torch.LongTensor` of shape `(batch_size, text_seq_len)`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, image_num_patches)`): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). image_attention_mask (`torch.FloatTensor` of shape `(batch_size, image_num_patches)`, *optional*): Mask to avoid performing attention on padding token indices specifically for images. 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) skip_unmasked_multimodal_encoder (*bool*, *optional*): Skip any calculations for multimodal encoder for unmasked inputs. FLAVA pretraining doesn't need unmasked multimodal embeddings or outputs as of now. mlm_labels (`torch.LongTensor` of shape `(batch_size, text_seq_len)`, *optional*): Labels for computing the left-to-right language and multimodal masked modeling loss (next word prediction). Indices should be in `[-100, 0, ..., text_config.vocab_size - 1]` (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., text_config.vocab_size - 1]`. mim_labels (`torch.LongTensor` of shape `(batch_size, image_num_patches)`, *optional*): Labels for computing the image and multimodal masked modeling loss. Indices should be in `[-100, 0, ..., image_config.vocab_size - 1]`. Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., image_config.vocab_size - 1]`. If not passed, they are generated automatically using the image codebook assigned to the model. By default, it uses [`FlavaImageCodebook`]. See [`FlavaImageCodebook`] to understand how to generate mim_labels. itm_labels (`torch.LongTensor` of shape `(batch_size, 1)`, *optional*): Labels for computing the image-text matching loss. 0 means the pairs don't match and 1 means they match. The pairs with 0 will be skipped for calculation of MMM and global contrastive losses as well. return_loss (`bool`, *optional*, default to None): Whether to return calculated loss or not. Examples: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import FlavaForPreTraining, AutoProcessor >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> model = FlavaForPreTraining.from_pretrained("facebook/flava-full") >>> processor = AutoProcessor.from_pretrained("facebook/flava-full") >>> text = ["a photo of a cat"] >>> inputs = processor( ... images=[image], ... text=text, ... return_masks=True, ... return_codebook_pixels=True, ... padding=True, ... max_length=77, ... return_tensors="pt", ... ) >>> output = model(**inputs) ``` Nz`input_ids_masked` isn't passed which means MLM loss won't be calculated correctlySetting it to `input_ids` so that model can work. Please pass it if this is unintentional. This is usually OKAY if you are doing inference on unmasked text...T) rrrrrrrrr4r5) rrrrrrrr4r5z`return_loss` is set to True but the image codebook is not initialized and no `mim_labels` have been passed. Reinstantiate the model with `init_codebook` set to True or pass in your custom `mim_labels`z`codebook_pixel_value` are required to generate `mim_labels` if loss is expected. Call `AutoProcessor` with `return_codebook_pixels` set to Truerrr|r}r)r@rArBrCrDrEc3(K|] }||nd ywrGr=)r+rMs r.r/z.FlavaForPreTraining.forward..Ks_T%5T1<_sc3&K|] }|| ywr^r=)r+rs r.r/z.FlavaForPreTraining.forward..lsrr4r5r?rv flava_outputflava_masked_outputpos_maskrr!rOrQrS total_lossmim_lossmlm_loss mmm_text_lossmmm_image_lossgc_lossitm_lossrUrVr[rZrWr r!sequence_for_image masked_tokensmim_labels_filteredsequence_for_textmlm_labels_filtered pos_pairs end_indextext_embeddingimage_embedding gc_labels gc_loss_image gc_loss_text flava_lossesrs7 r.rzFlavaForPreTraining.forwardsz D&1%eGV^GKK KZK/4D:>> >% # .2N2Z!k&&.&; )0$Y"00EEF[\  ??Q #:#FKgKo!8 %!// ; "&"4"4_"E7;7K7K ?--d34%7JOOA !}}::"D,A,ABDWD\D\]_D` H/H!]]+=>  ??Q #9#EJfJn 6 %!// ; $5a*//!:L9L9NPQ6Q$R! * d.B.B C &0&?#$5mQ6F$G!!]]+<= !}}::"D,@,@ACVC[C[\^C_ H/H!]]+<=  ??Q #?#K'CDJ%&MM!, $ '77!}}:::zRH/H/;3OPX3Y0)!+H!5J)!+H!5J&5h&?O ( 38M8MPQ8Q!= /44Q7!;I!3Aq1y=7H!4K!L %!// ; "&"4"4_"E7;7K7K ?--d34 * d.B.B C &0&?#%7 q8H%I"#'#6#67I#J %']]%@%@(--b$2G2GHJ]JbJbceJf&N#d&;&;;N#'#6#67I#J  ( 38L8Lq8P <  1!6L6Q6QRS6T5T5VXY2Y Z %!// ; * d.B.B C &0&?#$5mQ6F$G!"&"4"45F"G$&MM$?$?',,R1E1EFH[H`H`acHd%M"T%9%99M"&"4"45F"G  'O,GDLjLjmnLn!ZZ771a8PQN]]44^4LN"jj99:J1aQR7:STO mm55o25NO}} &&++223HJ_`;?;W;W1G1G< 8 oy ##3H#= "1(";%h/  " ; ;
  • Q>]>]>i#//88:os,&88D$55>>@  +F.<#8#8#:  > *F &9%J%J " "! ""# $*:% &)8' (.) *,+  r5r^)NNNNNNNNNNNNNNTNN)r6r7r8_tied_weights_keysrrr}rjr:rr9rrr;rJr2rLrrrs@r.r#r#s;00< !{!BII.2.2/304488<*.*.*.)-%)#'#'%p ##d*p  **T1p ''$. p %0047 p  t+ p  t+p ,p &&-p $llT1p +/+p LL4'p LL4'p LL4'p  $;p #!p "D[#p $D[%p ( u|| 8 8)p p r5r#)r#rrarXrVrKr)Lr9rrr dataclassesrtypingrr:rrrS activationsr masking_utilsr modeling_layersr modeling_outputsr r modeling_utilsrprocessing_utilsrutilsrrrrrrconfiguration_flavarrrrr get_loggerr6rCrrMrNrrr?rLr}rarprrrr rrr"r-rBrKrarrVrXrrrrrrRr rr#__all__r=r5r.rus2 #! &!69K-&jj   H %>&)99