hddlmZmZddlZddlmZddlmZddlm Z m Z ddl m Z ddl mZmZmZdd lmZmZmZmZGd d eZeGd d eZGddeeZedGddeeZgdZy))OptionalUnionN) IJepaConfig)BaseModelOutputWithPoolingImageClassifierOutput)Unpack)TransformersKwargsauto_docstring torch_int) ViTEmbeddingsViTForImageClassificationViTModelViTPreTrainedModelc eZdZddededdffd Zdejdededejfd Z dd ejd e ejd edejfd Z xZ S)IJepaEmbeddingsconfiguse_mask_tokenreturnNct||||`|jj}t j tjd||j|_ y)N) super__init__ cls_tokenpatch_embeddings num_patchesnn Parametertorchrandn hidden_sizeposition_embeddings)selfrrr __class__s e/var/www/html/eduruby.in/venv/lib/python3.12/site-packages/transformers/models/ijepa/modular_ijepa.pyrzIJepaEmbeddings.__init__sL 0 N++77 #%<< A{FL^L^0_#` embeddingsheightwidthc0|jd}|jjd}tjj s||k(r||k(r |jS|j}|jd}||j z}||j z} t |dz} |jd| | |}|jdddd}tjj||| fdd }|jddddjdd|}|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. 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 rg?rrr bicubicF)sizemode align_corners) shaper#r jit is_tracing patch_sizer reshapepermuter functional interpolateview) r$r(r)r*r num_positionspatch_pos_embeddim new_height new_widthsqrt_num_positionss r&interpolate_pos_encodingz(IJepaEmbeddings.interpolate_pos_encodings#!&&q) 0066q9 yy##%+*F6UZ?++ +22r"t. T__, &}c'9:)11!5GI[]`a)11!Q1=--33 i( 4 *11!Q1=BB1b#Nr' pixel_valuesbool_masked_posr@cx|j\}}}}|j||}|Z|jd} |jj|| d} |j dj | } |d| z z| | zz}|r||j |||z}n||jz}|j|}|S)N)r@rr,?) r1r mask_tokenexpand unsqueezetype_asr@r#dropout) r$rArBr@ batch_size_r)r*r( seq_length mask_tokensmasks r&forwardzIJepaEmbeddings.forward=s (4'9'9$ Avu**&>>J\\*- r')F)NF)__name__ __module__ __qualname__rboolrr Tensorintr@r BoolTensorrO __classcell__r%s@r&rrsa{aDaTa%5<<%%UX%]b]i]i%T7;). ll"%"2"23#'   r'rcdeZdZdeej ej ejfddfdZy)IJepaPreTrainedModelmodulerNclt|tjtjfrtjj |j jjtjd|jjj|j j|j _|j%|jjjyyt|tj rJ|jjj|j jj#dyt|t$rtjj |j&jjtjd|jjj|j&j|j&_|j(%|j(jjyyy)zInitialize the weightsg)meanstdNrD) isinstancerLinearConv2dinit trunc_normal_weightdatator float32rinitializer_rangedtypebiaszero_ LayerNormfill_rr#rE)r$r[s r& _init_weightsz"IJepaPreTrainedModel._init_weightsZs fryy"))4 5"$!6!6 ""%%emm43DKKDaDa"7"b$$% MM {{&   &&('  - KK   " " $ MM   $ $S )  0.0gg.C.C**//225==AKK11/D/b++112  & & +   ,!!&&,,.- 1r') rPrQrRrrr`rarlrnr'r&rZrZXs./E"))RYY *L$M/RV/r'rZc.eZdZddededeffd ZxZS) IJepaModelradd_pooling_layerrcVt||||_t|||_y)z add_pooling_layer (bool, *optional*, defaults to `True`): Whether to add a pooling layer use_mask_token (`bool`, *optional*, defaults to `False`): Whether to use a mask token for masked image modeling. )rN)rrrrr()r$rrrrr%s r&rzIJepaModel.__init__rs'   )&Pr')FF)rPrQrRrrSrrWrXs@r&rqrqqs( Q{ Qt Q]a Q Qr'rqa IJepa Model transformer with an image classification head on top (a linear layer on top of the final hidden states) e.g. for ImageNet. Note that it's possible to fine-tune IJepa on higher resolution images than the ones it has been trained on, by setting `interpolate_pos_encoding` to `True` in the forward of the model. This will interpolate the pre-trained position embeddings to the higher resolution. ) custom_introc eZdZdeffd Z d deejdeejdeejdeede e de f d Z xZ S) IJepaForImageClassificationrcht||t|d|_|j y)NF)rr)rrrqijepa post_init)r$rr%s r&rz$IJepaForImageClassification.__init__s(  %@  r'rA head_masklabelsr@kwargsrc |j|f||d|}|j}|j|jd}d} ||j|||j fi|} t | ||j|jS)a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). )rzr@r)r<N)losslogits hidden_states attentions) rxlast_hidden_state classifierr] loss_functionrrrr) r$rArzr{r@r|outputssequence_outputrr~s r&rOz#IJepaForImageClassification.forwards/9djj / %=/  /  "33!5!5!!5!<=  %4%%ffdkkLVLD$!//))   r')NNNN)rPrQrRrrrr rTrSr r rrOrWrXs@r&rvrv~s{04,0)-37 ! u||,! ELL)! & ! #+4. ! +, !  ! r'rv)rZrqrv)typingrrr torch.nnr-transformers.models.ijepa.configuration_ijeparmodeling_outputsrrprocessing_utilsr utilsr r r vit.modeling_vitrrrrrrZrqrv__all__ror'r&rs" EQ&BBeeGmGT/-//0 Q%x Q  ' "68Q'  ' T r'