S^(h\ dZddlmZmZmZddlZddlZddlmZddlm Z m Z m Z ddl m Z ddlmZmZmZmZdd lmZdd lmZmZmZmZmZdd lmZd d lmZej>e Z!dZ"dZ#gdZ$dZ%dZ&d3dejNde(de)dejNfdZ*GddejVZ,GddejVZ-GddejVZ.GddejVZ/Gdd ejVZ0Gd!d"ejVZ1Gd#d$ejVZ2Gd%d&eZ3d'Z4d(Z5ed)e4Gd*d+e3Z6ed,e4Gd-d.e3Z7ed/e4Gd0d1e3eZ8gd2Z9y)4zPyTorch ConvNextV2 model.)OptionalTupleUnionN)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)add_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) BackboneMixin)ConvNextV2Configrzfacebook/convnextv2-tiny-1k-224)rirztabby, tabby catinput drop_probtrainingreturnc|dk(s|s|Sd|z }|jdfd|jdz zz}|tj||j|j z}|j |j||z}|S)aF Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the argument. rr)r)dtypedevice)shapendimtorchrandr r!floor_div)rrr keep_probr" random_tensoroutputs p/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/convnextv2/modeling_convnextv2.py drop_pathr,:sCx I I [[^  Q 7 7E 5 ELL YYM YYy !M 1F McxeZdZdZd deeddffd ZdejdejfdZ de fdZ xZ S) ConvNextV2DropPathzXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).Nrrc0t|||_yN)super__init__r)selfr __class__s r+r3zConvNextV2DropPath.__init__Rs "r- hidden_statescDt||j|jSr1)r,rrr4r6s r+forwardzConvNextV2DropPath.forwardVs FFr-c8dj|jS)Nzp={})formatr)r4s r+ extra_reprzConvNextV2DropPath.extra_reprYs}}T^^,,r-r1) __name__ __module__ __qualname____doc__rfloatr3r$Tensorr9strr< __classcell__r5s@r+r/r/OsGb#(5/#T#GU\\GellG-C-r-r/c`eZdZdZdeffd ZdejdejfdZxZ S) ConvNextV2GRNz)GRN (Global Response Normalization) layerdimct|tjt j ddd||_tjt j ddd||_y)Nr)r2r3r Parameterr$zerosweightbias)r4rHr5s r+r3zConvNextV2GRN.__init__`sL ll5;;q!Q#<= LLQ1c!:; r-r6rctjj|ddd}||jdddzz }|j||zz|j z|z}|S)N)rrOT)ordrHkeepdim)rHrQư>)r$linalgnormmeanrLrM)r4r6global_features norm_featuress r+r9zConvNextV2GRN.forwardesi,,++MqfVZ+['?+?+?BPT+?+UX\+\]  }}'DE QTaa r-) r=r>r?r@intr3r$ FloatTensorr9rDrEs@r+rGrG]s13#?@ LL-=!>? &   #F F%(A$BRBRAS&TU U!1 3r-xrcd|jdk(rWtjjj ||j |j |j|j}|S|jdk(r|j}|j}|jdd}||z jdjdd}||z tj||jzz }|j|}|j ddddf|z|jddddfz}|S)Nr^r_rT)rQrO)r )rbr$r functional layer_normrdrLrMrar rArVpowsqrtto)r4re input_dtypeuss r+r9zConvNextV2LayerNorm.forwards     .##..q$2G2GVZV_V_aeaiaijA  !1 1''K Aq$'AQ A##At#4AQ%**Q\22A;'A AtTM*Q.1dD=1IIAr-)rSr^) r=r>r?r@r3r$rBr9rDrEs@r+r\r\os( 4  %,, r-r\cZeZdZdZfdZdej dejfdZxZ S)ConvNextV2EmbeddingszThis class is comparable to (and inspired by) the SwinEmbeddings class found in src/transformers/models/swin/modeling_swin.py. ct|tj|j|j d|j |j |_t|j ddd|_ |j|_y)Nr kernel_sizestriderSr_rarb) r2r3rConv2d num_channels hidden_sizes patch_sizepatch_embeddingsr\ layernormr4configr5s r+r3zConvNextV2Embeddings.__init__sr  "   !4!4Q!7VEVEV_e_p_p! -V-@-@-C[kl"//r- pixel_valuesrc|jd}||jk7r td|j|}|j |}|S)NrzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)r"rw ValueErrorrzr{)r4r~rw embeddingss r+r9zConvNextV2Embeddings.forwardsV#))!, 4,, ,w **<8 ^^J/ r- r=r>r?r@r3r$rZrBr9rDrEs@r+rprps*0E$5$5%,,r-rpc\eZdZdZdfd Zdej dejfdZxZ S)ConvNextV2Layera5This corresponds to the `Block` class in the original implementation. There are two equivalent implementations: [DwConv, LayerNorm (channels_first), Conv, GELU,1x1 Conv]; all in (N, C, H, W) (2) [DwConv, Permute to (N, H, W, C), LayerNorm (channels_last), Linear, GELU, Linear]; Permute back The authors used (2) as they find it slightly faster in PyTorch. Args: config ([`ConvNextV2Config`]): Model configuration class. dim (`int`): Number of input channels. drop_path (`float`): Stochastic depth rate. Default: 0.0. ct|tj||dd||_t |d|_tj|d|z|_t|j|_ td|z|_ tjd|z||_|dkDrt||_ytj |_y)Nrr )rspaddinggroupsrSrar)r2r3rrvdwconvr\r{Linearpwconv1r hidden_actactrGgrnpwconv2r/Identityr,)r4r}rHr,r5s r+r3zConvNextV2Layer.__init__s iiSa3O ,Sd;yya#g. &++, S)yyS#. :Cc/+I6r{{}r-r6rcN|}|j|}|jdddd}|j|}|j|}|j |}|j |}|j |}|jdddd}||j|z}|S)NrrOr r)rpermuter{rrrrr,)r4r6rres r+r9zConvNextV2Layer.forwards KK & IIaAq ! NN1  LLO HHQK HHQK LLO IIaAq ! DNN1% %r-)rrrEs@r+rrs+  ]U%6%65<<r-rc\eZdZdZdfd Zdej dejfdZxZ S)ConvNextV2StageaConvNeXTV2 stage, consisting of an optional downsampling layer + multiple residual blocks. Args: config ([`ConvNextV2Config`]): Model configuration class. in_channels (`int`): Number of input channels. out_channels (`int`): Number of output channels. depth (`int`): Number of residual blocks. drop_path_rates(`List[float]`): Stochastic depth rates for each layer. c ~t |||k7s|dkDr?tjt |ddtj |||||_ntj|_|xsdg|z}tjt|Dcgc]}t||||c}|_ ycc}w)NrrSr_rurrr)rHr,) r2r3r Sequentialr\rvdownsampling_layerrrangerlayers) r4r} in_channels out_channelsrsrtdepthdrop_path_ratesjr5s r+r3zConvNextV2Stage.__init__s  , &&1*&(mm#KTGWX +|U[\'D # ')kkmD #):cUU]mm_dej_k lZ[of,/RSBTU l  lsB:r6rcJ|j|}|j|}|Sr1)rrr8s r+r9zConvNextV2Stage.forwards&// >  M2 r-)rOrOrONrrEs@r+rrs*  U%6%65<<r-rc feZdZfdZ ddej deedeedee e ffdZ xZ S)ConvNextV2Encoderc (t|tj|_t j d|jt|jj|jDcgc]}|j}}|jd}t|jD]V}|j|}t||||dkDrdnd|j|||}|jj!||}Xycc}w)NrrOr)rrrtrr)r2r3r ModuleListstagesr$linspacedrop_path_ratesumdepthssplittolistrxr num_stagesrappend) r4r}rerprev_chsiout_chsstager5s r+r3zConvNextV2Encoder.__init__s mmo %q&2G2GV]]I[ \ b bcicpcp q AHHJ  &&q)v(() A))!,G#$$EqqmmA& / 2 E KK  u %H   s8Dr6output_hidden_states return_dictrc|rdnd}t|jD]\}}|r||fz}||}|r||fz}|std||fDSt||S)Nc3&K|] }|| ywr1r).0vs r+ z,ConvNextV2Encoder.forward..sXq!-Xs)last_hidden_stater6) enumeratertupler )r4r6rrall_hidden_statesr layer_modules r+r9zConvNextV2Encoder.forwards #7BD(5 8OA|#$58H$H!(7M  8 1]4D D X]4E$FXX X-++  r-)FT) r=r>r?r3r$rZrboolrrr r9rDrEs@r+rrsT.05&*  (( 'tn d^  u44 5  r-rc(eZdZdZeZdZdZdgZdZ y)ConvNextV2PreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. convnextv2r~rct|tjtjfrm|jj j d|jj|j%|jj jyyt|tjtfrJ|jj j|jj jdyt|trI|jj j|jj jyy)zInitialize the weightsr)rVstdNg?) isinstancerrrvrLdatanormal_r}initializer_rangerMzero_ LayerNormr\fill_rG)r4modules r+ _init_weightsz'ConvNextV2PreTrainedModel._init_weights-s fryy"))4 5 MM   & &CT[[5R5R & S{{&   &&(' /B C D KK   " " $ MM   $ $S )  . MM   $ $ & KK   " " $/r-N) r=r>r?r@r config_classbase_model_prefixmain_input_name_no_split_modulesrrr-r+rr"s' $L$$O*+ %r-raL This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`ConvNextV2Config`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights. aI Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`ConvNextImageProcessor`]. See [`ConvNextImageProcessor.__call__`] for details. 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. zSThe bare ConvNextV2 model outputting raw features without any specific head on top.c eZdZfdZeeeeee de  d de e jde ede edeeeffdZxZS) ConvNextV2Modelct||||_t||_t ||_tj|jd|j|_ |jy)NrRr) r2r3r}rprrencoderrrrxlayer_norm_epsr{ post_initr|s r+r3zConvNextV2Model.__init__[s`   .v6(0 f&9&9"&=6CXCXY r-vision) checkpoint output_typermodalityexpected_outputr~rrrcd||n|jj}||n|jj}| td|j |}|j |||}|d}|j |jddg}|s ||f|ddzSt|||jS)Nz You have to specify pixel_valuesrrrrRr)r pooler_outputr6) r}ruse_return_dictrrrr{rVrr6)r4r~rrembedding_outputencoder_outputsr pooled_outputs r+r9zConvNextV2Model.forwardhs%9$D $++JjJj &1%r?r3rCONVNEXTV2_INPUTS_DOCSTRINGr_CHECKPOINT_FOR_DOCr_CONFIG_FOR_DOC_EXPECTED_OUTPUT_SHAPErr$rZrrrr9rDrEs@r+rrUs ++FG&<$. 59/3&* " u001" 'tn" d^ " u>> ? " H" r-rz ConvNextV2 Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. ceZdZfdZeeeeee e  d de e jde e jde ede edeeeff dZxZS) ConvNextV2ForImageClassificationc0t|||j|_t||_|jdkDr-t j |jd|jnt j|_ |jy)NrrR) r2r3 num_labelsrrrrrxr classifierrr|s r+r3z)ConvNextV2ForImageClassification.__init__sy   ++)&1FLEVEVYZEZBIIf))"-v/@/@ A`b`k`k`m  r-)rrrrr~labelsrrrc||n|jj}|j|||}|r |jn|d}|j |}d}||jj |j dk(rd|j_nl|j dkDrL|jtjk(s|jtjk(rd|j_nd|j_|jj dk(rIt} |j dk(r&| |j|j}n| ||}n|jj dk(r=t} | |jd|j |jd}n,|jj dk(rt} | ||}|s|f|ddz} ||f| zS| St!|||j" S) a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Nrr regressionsingle_label_classificationmulti_label_classificationrRrO)losslogitsr6)r}rrrr problem_typerr r$longrYr squeezerviewrrr6) r4r~rrroutputsrrrloss_fctr*s r+r9z(ConvNextV2ForImageClassification.forwards(&1%r?r3rrr_IMAGE_CLASS_CHECKPOINTrr_IMAGE_CLASS_EXPECTED_OUTPUTrr$rZ LongTensorrrrr9rDrEs@r+rrs ++FG*8$4 59-1/3&* 3 u0013 ))*3 'tn 3 d^ 3 u:: ; 3 H3 r-rzT ConvNeXT V2 backbone, to be used with frameworks like DETR and MaskFormer. c eZdZfdZeeeee dde jde e de e defdZ xZS) ConvNextV2Backbonect||t| |t||_t ||_|jdg|jz|_i}t|j|jD]\}}t|d||<tj||_|j!y)Nrr_)rb)r2r3_init_backbonerprrrrx num_featureszip _out_featureschannelsr\r ModuleDicthidden_states_normsr)r4r}rrrwr5s r+r3zConvNextV2Backbone.__init__s   v&.v6(0 #0034v7J7JJ!#&t'9'94==#I i E<)<\Wg)h  & i#%==1D#E  r-)rrr~rrrc||n|jj}||n|jj}|j|}|j |d|}|r |j n|d}d}t |j|D]/\}} ||jvs|j|| } || fz }1|s |f} |r| |fz } | St||r|dSddS)a Returns: Examples: ```python >>> from transformers import AutoImageProcessor, AutoBackbone >>> import torch >>> from PIL import Image >>> import requests >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = AutoImageProcessor.from_pretrained("facebook/convnextv2-tiny-1k-224") >>> model = AutoBackbone.from_pretrained("facebook/convnextv2-tiny-1k-224") >>> inputs = processor(image, return_tensors="pt") >>> outputs = model(**inputs) ```NTrrr) feature_mapsr6 attentions) r}rrrrr6r stage_names out_featuresrr ) r4r~rrrrr6r r hidden_stater*s r+r9zConvNextV2Backbone.forwards8&1%t77>|L /  0 "_F#=**M%+?-  EI  r-)NN)r=r>r?r3rrrr rr$rBrrr9rDrEs@r+rrsm"++FG>X04&* 9 ll9 'tn9 d^ 9  9 YH9 r-r)rrrr)rF):r@typingrrrr$torch.utils.checkpointrtorch.nnrrr activationsr modeling_outputsr r rrmodeling_utilsrutilsrrrrrutils.backbone_utilsrconfiguration_convnextv2r get_loggerr=loggerrrrrrrBrArr,Moduler/rGr\rprrrrCONVNEXTV2_START_DOCSTRINGrrrr__all__rr-r+rs )) AA! .26   H %%8'<1U\\eTV[VbVb* - -BII$"))>2990(bii(XbiiB,  , ^%%6  Y 8 /8  8 v I '@I I X M 2MM M ` ur-