o *ju@sddlZddlZddlmZddlmZmZddlZddl Z ddl m m Z ddl m Z ddlmZGddde jZGd d d e jZGd d d e jZGd dde jZGddde jZeddddddedddddddZdddZedkreZdSdS)N) OrderedDict)TupleUnion)nn)ViMcs$eZdZdejffdd ZZS) LayerNormxcs$|j}t|tj}||SN)dtypesuperforwardtypetorchZfloat32)selfr orig_typeret __class__p/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/modelscope/models/cv/vision_middleware/backbone.pyr s zLayerNorm.forward)__name__ __module__ __qualname__rTensorr __classcell__rrrrrsrc@seZdZdejfddZdS) QuickGELUr cCs|td|S)NgZd;?)rZsigmoidrr rrrr szQuickGELU.forwardN)rrrrrr rrrrrsrcsTeZdZ d dededejffdd Zdejfdd Zdejd efd d Z Z S)ResidualAttentionBlockNd_modeln_head attn_maskc stt|||_t||_ttdt ||dfdt fdt |d|fg|_ t||_ ||_ t|_t|_dS)NZc_fcZgeluZc_proj)r __init__rZMultiheadAttentionattnrln_1Z SequentialrZLinearrmlpln_2r!rvim_attvim_mlp)rrr r!rrrr#!s    zResidualAttentionBlock.__init__r cCs>|jdur|jj|j|jdnd|_|j|||d|jddS)N)r deviceF)Z need_weightsr!r)r!tor r*r$rrrr attention3s  z ResidualAttentionBlock.attention task_namecCsT||}|||}||||}||}|||}||||}|Sr )r%r,r(r'r&r))rr r-Z x_normed_1Z x_normed_2rrrr :s  zResidualAttentionBlock.forwardr ) rrrintrrr#r,strr rrrrrrsrc sDeZdZ d dedededejffdd Zdejfd d ZZS) TransformerNwidthlayersheadsr!cs<t|_||_tfddt|D|_dS)Ncsg|]}tqSr)r).0_r!r3r1rr Ps z(Transformer.__init__..)r r#r1r2rZ ModuleListrange resblocks)rr1r2r3r!rr6rr#Hs zTransformer.__init__r c KsD|\}}}g}t|jD]\}}||fi|}||q|Sr )size enumerater9append) rr kwargsLBDfeaturesiZblkrrrr Us  zTransformer.forwardr ) rrrr.rrr#r rrrrrr0Fs r0c sNeZdZdZ ddedededededef fd d Zd ejfd d ZZ S)VisionTransformeraf The Vision Transformer (ViT) model Args: - input_resolution (int): shape of input image - patch_width (int): size of patch tokens - width (int): feature channels - layers (int): number of transformer layers - heads (int): number of multi-head attention - output_dim (int): output feature channels input_resolution patch_sizer1r2r3 output_dimcst||_tjd|||dd|_|d}t|t||_ t|t||dd||_ t ||_ |||_ t||||_t ||_t|t|||_||_dS)NF)Z in_channelsZ out_channelsZ kernel_sizeZstrideZbiasgr)r r#rErZConv2dconv1 ParameterrZrandnclass_embeddingpositional_embeddingrln_preZpatch_per_sider0 transformerln_postprojrG)rrErFr1r2r3rGscalerrrr#js(      zVisionTransformer.__init__r c KsN||}|d}|d}||jd|jdd}|ddd}|j|jddd|dd}t j ||gdd}||j |j}| |}|ddd}|j |fi|}|d}|ddd}||dddddf}|jdur~||j}g}|D]}||ddddddfddd|d||q|||S)NrrIr)dim)rJr:ZreshapeshapeZpermuterLr+r repeatrcatrMrNrOrPrQr<) rr r=r?PZ cls_tokenZ x_per_layerZoutputsoutputrrrr s:       "  zVisionTransformer.forward)rD) rrr__doc__r.r#rrr rrrrrrC^s"rCi )rErFr1r2r3 ) vit_b16_224Z vit_b32_224r_cCs$t|}tdi|}|||S)z build a ViT + ViM model Args: arch: name of backbone pretrained: weights of pretrained model Nr) model_dictrCZload_state_dict)archZ pretrainedZ model_argsmodelrrrbuild_backbones rc__main__)r_N)mathos collectionsrtypingrrnumpynprZtorch.nn.functionalrZ functionalFvimrrModulerrr0rCdictr`rcrrbrrrrs*   'L