o پi7 @s|UdZddlmZmZmZmZmZmZmZddl Z ddl m Z ddl m m ZddlmZmZddlmZddlmZddlmZdd lmZmZd gZgd gd gd gddZeeeeeeffed<eGddde j Z!Gdd d e j Z"de#deee j$ffddZ%dede&dede"fddZ'd5dedeeeffdd Z(ee(d!d"e(d!d"e(d!d"e(d!d"e(d!d"e(d!d"e(d!d"e(d!d"d#Z)ed6de&dede"fd%d&Z*ed6de&dede"fd'd(Z+ed6de&dede"fd)d*Z,ed6de&dede"fd+d,Z-ed6de&dede"fd-d.Z.ed6de&dede"fd/d0Z/ed6de&dede"fd1d2Z0ed6de&dede"fd3d4Z1dS)7zVGG Adapted from https://github.com/pytorch/vision 'vgg.py' (BSD-3-Clause) with a few changes for timm functionality. Copyright 2021 Ross Wightman )AnyDictListOptionalTypeUnioncastNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)ClassifierHead)build_model_with_cfg)register_notrace_module)register_modelgenerate_default_cfgsVGG) @Mrrrrrrrr)rrrrrrrrrrrrrrr)rrrrrrrrrrrrrrrrrr)rrrrrrrrrrrrrrrrrrrrr)vgg11vgg13vgg16vgg19cfgscsteZdZdZdddddejejfdeded ed ed ed e ej d e ej ffdd Z de j de j fddZZS)ConvMlpzxConvolutional MLP block for VGG head. Replaces traditional Linear layers with Conv2d layers in the classifier. r?g? in_features out_features kernel_size mlp_ratio drop_rate act_layer conv_layerc shtt|||_t||}||||dd|_|d|_t||_ |||ddd|_ |d|_ dS)aInitialize ConvMlp. Args: in_features: Number of input features. out_features: Number of output features. kernel_size: Kernel size for first conv layer. mlp_ratio: Ratio for hidden layer size. drop_rate: Dropout rate. act_layer: Activation layer type. conv_layer: Convolution layer type. T)biasr N) superr__init__input_kernel_sizeintfc1act1nnDropoutdropfc2act2) selfr!r"r#r$r%r&r' mid_features __class__C/home/ubuntu/.local/lib/python3.10/site-packages/timm/models/vgg.pyr*&s   zConvMlp.__init__xreturncCs|jd|jks|jd|jkr(t|j|jdt|j|jdf}t||}||}||}||}||}| |}|S)zoForward pass. Args: x: Input tensor. Returns: Output tensor. ) shaper+maxFadaptive_avg_pool2dr-r.r1r2r3)r4r: output_sizer8r8r9forwardDs $      zConvMlp.forward)__name__ __module__ __qualname____doc__r/ReLUConv2dr,floatrModuler*torchTensorrC __classcell__r8r8r6r9rs2rcsFeZdZdZddddejejdddf d eed e d e d e d e de ej de ej de e ej dede ddffdd Zejjd,dedeeeffddZejjd-deddfddZejjdej fddZd.d e de eddfd d!Zd"ejdejfd#d$Zd,d"ejd%edejfd&d'Zd"ejdejfd(d)Zd/d*d+ZZS)0rzVGG model architecture. Based on `Very Deep Convolutional Networks for Large-Scale Image Recognition` - https://arxiv.org/abs/1409.1556  r Navggcfg num_classesin_chans output_strider$r&r' norm_layer global_poolr%r;c  sltt||dks J||_| |_d|_|du|_g|_|} d} tj } g}|D]Q}t |d}|dkrQ|j t | | d|d|| ddd g7}| d9} q)t t|}|| |d dd }|duro|||||d d g7}n |||d d g7}|} q)tj||_|j t | | dt |dd| |_d|_t| |jd|| ||d|_t|j|| | d|_|dS)a/Initialize VGG model. Args: cfg: Configuration list defining network architecture. num_classes: Number of classes for classification. in_chans: Number of input channels. output_stride: Output stride of network. mlp_ratio: Ratio for MLP hidden layer size. act_layer: Activation layer type. conv_layer: Convolution layer type. norm_layer: Normalization layer type. global_pool: Global pooling type. drop_rate: Dropout rate. rQFNr rz features.)num_chs reductionmodule)r#striderP)r#paddingT)inplacerr)r$r%r&r') pool_typer%)r)rr*rTr%grad_checkpointinguse_norm feature_infor/ MaxPool2dlenappenddictrr, Sequentialfeatures num_featureshead_hidden_sizer pre_logitsr head_initialize_weights)r4rSrTrUrVr$r&r'rWrXr%prev_chs net_stride pool_layerlayersvlast_idxconv2dr6r8r9r*`sV      $  z VGG.__init__FcoarsecCs tdddS)zGroup matcher for parameter groups. Args: coarse: Whether to use coarse grouping. Returns: Dictionary of grouped parameters. z ^features\.0z^features\.(\d+))stemblocks)rg)r4rvr8r8r9 group_matchers zVGG.group_matcherTenablecCs|rJddS)zEnable or disable gradient checkpointing. Args: enable: Whether to enable gradient checkpointing. z$gradient checkpointing not supportedNr8)r4rzr8r8r9set_grad_checkpointingszVGG.set_grad_checkpointingcCs|jjS)zTGet the classifier module. Returns: Classifier module. )rmfc)r4r8r8r9get_classifierszVGG.get_classifiercCs||_|j||dS)zReset the classifier. Args: num_classes: Number of classes for new classifier. global_pool: Global pooling type. N)rTrmreset)r4rTrXr8r8r9reset_classifierszVGG.reset_classifierr:cCs||}|S)zForward pass through feature extraction layers. Args: x: Input tensor. Returns: Feature tensor. )rir4r:r8r8r9forward_featuress zVGG.forward_featuresrlcCs&||}|r|j||dS||S)zForward pass through head. Args: x: Input features. pre_logits: Return features before final linear layer. Returns: Classification logits or features. )rl)rlrm)r4r:rlr8r8r9 forward_heads zVGG.forward_headcCs||}||}|S)zoForward pass. Args: x: Input tensor. Returns: Output logits. )rrrr8r8r9rCs z VGG.forwardcCs|D]N}t|tjr$tjj|jddd|jdur#tj|jdqt|tj r;tj|jdtj|jdqt|tj rRtj |jddtj|jdqdS)zInitialize model weights.fan_outrelu)mode nonlinearityNrr g{Gz?) modules isinstancer/rIinitkaiming_normal_weightr( constant_ BatchNorm2dLinearnormal_)r4mr8r8r9rns     zVGG._initialize_weightsF)T)N)r;N)rDrErFrGr/rHrIrrr,rJrrKrstrr*rLjitignoreboolrryr{r}rrMrrrCrnrNr8r8r6r9rYsZ     L     state_dictr;cCszi}|D]4\}}|}|dd}|dd}|dd}d|vr*|dd d d }d |vr6|dd d d }|||<q|S)zConvert patch embedding weight from manual patchify + linear proj to conv. Args: state_dict: State dictionary to filter. Returns: Filtered state dictionary. z classifier.0zpre_logits.fc1z classifier.3zpre_logits.fc2z classifier.6head.fczclassifier.0.weightr=rrzclassifier.3.weightrr )itemsreplacereshape)rout_dictkrsk_rr8r8r9 _filter_fn s     rvariant pretrainedkwargscKsF|dd}|dd}tt||ft|td|dtd|}|S)zCreate a VGG model. Args: variant: Model variant name. pretrained: Load pretrained weights. **kwargs: Additional model arguments. Returns: VGG model instance. _r out_indices)rr r\rPT)flatten_sequentialr) model_cfg feature_cfgpretrained_filter_fn)splitpoprrrrgr)rrrrSrmodelr8r8r9 _create_vgg!s   rurlc Ks|dddddttddd |S) zCreate default configuration dictionary. Args: url: Model weight URL. **kwargs: Additional configuration options. Returns: Configuration dictionary. rO)rPr)rrg?bilinearz features.0r) rrT input_size pool_sizecrop_pct interpolationmeanstd first_conv classifierr )rrr8r8r9_cfg;s rztimm/) hf_hub_id)z vgg11.tv_in1kz vgg13.tv_in1kz vgg16.tv_in1kz vgg19.tv_in1kzvgg11_bn.tv_in1kzvgg13_bn.tv_in1kzvgg16_bn.tv_in1kzvgg19_bn.tv_in1kFcK tdi|}tdd|i|S)zVGG 11-layer model (configuration "A") from `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rrNr8)rrgrrr model_argsr8r8r9r[rcK&tddtji|}tdd|i|S)zVGG 11-layer model (configuration "A") with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rWvgg11_bnrNr8)rrgr/rrrr8r8r9rdrcKr)zVGG 13-layer model (configuration "B") `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rrNr8)rrrr8r8r9rmrrcKr)zVGG 13-layer model (configuration "B") with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rWvgg13_bnrNr8)rrrr8r8r9rvrrcKr)zVGG 16-layer model (configuration "D") `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rrNr8)rrrr8r8r9rrrcKr)zVGG 16-layer model (configuration "D") with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rWvgg16_bnrNr8)rrrr8r8r9rrrcKr)zVGG 19-layer model (configuration "E") `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rrNr8)rrrr8r8r9rrrcKr)zVGG 19-layer model (configuration 'E') with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" `._ rWvgg19_bnrNr8)rrrr8r8r9rrr)rr)2rGtypingrrrrrrrrLtorch.nnr/torch.nn.functional functionalr@ timm.datar r timm.layersr _builderr _features_fxr _registryrr__all__rrr,__annotations__rKrrrgrMrrrr default_cfgsrrrrrrrrr8r8r8r9sb$    "92