o ,hn}@sddlZddlmZddlmZddlmZddlmZm Z m Z ddl Z ddl Z ddlmmZddl mZmZddlmZmZmZddlmZdd lmZmZdd lmZmZdd lm Z dd l!m"Z"m#Z#dd l$m%Z%gdZ&de'e(e(fde(de(de(de'e(e(ff ddZ)de'e(e(fde(de*e'e(e(ffddZ+de(de(de jfddZ,Gdddej-Z.Gdd d ej-Z/Gd!d"d"ej-Z0Gd#d$d$ej-Z1Gd%d&d&ej-Z2Gd'd(d(ej-Z3Gd)d*d*ej-Z4Gd+d,d,ej-Z5Gd-d.d.ej-Z6  /dCd0e(d1e*e(d2e*e(d3e7d4e(d5e(d6e ed7e8d8ede6fd9d:Z9Gd;d<ddd?d@d6e e:d7e8d8ede6fdAdBZr?rrEconv_c)super__init__rConv2d AvgPool2d SequentialprojIdentityintrstochastic_depthrrrSiLUlayers)selfr>r?r@rArrBrCrD should_projrW mid_channels sqz_channels_layers __class__r r!rSDsP          zMBConv.__init__xcCs"||}|||}||S)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. Returns: Tensor: Output tensor with expected layout of [B, C, H / stride, W / stride]. )rWrZr\r]rdresr r r!forwards zMBConv.forward)r=) __name__ __module__ __qualname____doc__rYfloatrrModulerSr rg __classcell__r r rbr!r<6s.    =r<csTeZdZdZdedededdffdd Zdejfd d Zd edefd d Z Z S)$RelativePositionalMultiHeadAttentionzRelative Positional Multi-Head Attention. Args: feat_dim (int): Number of input features. head_dim (int): Number of features per head. max_seq_len (int): Maximum sequence length. feat_dimhead_dim max_seq_lenrNcst||dkrtd|d||||_||_tt||_||_ t ||j|jd|_ |d|_ t |j|j||_t jtjd|jdd|jd|jftjd|_|d t|j|jtj jj|jd d dS) Nrz feat_dim: z must be divisible by head_dim: r$grr)dtyperelative_position_index{Gz?std)rRrS ValueErrorn_headsrqrYmathsqrtsizerrrLinearto_qkv scale_factormerge parameter Parameterr0emptyfloat32relative_position_bias_tableregister_bufferr;init trunc_normal_)r]rprqrrrbr r!rSs    ,z-RelativePositionalMultiHeadAttention.__init__cCs@|jd}|j||j|jd}|ddd}|dS)Nr/rrr)rtviewrrrr5r6 unsqueeze)r] bias_index relative_biasr r r!get_relative_positional_biass  zARelativePositionalMultiHeadAttention.get_relative_positional_biasrdcCs|j\}}}}|j|j}}||}tj|ddd\} } } | |||||ddddd} | |||||ddddd} | |||||ddddd} | |j} t d| | } | } t j | | dd} t d | | }|ddddd||||}| |}|S) z Args: x (Tensor): Input tensor with expected layout of [B, G, P, D]. Returns: Tensor: Output tensor with expected layout of [B, G, P, D]. r$r/)dimrrrz!B G H I D, B G H J D -> B G H I Jz!B G H I J, B G H J D -> B G H I D)shaperyrqr~r0chunkreshaper5reinsumrFsoftmaxr)r]rdBGPDHDHqkvqkvdot_prodpos_biasoutr r r!rgs      z,RelativePositionalMultiHeadAttention.forward) rhrirjrkrYrSr0r rrgrnr r rbr!rosrocsDeZdZdZdededdffdd Zdejdejfd d ZZ S) SwapAxeszPermute the axes of a tensor.abrNcst||_||_dSN)rRrSrr)r]rrrbr r!rSs  zSwapAxes.__init__rdcCst||j|j}|Sr)r0swapaxesrrrer r r!rgszSwapAxes.forward) rhrirjrkrYrSr0r rgrnr r rbr!rsrcs8eZdZdZd fdd Zdededefdd ZZS) WindowPartitionzB Partition the input tensor into non-overlapping windows. rNctdSrrRrSr]rbr r!rSzWindowPartition.__init__rdpcCsf|j\}}}}|}|||||||||}|dddddd}|||||||||}|S)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. p (int): Number of partitions. Returns: Tensor: Output tensor with expected layout of [B, H/P, W/P, P*P, C]. rrrr$rrrr5)r]rdrrCrWrr r r!rgs  zWindowPartition.forwardrN rhrirjrkrSr rYrgrnr r rbr!rsrc s@eZdZdZd fdd Zdededed edef d d ZZS) WindowDepartitionzo Departition the input tensor of non-overlapping windows into a feature volume of layout [B, C, H, W]. rNcrrrrrbr r!rSrzWindowDepartition.__init__rdr h_partitions w_partitionsc Cs`|j\}}}}|} ||} } ||| | | | |}|dddddd}|||| | | | }|S)ar Args: x (Tensor): Input tensor with expected layout of [B, (H/P * W/P), P*P, C]. p (int): Number of partitions. h_partitions (int): Number of vertical partitions. w_partitions (int): Number of horizontal partitions. Returns: Tensor: Output tensor with expected layout of [B, C, H, W]. rrrr$rrr) r]rdrrrrrPPrrHPWPr r r!rgs  zWindowDepartition.forwardrrr r rbr!rs&rcseZdZdZdededededeeefdeded ej fd ed ej fd e d e d e ddffdd Z de de fddZ ZS)PartitionAttentionLayera Layer for partitioning the input tensor into non-overlapping windows and applying attention to each window. Args: in_channels (int): Number of input channels. head_dim (int): Dimension of each attention head. partition_size (int): Size of the partitions. partition_type (str): Type of partitioning to use. Can be either "grid" or "window". grid_size (Tuple[int, int]): Size of the grid to partition the input tensor into. mlp_ratio (int): Ratio of the feature size expansion in the MLP layer. activation_layer (Callable[..., nn.Module]): Activation function to use. norm_layer (Callable[..., nn.Module]): Normalization function to use. attention_dropout (float): Dropout probability for the attention layer. mlp_dropout (float): Dropout probability for the MLP layer. p_stochastic_dropout (float): Probability of dropping out a partition. r>rqpartition_sizepartition_type grid_size mlp_ratiorB.rCattention_dropout mlp_dropoutrDrNc s(t|||_||_|d||_||_||_|dvr"td|dkr/||j|_|_ n|j||_|_ t |_ t |_ |dkrHtddnt|_|dkrVtddnt|_t||t|||dt| |_tt|t||||t|||t| |_t| d d |_dS) Nr)gridwindowz0partition_type must be either 'grid' or 'window'rrrrGrH)rRrSryrq n_partitionsrrrxrgr partition_oprdepartition_oprrrXpartition_swapdepartition_swaprVroDropout attn_layer LayerNormr} mlp_layerrstochastic_dropout) r]r>rqrrrrrBrCrrrDrbr r!rS-s8    z PartitionAttentionLayer.__init__rdcCs|jd|j|jd|j}}t|jd|jdko&|jd|jdkd|j|j|||j}||}||||}||| |}| |}| ||j||}|S)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. Returns: Tensor: Output tensor with expected layout of [B, C, H, W]. rrz[Grid size must be divisible by partition size. Got grid size of {} and partition size of {}) rrr0_assertformatrrrrrrr)r]rdghgwr r r!rggs" &  zPartitionAttentionLayer.forward)rhrirjrkrYstrtuplerrrmrlrSr rgrnr r rbr!rs8     :rcseZdZdZdededededededejfd edejfd ed ed ed ededede eefddffdd Z de de fddZ Z S) MaxVitLayera MaxVit layer consisting of a MBConv layer followed by a PartitionAttentionLayer with `window` and a PartitionAttentionLayer with `grid`. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. expansion_ratio (float): Expansion ratio in the bottleneck. squeeze_ratio (float): Squeeze ratio in the SE Layer. stride (int): Stride of the depthwise convolution. activation_layer (Callable[..., nn.Module]): Activation function. norm_layer (Callable[..., nn.Module]): Normalization function. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Ratio of the MLP layer. mlp_dropout (float): Dropout probability for the MLP layer. attention_dropout (float): Dropout probability for the attention layer. p_stochastic_dropout (float): Probability of stochastic depth. partition_size (int): Size of the partitions. grid_size (Tuple[int, int]): Size of the input feature grid. r>r?rAr@rrC.rBrqrrrrDrrrNc stt}t|||||||| d|d<t||| d|| |tj| | | d |d<t||| d|| |tj| | | d |d<t||_dS)N)r>r?r@rArrBrCrDMBconvr) r>rqrrrrrBrCrrrDwindow_attentionrgrid_attention) rRrSrr<rrrrVr\)r]r>r?rAr@rrCrBrqrrrrDrrr\rbr r!rSsN       zMaxVitLayer.__init__rdcCs||}|Sz Args: x (Tensor): Input tensor of shape (B, C, H, W). Returns: Tensor: Output tensor of shape (B, C, H, W). r\)r]rdr r r!rgs zMaxVitLayer.forward)rhrirjrkrYrlrrrmrrSr rgrnr r rbr!rsD     ArcseZdZdZdedededededejfdedejfd ed ed ed ed ede eefdede eddffdd Z de de fddZ ZS) MaxVitBlocka( A MaxVit block consisting of `n_layers` MaxVit layers. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. expansion_ratio (float): Expansion ratio in the bottleneck. squeeze_ratio (float): Squeeze ratio in the SE Layer. activation_layer (Callable[..., nn.Module]): Activation function. norm_layer (Callable[..., nn.Module]): Normalization function. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Ratio of the MLP layer. mlp_dropout (float): Dropout probability for the MLP layer. attention_dropout (float): Dropout probability for the attention layer. p_stochastic_dropout (float): Probability of stochastic depth. partition_size (int): Size of the partitions. input_grid_size (Tuple[int, int]): Size of the input feature grid. n_layers (int): Number of layers in the block. p_stochastic (List[float]): List of probabilities for stochastic depth for each layer. r>r?rAr@rC.rBrqrrrrinput_grid_sizen_layers p_stochasticrNcstt|| kstd| d|dt|_t| dddd|_t |D]+\}}|dkr2dnd}|jt |dkr>|n|||||||||| | | |j|d g7_q(dS) Nz'p_stochastic must have length n_layers=z, got p_stochastic=.r$rrrFr)r>r?rAr@rrCrBrqrrrrrrD) rRrSlenrxr ModuleListr\r"r enumerater)r]r>r?rAr@rCrBrqrrrrrrridxrrrbr r!rSs4    zMaxVitBlock.__init__rdcCs|jD]}||}q|Srr)r]rdlayerr r r!rg-s  zMaxVitBlock.forward)rhrirjrkrYrlrrrmrlistrSr rgrnr r rbr!rsD      3rc!seZdZdZdejddddddfdeeefded ed eed eed ed e de e dej fde dej fde de dede de deddf fdd Z dedefddZddZZS)ray Implements MaxVit Transformer from the `MaxViT: Multi-Axis Vision Transformer `_ paper. Args: input_size (Tuple[int, int]): Size of the input image. stem_channels (int): Number of channels in the stem. partition_size (int): Size of the partitions. block_channels (List[int]): Number of channels in each block. block_layers (List[int]): Number of layers in each block. stochastic_depth_prob (float): Probability of stochastic depth. Expands to a list of probabilities for each layer that scales linearly to the specified value. squeeze_ratio (float): Squeeze ratio in the SE Layer. Default: 0.25. expansion_ratio (float): Expansion ratio in the MBConv bottleneck. Default: 4. norm_layer (Callable[..., nn.Module]): Normalization function. Default: None (setting to None will produce a `BatchNorm2d(eps=1e-3, momentum=0.01)`). activation_layer (Callable[..., nn.Module]): Activation function Default: nn.GELU. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Expansion ratio of the MLP layer. Default: 4. mlp_dropout (float): Dropout probability for the MLP layer. Default: 0.0. attention_dropout (float): Dropout probability for the attention layer. Default: 0.0. num_classes (int): Number of classes. Default: 1000. Ng?rr=ir stem_channelsrblock_channels block_layersrqstochastic_depth_probrC.rBrAr@rrr num_classesrcstt|d}|durttjddd}t|t|}t|D]%\}}|d|dks6|d|dkrGt d|d|d |d |d q"t t ||dd || d ddt ||dddddd|_ t |dd dd}||_t|_|g|dd}|}td|t|}d}t|||D]+\}}}|jt||| | || || | ||||||||d|jdj}||7}qt tdtt|dt|d|dttj|d|d d|_|dS)Nr$gMbP?g{Gz?)epsmomentumrrz Input size z of block z$ is not divisible by partition size zx. Consider changing the partition size or the input size. Current configuration yields the following block input sizes: rrF)rrCrBrErKT)rrCrBrErFr/)r>r?rAr@rCrBrqrrrrrrr)rE) rRrSrrr BatchNorm2dr*rrrxrVrstemr"rrblocksnplinspacer7tolistzipr&rrAdaptiveAvgPool2dFlattenrr}Tanh classifier _init_weights)r]rrrrrrqrrCrBrAr@rrrrinput_channelsblock_input_sizesrblock_input_sizer>r?rp_idx in_channel out_channel num_layersrbr r!rSNs        zMaxVit.__init__rdcCs,||}|jD]}||}q||}|Sr)rrr)r]rdblockr r r!rgs    zMaxVit.forwardcCs|D]P}t|tjr"tjj|jdd|jdur!tj|jqt|tj r9tj |jdtj |jdqt|tj rTtjj|jdd|jdurTtj|jqdS)Nrurvrr) modules isinstancerrTrnormal_weightrEzeros_r constant_r})r]mr r r!rs      zMaxVit._init_weights)rhrirjrkrGELUrrYrrlrrrmrSr rgrrnr r rbr!r9sZ%      vrFrrrrrrqweightsprogresskwargsc Ks|dur(t|dt|jd|jdd|jddksJt|d|jd|dd} td ||||||| d|} |durK| |j|d d | S) Nr categoriesmin_sizerrrr)rrrrrqrrT)r  check_hashr )rrmetapoprload_state_dictget_state_dict) rrrrrrqr r r rmodelr r r!_maxvits&   rc @sHeZdZedeeddejdedddddd d id d d ddZ e Z dS)rz9https://download.pytorch.org/models/maxvit_t-bc5ab103.pthr) crop_size resize_size interpolationirzLhttps://github.com/pytorch/vision/tree/main/references/classification#maxvitz ImageNet-1KgT@g|?5.X@)zacc@1zacc@5gZd;@gK7]@zThese weights reproduce closely the results of the paper using a similar training recipe. They were trained with a BatchNorm2D momentum of 0.99 instead of the more correct 0.01.)r  num_paramsrrecipe_metrics_ops _file_size_docs)url transformsrN) rhrirjr rrrBICUBICr IMAGENET1K_V1DEFAULTr r r r!rs* r pretrained)r T)r r c Ks2t|}td dgdgdddd||d|S) a Constructs a maxvit_t architecture from `MaxViT: Multi-Axis Vision Transformer `_. Args: weights (:class:`~torchvision.models.MaxVit_T_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.MaxVit_T_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.maxvit.MaxVit`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.MaxVit_T_Weights :members: @)r'i)rrrr g?)rrrrqrrr r Nr )rverifyr)r r r r r r!rs  r)NF)=rz collectionsrcollections.abcr functoolsrtypingrrrnumpyrr0torch.nn.functionalr functionalrr torchvision.models._apir r r torchvision.models._metar torchvision.models._utilsrrtorchvision.ops.miscrr torchvision.ops.stochastic_depthrtorchvision.transforms._presetsrrtorchvision.utilsr__all__rrYr"rr*r;rmr<rorrrrrrrrlboolrrr$rr r r r!sr      .*  WI haU.    *.