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$ej-Z1Gd%d&d&ej-Z2Gd'd(d(ej-Z3Gd)d*d*ej-Z4Gd+d,d,ej-Z5Gd-d.d.ej-Z6  /dCd0e(d1e*e(d2e*e(d3e7d4e(d5e(d6e ed7e8d8ede6fd9d:Z9Gd;d<ddd?d@d6e e:d7e8d8ede6fdAdBZ B G H I Jz!B G H I J, B G H J D -> B G H I D)shaperbr\rgr,chunkreshaper/rhZeinsumroFZsoftmaxri)rLrOBGPDHZDHZqkvqkvZdot_prodZpos_biasoutr r r!rRs      z,RelativePositionalMultiHeadAttention.forward) rSrTrUrVrIrDr,r rorRrYr r rMr!rZsrZcsDeZdZdZdededdffdd Zdejdejfd d ZZ S) SwapAxeszPermute the axes of a tensor.abrNcst||_||_dSN)rCrDrr)rLrrrMr r!rDs  zSwapAxes.__init__rOcCst||j|j}|Sr)r,ZswapaxesrrrPr r r!rRszSwapAxes.forward) rSrTrUrVrIrDr,r rRrYr r rMr!rsrcs8eZdZdZd fdd Zdededefdd ZZS) WindowPartitionzB Partition the input tensor into non-overlapping windows. rNctdSrrCrDrLrMr r!rDzWindowPartition.__init__rOpcCsf|j\}}}}|}|||||||||}|dddddd}|||||||||}|S)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. p (int): Number of partitions. Returns: Tensor: Output tensor with expected layout of [B, H/P, W/P, P*P, C]. rrrqr$rrrrtr/)rLrOrrvCrzWrxr r r!rRs  zWindowPartition.forwardrN rSrTrUrVrDr rIrRrYr r rMr!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]. rNcrrrrrMr r!rDrzWindowDepartition.__init__rOr 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$rrqr) rLrOrrrrvrwZPPrrxZHPZWPr r r!rRs  zWindowDepartition.forwardrrr r rMr!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. r5r\partition_sizepartition_type grid_size mlp_ratior9.r:attention_dropout mlp_dropoutr;rNc 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'rrrr>r?)rCrDrbr\Z n_partitionsrrrargr partition_oprdepartition_oprrrHpartition_swapdepartition_swaprFrZZDropout attn_layer LayerNormrf mlp_layerrstochastic_dropout) rLr5r\rrrrr9r:rrr;rMr r!rD-s8    z PartitionAttentionLayer.__init__rOcCs|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 {}) rrr,Z_assertformatrrrrrrr)rLrOghZgwr r r!rRgs" &  zPartitionAttentionLayer.forward)rSrTrUrVrIstrtuplerrrXrWrDr rRrYr r rMr!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. r5r6r8r7rr:.r9r\rrrr;rrrNc stt}t|||||||| d|d<t||| d|| |tj| | | d |d<t||| d|| |tj| | | d |d<t||_dS)N)r5r6r7r8rr9r:r;ZMBconvr) r5r\rrrrr9r:rrr;Zwindow_attentionrZgrid_attention) rCrDrr3rrrrFrK)rLr5r6r8r7rr:r9r\rrrr;rrrKrMr r!rDsN       zMaxVitLayer.__init__rOcCs||}|Sz Args: x (Tensor): Input tensor of shape (B, C, H, W). Returns: Tensor: Output tensor of shape (B, C, H, W). rK)rLrOr r r!rRs zMaxVitLayer.forward)rSrTrUrVrIrWrrrXrrDr rRrYr r rMr!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. r5r6r8r7r:.r9r\rrrrinput_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$rrr=r)r5r6r8r7rr:r9r\rrrrrr;) rCrDlenrar ModuleListrKr"r enumerater)rLr5r6r8r7r:r9r\rrrrrrridxrrrMr r!rDs4    zMaxVitBlock.__init__rOcCs|jD]}||}q|Srr)rLrOlayerr r r!rR-s  zMaxVitBlock.forward)rSrTrUrVrIrWrrrXrlistrDr rRrYr r rMr!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?rqr4ir stem_channelsrblock_channels block_layersr\stochastic_depth_probr:.r9r8r7rrr 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$gMbP?g{Gz?)epsZmomentumrrz Input size z of block z$ 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)rr:r9r<rAT)rr:r9r<r=r+)r5r6r8r7r:r9r\rrrrrrr)r<) rCrDrrr BatchNorm2dr(rrrarFrstemr"rrblocksnpZlinspacer1tolistzipr&rrZAdaptiveAvgPool2dZFlattenrrfZTanh classifier _init_weights)rLrrrrrr\rr:r9r8r7rrrrZinput_channelsZblock_input_sizesrZblock_input_sizer5r6rZp_idxZ in_channelZ out_channelZ num_layersrMr r!rDNs        zMaxVit.__init__rOcCs,||}|jD]}||}q||}|Sr)rrr)rLrOblockr r r!rRs    zMaxVit.forwardcCs|D]P}t|tjr"tjj|jdd|jdur!tj|jqt|tj r9tj |jdtj |jdqt|tj rTtjj|jdd|jdurTtj|jqdS)Nr_r`rr) modules isinstancerrErmZnormal_weightr<Zzeros_rZ constant_rf)rLmr r r!rs      zMaxVit._init_weights)rSrTrUrVrZGELUrrIrrWrrrXrDr rRrrYr r rMr!r9sZ%      vrFrrrrrr\weightsprogresskwargsc Ks|dur(t|dt|jd|jdd|jddksJt|d|jd|dd} td ||||||| d|} |durK| |j|d d | S) Nr categoriesmin_sizerrrr)rrrrr\rrT)rZ check_hashr )rrmetapoprZload_state_dictZget_state_dict) rrrrrr\rrrrmodelr r r!_maxvits&   rc @sHeZdZedeeddejdedddddd d id d d ddZ e Z dS)rz9https://download.pytorch.org/models/maxvit_t-bc5ab103.pthr)Z crop_sizeZ 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.)rZ num_paramsrZrecipeZ_metricsZ_ops _file_sizeZ_docs)urlZ transformsrN) rSrTrUr rrrZBICUBICr IMAGENET1K_V1DEFAULTr r r r!rs* rZ pretrained)rT)rrc 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: @)ri)rrrr g?)rrrr\rrrrNr )rverifyr)rrrr r r!rs  r)NF)=rc collectionsrcollections.abcr functoolsrtypingrrrnumpyrr,Ztorch.nn.functionalrZ functionalrur Ztorchvision.models._apir r r Ztorchvision.models._metar Ztorchvision.models._utilsrrZtorchvision.ops.miscrrZ torchvision.ops.stochastic_depthrZtorchvision.transforms._presetsrrZtorchvision.utilsr__all__rrIr"rr(r2rXr3rZrrrrrrrrWboolrrrrr r r r!sr      .*  WI haU.    *.