models/networks/generator_module/HEtransformerEncoder.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint

from .ViT_helper import to_2tuple, to_ntuple,DropPath

class Mlp(nn.Module):
    """MLP as implemented in timm
    """

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        drops = to_2tuple(drop)

        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.drop1 = nn.Dropout(drops[0])
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop2 = nn.Dropout(drops[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x


class Attention(nn.Module):
    """Self Attention as implemented in timm
    """

    def __init__(self, d_model, nhead=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        assert d_model % nhead == 0, 'd_model needs to be divisible by nhead'
        self.nhead = nhead
        self.scale = (d_model // nhead) ** -0.5

        self.to_qkv = nn.Linear(d_model, d_model * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.size()
        qkv = self.to_qkv(x).reshape(B, N, 3, self.nhead, C // self.nhead).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)

        attn = (q @ k.transpose(-1, -2)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        return x


class Attention_Cross(nn.Module):
    """Attention for decoder layer.Some palce may be called "inter attention"
    """

    def __init__(self, d_model, nhead=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        assert d_model % nhead == 0, 'd_model needs to be divisible by nhead'
        self.nhead = nhead
        self.scale = (d_model // nhead) ** -0.5

        self.to_q = nn.Linear(d_model, d_model, bias=qkv_bias)
        self.to_kv = nn.Linear(d_model, d_model * 2, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x, y):
        """
          Args:
            x: output of the former layer
            y: memery of the encoder layer
        """
        B, Nx, C = x.size()
        _, Ny, _ = y.size()
        q = self.to_q(x).reshape(B, Nx, self.nhead, C // self.nhead).permute(0, 2, 1, 3)
        kv = self.to_kv(y).reshape(B, Ny, 2, self.nhead, C // self.nhead).permute(2, 0, 3, 1, 4)
        k, v = kv.unbind(0)

        attn = (q @ k.transpose(-1, -2)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, Nx, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        return x

def split_int(num):
    """Split an integer into 2 integers evenly
    Args:
      num (int): The input integer
    Returns:
      num_1 (int)
      num_2 (int)
    """
    if num % 2 == 0:
        num_1 = num_2 = num // 2
    else:
        num_1 = num // 2
        num_2 = num_1 + 1
    return num_1, num_2


def unpad2D(input, pad):
    """Crop the input tensor according to pad.(Inverse operation for padding)
    Args:
      input (Tensor): (B, C, H, W)
      pad (Tuple of int): (left, right, top, bottom)
    Returns:
      output (Tensor): (B, C, new_H, new_W)
    """
    pad_W_left, pad_W_right, pad_H_top, pad_H_bottom = pad
    if pad_H_top == 0 and pad_H_bottom == 0 and not (pad_W_left == 0 and pad_W_right == 0):
        output = input[:, :, :, pad_W_left:-pad_W_right]
    elif pad_W_left == 0 and pad_W_right == 0 and not (pad_H_top == 0 and pad_H_bottom == 0):
        output = input[:, :, pad_H_top:-pad_H_bottom, :]
    elif pad_H_top == 0 and pad_H_bottom == 0 and pad_W_left == 0 and pad_W_right == 0:
        output = input
    else:
        output = input[:, :, pad_H_top:-pad_H_bottom, pad_W_left:-pad_W_right]
    return output


def seq_padding(x, dividable_size, input_resolution, pad_mode='constant'):
    """Padding for sequential data
    Args:
      x (Tensor): (B, L, C)
      dividable_size (Tuple | int): dividable size
      input_resolution (Tuple): resolution of x
    Returns:
      x (Tensor): (B, new_L, C)
      output_resolution (Tuple): new resolution of x
      pad (Tuple of int): (left, right, top, bottom)
    """
    H, W = input_resolution
    B, L, C = x.shape
    assert L == H * W, 'Input of wrong size.'
    dividable_size = to_2tuple(dividable_size)
    x = x.permute(0, 2, 1).reshape(B, C, H, W)

    rema_H, rema_W = H % dividable_size[0], W % dividable_size[1]
    pad_H, pad_W = dividable_size[0] - rema_H, dividable_size[1] - rema_W

    pad_H_top, pad_H_bottom = split_int(pad_H) if rema_H != 0 else (0, 0)
    pad_W_left, pad_W_right = split_int(pad_W) if rema_W != 0 else (0, 0)

    x = F.pad(x, (pad_W_left, pad_W_right, pad_H_top, pad_H_bottom), pad_mode, 0)

    padded_H, padded_W = x.shape[-2:]
    x = x.reshape(B, C, -1).permute(0, 2, 1)
    return x, (padded_H, padded_W), (pad_W_left, pad_W_right, pad_H_top, pad_H_bottom)


# Example
# x = torch.randn(1, 14, 768)
# y = seq_padding(x, dividable_size=7, input_resolution=(2, 7))
# print(y[0].shape, y[1], y[2])


def seq_unpad(x, input_resolution, pad):
    """Unpadding for sequential data
    Args:
      x (Tensor): (B, L, C)
      input_resolution (Tuple): resolution of x
      pad (Tuple of int): (left, right, top, bottom)
    Returns:
      x (Tensor): (B, new_L, C)
      output_resolution (Tuple): new resolution of x
    """
    padded_H, padded_W = input_resolution
    B, L, C = x.shape
    assert L == padded_H * padded_W, 'Input of wrong size.'
    x = x.permute(0, 2, 1).reshape(B, C, padded_H, padded_W)

    x = unpad2D(x, pad=pad)

    H, W = x.shape[-2:]
    x = x.reshape(B, C, -1).permute(0, 2, 1)
    return x, (H, W)
def window_partition(x, window_size):
    """Slightly modified for arbitrary window_size & resolution combination
    Args:
      x: (B,H,W,C)
      window_size (tuple[int] | int): window size
    Returns:
      windows: (num_windows*B, window_size, window_size, C)
    """
    window_size = to_2tuple(window_size)
    B, H, W, C = x.shape
    n_win_H = H // window_size[0]
    n_win_W = W // window_size[1]
    if not (H % window_size[0] == 0 and W % window_size[1] == 0):
        x = x[:, :n_win_H * window_size[0], :n_win_W * window_size[1], :]
    x = x.view(B, n_win_H, window_size[0], n_win_W, window_size[1], C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
    return windows


def window_reverse(windows, window_size, H, W):
    """Slightly modified for arbitrary window_size & resolution combination
    Args:
      windows: (num_windows*B, window_size, window_size, C)
      window_size (tuple[int] | int): Window size
      H (int): Height of image
      W (int): Width of image
    Returns:
      x: (B, H, W, C)
    """
    window_size = to_2tuple(window_size)
    n_win_H = H // window_size[0]
    n_win_W = W // window_size[1]
    B = windows.shape[0] // (n_win_H * n_win_W)
    x = windows.view(B, n_win_H, n_win_W, window_size[0], window_size[1], -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, n_win_H * window_size[0], n_win_W * window_size[1], -1)
    return x


# Example
# H, W = 6, 6
# window_size = 2
# x = torch.randn(1, H, W, 3)
# win = window_partition(x, window_size=window_size)
# y = window_reverse(win, window_size=window_size, H=H, W=W)
# print(x.shape)
# print(win.shape)
# print(y.shape)
# print(x[0].permute(2,0,1)[0])
# print(y[0].permute(2,0,1)[0])


def seq_crop(x, dividable_size, input_resolution):
    """
    Arg:
      x (Tensor): (B, L, C)
      dividable_size (Tuple | int): dividable size
      input_resolution (Tuple): resolution of x
    Returns:
      x (Tensor): (B, new_L, C)
      output_resolution (Tuple): new resolution of x
    """
    H, W = input_resolution
    B, L, C = x.shape
    assert L == H * W, 'Input of wrong size.'
    dividable_size = to_2tuple(dividable_size)
    x = x.reshape(B, H, W, C)

    rema_H, rema_W = H % dividable_size[0], W % dividable_size[1]
    new_H, new_W = H - rema_H, W - rema_W
    if rema_H != 0 or rema_W != 0:
        x = x[:, :new_H, :new_W, :]

    x = x.reshape(B, -1, C)
    return x, (new_H, new_W)


# Example
# H, W = 22, 34
# x = torch.randn(2, H*W, 96)
# x, new_size = seq_crop(x, dividable_size=7, input_resolution=(H, W))
# print(x.shape, new_size)


class PatchEmbed_Kai(nn.Module):
    """ 2D Image to Patch Embedding
    """

    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None, flatten=True):
        super().__init__()
        patch_size = to_2tuple(patch_size)
        self.in_chans = in_chans
        self.flatten = flatten

        self.proj = nn.Conv2d(self.in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        B, C, H, W = x.shape
        assert C == self.in_chans, 'Input image need to have same numbers of channels with the initialed.'
        x = self.proj(x)
        H, W = x.shape[2], x.shape[3]
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)
        return x, (H, W)


# Example
# model = PatchEmbed_Kai(patch_size=4, in_chans=3, embed_dim=96)
# x = torch.randn(1, 3, 224, 224)
# y = model(x)
# torch.save(model.state_dict(), "./PatchEmbed.pkl")
# print(y[0].shape, y[1])


class PatchMerging_Kai(nn.Module):
    """ Patch Merging Layer.
    Args:
      input_resolution (tuple[int] | int): Resolution of input feature.
      d_model (int): Number of input channels.
      norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, input_resolution, d_model, norm_layer=nn.LayerNorm):
        super().__init__()
        self.input_resolution = to_2tuple(input_resolution)
        self.d_model = d_model
        self.reduction = nn.Linear(4 * d_model, 2 * d_model, bias=False)
        self.norm = norm_layer(4 * d_model)

    def forward(self, x):
        """
        Args:
          x (Tuple): (Tensor, arbitrary_input, (H,W)), arbitrary_input (bool)
            if arbitrary_input=False, (H,W) will not be required
            B, H*W, C -> B, H/2*W/2, 4*C
        """
        arbitrary_input = x[1]
        if arbitrary_input:
            H, W = x[2]
        else:
            H, W = self.input_resolution

        x = x[0]
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        x = x.view(B, H, W, C)
        if H % 2 != 0:
            x = x[:, 0:-1, :, :]
        if W % 2 != 0:
            x = x[:, :, 0:-1, :]

        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
        H, W = x.shape[1], x.shape[2]
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)

        return x, arbitrary_input, (H, W)


# Example
# model = PatchMerging_Kai(input_resolution=(5,4), d_model=3)
# arbitrary_input = True
# # x = torch.randn(1, 20, 3)
# # y = model((x, arbitrary_input))
# x = torch.randn(1, 45, 3)
# y = model((x, arbitrary_input, (5,9)))
# print(y[0].shape, y[2])


class WindowAttention_Kai(nn.Module):
    def __init__(self, d_model, window_size, nhead, qkv_bias=True, attn_drop=0., proj_drop=0.):
        super().__init__()
        assert d_model % nhead == 0, 'd_model needs to be divisible by nhead'
        self.window_size = to_2tuple(window_size)
        self.nhead = nhead
        self.scale = (d_model // nhead) ** -0.5

        # Relative Position Bias's parameter Table
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1), nhead)
        )

        # Compute indice of relative_position_bias_table for attention matrix
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
        coords_flatten = torch.flatten(coords, 1)
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
        relative_coords[:, :, 0] += self.window_size[0] - 1
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)
        self.register_buffer("relative_position_index", relative_position_index)

        self.qkv = nn.Linear(d_model, d_model * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

        nn.init.trunc_normal_(self.relative_position_bias_table, std=.02)

    def forward(self, x, shape):
        H, W = shape
        Bi, Ni, Ci = x.size()
        assert Ni == H * W, "Inputs with wrong size."
        x = x.reshape(Bi, H, W, Ci)

        # print(self.window_size)
        x = window_partition(x, self.window_size)
        x = x.reshape(-1, self.window_size[0] * self.window_size[1], Ci)

        B_, N, C = x.shape
        qkv = self.qkv(x).reshape(B_, N, 3, self.nhead, C // self.nhead).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)

        attn = (q @ k.transpose(-2, -1)) * self.scale

        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], self.nhead)
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
        attn = attn + relative_position_bias.unsqueeze(0)

        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        x = window_reverse(x, self.window_size, H, W)
        x = x.reshape(Bi, Ni, Ci)
        return x

    # Example


# model = WindowAttention_Kai(
#   d_model=768,
#   window_size=7,
#   nhead=8
# )
# x = torch.randn(1, 784, 768)
# y = model(x, shape=(28, 28))
# print(y.shape)


class StripAttention(nn.Module):
    def __init__(self, d_model, nhead=8, strip_width=7, is_vertical=False, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.d_model = d_model
        self.strip_width = strip_width
        self.is_vertical = is_vertical

        self.attn = Attention(
            d_model=d_model,
            nhead=nhead,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
        )

    def forward(self, x, shape):
        H, W = shape
        B, N, C = x.size()
        assert N == H * W, "Inputs with wrong size."
        x = x.reshape(B, H, W, C)

        # print(self.strip_width)
        if self.is_vertical:
            x = window_partition(x, (H, self.strip_width))
            x = x.reshape(-1, H * self.strip_width, C)
        else:
            x = window_partition(x, (self.strip_width, W))
            x = x.reshape(-1, W * self.strip_width, C)

        wins = self.attn(x)

        if self.is_vertical:
            x = window_reverse(wins, (H, self.strip_width), H, W)
        else:
            x = window_reverse(wins, (self.strip_width, W), H, W)

        x = x.reshape(B, N, C)
        return x


# Example
# model = StripAttention(
#   d_model=768,
#   nhead=8,
#   strip_width=7,
#   is_vertical=False
# )
# x = torch.randn(1, 784, 768)
# y = model(x, (28, 28))
# print(y.shape)


class StripAttentionBlock(nn.Module):
    def __init__(self, d_model, input_resolution, nhead=8, strip_width=7,
                 mlp_ratio=4, qkv_bias=False, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.d_model = d_model
        self.input_resolution = to_2tuple(input_resolution)
        self.strip_width = strip_width

        self.norm1 = norm_layer(d_model)

        self.attn1 = StripAttention(
            d_model=d_model,
            nhead=nhead,
            strip_width=strip_width,
            is_vertical=False,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop
        )
        self.attn2 = StripAttention(
            d_model=d_model,
            nhead=nhead,
            strip_width=strip_width,
            is_vertical=True,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop
        )
        self.attn3 = WindowAttention_Kai(
            d_model=d_model,
            window_size=(strip_width * 2, strip_width * 2),
            nhead=nhead,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(d_model)
        mlp_hidden_dim = int(d_model * mlp_ratio)
        self.mlp = Mlp(d_model, hidden_features=mlp_hidden_dim, out_features=d_model, act_layer=act_layer, drop=drop)

    def forward(self, x):
        arbitrary_input = x[1]
        if arbitrary_input:
            H, W = x[2]
            # x, (H, W) = seq_crop(x[0], dividable_size=self.strip_width*2, input_resolution=(H, W))
            x, (H, W), pad = seq_padding(x[0], dividable_size=self.strip_width * 2, input_resolution=(H, W),
                                         pad_mode='constant')
        else:
            H, W = self.input_resolution
            x = x[0]

        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)

        x1 = self.attn1(x, shape=(H, W))
        x2 = self.attn2(x, shape=(H, W))
        x3 = self.attn3(x, shape=(H, W))

        # Method 1
        # x = x1 + x2 + x3
        # Method 2
        q_x = x.unsqueeze(dim=2)
        k_x = torch.stack([x, x1, x2, x3], dim=2)
        attn_x = (q_x @ k_x.transpose(-1, -2)).softmax(dim=-1)
        x = attn_x @ k_x
        x = x.squeeze(dim=2)
        x = shortcut + self.drop_path(x)

        # FFN
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        if arbitrary_input:
            x, (H, W) = seq_unpad(x, (H, W), pad)

        return (x, arbitrary_input, (H, W))


# Example
# model = StripAttentionBlock(
#   d_model=96,
#   input_resolution=28,
#   nhead=8,
#   strip_width=7
# )
# arbitrary_input = True
# # x = (torch.randn(1, 784, 96), arbitrary_input, (28,28))
# # y = model(x)
# x = (torch.randn(1, 840, 96), arbitrary_input, (28, 30))
# y = model(x)
# print(y[0].shape, y[2])


class BasicLayer_SA(nn.Module):
    def __init__(self, d_model, input_resolution, depth, nhead, strip_width,
                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
        super().__init__()
        self.d_model = d_model
        self.input_resolution = to_2tuple(input_resolution)
        self.depth = depth
        self.strip_width = list(to_ntuple(self.depth)(strip_width))
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList([
            StripAttentionBlock(
                d_model=d_model,
                input_resolution=self.input_resolution,
                nhead=nhead,
                strip_width=self.strip_width[i],
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                drop=drop,
                attn_drop=attn_drop,
                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                norm_layer=norm_layer
            )
            for i in range(self.depth)
        ])

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(self.input_resolution, d_model=d_model, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        # print(x[0].shape, x[1], x[2])
        if self.downsample is not None:
            x = self.downsample(x)
        return x


# Example
# model = BasicLayer_SA(
#   d_model=768,
#   input_resolution=112,
#   depth=3,
#   nhead=8,
#   strip_width=[7, 2, 7],
#   drop_path=0.,
#   downsample=PatchMerging_Kai,
#   use_checkpoint=False
# )
# arbitrary_input = True
# # x = (torch.randn(1, 12544, 768), arbitrary_input, (112,112))
# # y = model(x)
# x = (torch.randn(1, 810, 768), arbitrary_input, (27,30))
# y = model(x)
# print(y[0].shape, y[2])

########################################## S2WAT ##########################################

class HeTransformerEncoder(nn.Module):
    def __init__(self, img_size=224, patch_size=4, in_chans=3,
                 embed_dim=96, depths=[2, 2, 6, 2], nhead=[3, 6, 12, 24],
                 strip_width=7, mlp_ratio=4., qkv_bias=True,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
                 use_checkpoint=False):
        super().__init__()

        self.img_size = to_2tuple(img_size)
        self.patch_size = to_2tuple(patch_size)
        self.num_layers = len(depths)
        self.strip_width = list(to_ntuple(self.num_layers)(strip_width))
        self.embed_dim = embed_dim
        self.ape = ape
        self.printed_modes = set()
        self.patch_norm = patch_norm
        self.device="cuda" if torch.cuda.is_available() else "cpu"

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed_Kai(
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None
        )
        self.patches_resolution = (self.img_size[0] // self.patch_size[0], self.img_size[1] // self.patch_size[1])
        self.num_patches = self.patches_resolution[0] * self.patches_resolution[1]

        # absolute position embedding
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, self.num_patches, embed_dim))
            nn.init.trunc_normal_(self.absolute_pos_embed, std=.02)
        self.pos_drop = nn.Dropout(drop_rate)

        # stochastic depth
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        for i in range(self.num_layers):
            layer = BasicLayer_SA(
                d_model=int(self.embed_dim * 2 ** i),
                input_resolution=(self.patches_resolution[0] // (2 ** i),
                                  self.patches_resolution[1] // (2 ** i)),
                depth=depths[i],
                nhead=nhead[i],
                strip_width=self.strip_width[i],
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
                norm_layer=norm_layer,
                downsample=PatchMerging_Kai if (i < self.num_layers - 1) else None,
                use_checkpoint=use_checkpoint
            )
            self.layers.append(layer)

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            nn.init.trunc_normal_(m.weight, std=.02)
            if m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'absolute_pos_embed'}

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {'relative_position_bias_table'}


    def add_z(self, x, alpha, mode):
        if mode not in self.printed_modes:
            if mode == 1:
                print("Using Hadamard Product => Element-wise multiplication")
            elif mode == 2:
                print("Using Addition")
            elif mode == 3:
                print("Joint Embedding (concatenation along a new dimension)")
            else:
                raise ValueError("Invalid mode. Please choose 1, 2, or 3")

            self.printed_modes.add(mode)

        size = x[0].size()
        alpha_exp = alpha.expand(size)

        if mode == 1:
            x_concat_alpha = x[0].to(self.device) * alpha_exp.to(self.device)
        elif mode == 2:
            x_concat_alpha = x[0].to(self.device) + alpha_exp.to(self.device)
        elif mode == 3:
            x_concat_alpha = torch.cat((x[0].to(self.device), alpha_exp.to(self.device)))
        else:
            raise ValueError("Invalid mode. Please choose 1, 2, or 3")

        return x_concat_alpha.to(self.device)


    def forward_features(self, x,alpha,mode):
        x, arbitrary_input = x[0], x[1]
        x, (H, W) = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)

        x = (x, arbitrary_input, (H, W))
        for layer in self.layers:
            x = layer(x)
            x = (self.add_z(x, alpha, mode), x[1], x[2])

        return x

    def forward(self, x,arbitrary_input=False,alpha=None,mode=2):
        if arbitrary_input:
            H, W = x.shape[2], x.shape[3]
            x = (x, arbitrary_input, (H, W))
        else:
            x = (x, arbitrary_input)

        x = self.forward_features(x,alpha,mode)

        return x