Upload folder using huggingface_hub

.gitattributes

@@ -341,3 +341,4 @@ models/sapiens/depth/sapiens_1b_render_people_epoch_88_torchscript.pt2 filter=lf
 models/sapiens/normal/sapiens_1b_normal_render_people_epoch_115_torchscript.pt2 filter=lfs diff=lfs merge=lfs -text
 models/sapiens/pose/sapiens_1b_goliath_best_goliath_AP_639_torchscript.pt2 filter=lfs diff=lfs merge=lfs -text
 models/sapiens/seg/sapiens_1b_goliath_best_goliath_mIoU_7994_epoch_151_torchscript.pt2 filter=lfs diff=lfs merge=lfs -text
+models/RMBG/SAM/Moonlit[[:space:]]Serenade.mp3 filter=lfs diff=lfs merge=lfs -text

models/RMBG/BEN/BEN_Base.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f00a5804c96afed6e19be09dbbbe56ccaf82cff3d751f44aadb9626b77facfa
+size 1134588350

models/RMBG/BEN/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/BEN/model.py

@@ -0,0 +1,951 @@
+import math
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from einops import rearrange
+from PIL import Image, ImageFilter, ImageOps
+from timm.layers import DropPath, to_2tuple, trunc_normal_
+from torchvision import transforms
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        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], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_chans=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+
+        outs = [x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+
+
+        return tuple(outs)
+
+
+
+
+
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "gelu":
+        return F.gelu
+
+    raise RuntimeError(F"activation should be gelu, not {activation}.")
+
+
+def make_cbr(in_dim, out_dim):
+    return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU())
+
+
+def make_cbg(in_dim, out_dim):
+    return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU())
+
+
+def rescale_to(x, scale_factor: float = 2, interpolation='nearest'):
+    return F.interpolate(x, scale_factor=scale_factor, mode=interpolation)
+
+
+def resize_as(x, y, interpolation='bilinear'):
+    return F.interpolate(x, size=y.shape[-2:], mode=interpolation)
+
+
+def image2patches(x):
+    """b c (hg h) (wg w) -> (hg wg b) c h w"""
+    x = rearrange(x, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2)
+    return x
+
+
+def patches2image(x):
+    """(hg wg b) c h w -> b c (hg h) (wg w)"""
+    x = rearrange(x, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2)
+    return x
+class PositionEmbeddingSine:
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+        self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32)
+
+    def __call__(self, b, h, w):
+        device = self.dim_t.device
+        mask = torch.zeros([b, h, w], dtype=torch.bool, device=device)
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(dim=1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(dim=2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+
+        return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+
+
+class MCLM(nn.Module):
+    def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]):
+        super(MCLM, self).__init__()
+        self.attention = nn.ModuleList([
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1)
+        ])
+
+        self.linear1 = nn.Linear(d_model, d_model * 2)
+        self.linear2 = nn.Linear(d_model * 2, d_model)
+        self.linear3 = nn.Linear(d_model, d_model * 2)
+        self.linear4 = nn.Linear(d_model * 2, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout = nn.Dropout(0.1)
+        self.dropout1 = nn.Dropout(0.1)
+        self.dropout2 = nn.Dropout(0.1)
+        self.activation = get_activation_fn('gelu')
+        self.pool_ratios = pool_ratios
+        self.p_poses = []
+        self.g_pos = None
+        self.positional_encoding = PositionEmbeddingSine(num_pos_feats=d_model // 2, normalize=True)
+
+    def forward(self, l, g):
+        """
+        l: 4,c,h,w
+        g: 1,c,h,w
+        """
+        b, c, h, w = l.size()
+        # 4,c,h,w -> 1,c,2h,2w
+        concated_locs = rearrange(l, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2)
+
+        pools = []
+        for pool_ratio in self.pool_ratios:
+             # b,c,h,w
+            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
+            pool = F.adaptive_avg_pool2d(concated_locs, tgt_hw)
+            pools.append(rearrange(pool, 'b c h w -> (h w) b c'))
+            if self.g_pos is None:
+                pos_emb = self.positional_encoding(pool.shape[0], pool.shape[2], pool.shape[3])
+                pos_emb = rearrange(pos_emb, 'b c h w -> (h w) b c')
+                self.p_poses.append(pos_emb)
+        pools = torch.cat(pools, 0)
+        if self.g_pos is None:
+            self.p_poses = torch.cat(self.p_poses, dim=0)
+            pos_emb = self.positional_encoding(g.shape[0], g.shape[2], g.shape[3])
+            self.g_pos = rearrange(pos_emb, 'b c h w -> (h w) b c')
+
+        device = pools.device
+        self.p_poses = self.p_poses.to(device)
+        self.g_pos = self.g_pos.to(device)
+
+
+        # attention between glb (q) & multisensory concated-locs (k,v)
+        g_hw_b_c = rearrange(g, 'b c h w -> (h w) b c')
+
+
+        g_hw_b_c = g_hw_b_c + self.dropout1(self.attention[0](g_hw_b_c + self.g_pos, pools + self.p_poses, pools)[0])
+        g_hw_b_c = self.norm1(g_hw_b_c)
+        g_hw_b_c = g_hw_b_c + self.dropout2(self.linear2(self.dropout(self.activation(self.linear1(g_hw_b_c)).clone())))
+        g_hw_b_c = self.norm2(g_hw_b_c)
+
+        # attention between origin locs (q) & freashed glb (k,v)
+        l_hw_b_c = rearrange(l, "b c h w -> (h w) b c")
+        _g_hw_b_c = rearrange(g_hw_b_c, '(h w) b c -> h w b c', h=h, w=w)
+        _g_hw_b_c = rearrange(_g_hw_b_c, "(ng h) (nw w) b c -> (h w) (ng nw b) c", ng=2, nw=2)
+        outputs_re = []
+        for i, (_l, _g) in enumerate(zip(l_hw_b_c.chunk(4, dim=1), _g_hw_b_c.chunk(4, dim=1))):
+            outputs_re.append(self.attention[i + 1](_l, _g, _g)[0])  # (h w) 1 c
+        outputs_re = torch.cat(outputs_re, 1)  # (h w) 4 c
+
+        l_hw_b_c = l_hw_b_c + self.dropout1(outputs_re)
+        l_hw_b_c = self.norm1(l_hw_b_c)
+        l_hw_b_c = l_hw_b_c + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(l_hw_b_c)).clone())))
+        l_hw_b_c = self.norm2(l_hw_b_c)
+
+        l = torch.cat((l_hw_b_c, g_hw_b_c), 1)  # hw,b(5),c
+        return rearrange(l, "(h w) b c -> b c h w", h=h, w=w)  ## (5,c,h*w)
+
+
+
+
+
+
+
+
+
+class MCRM(nn.Module):
+    def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None):
+        super(MCRM, self).__init__()
+        self.attention = nn.ModuleList([
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1)
+        ])
+        self.linear3 = nn.Linear(d_model, d_model * 2)
+        self.linear4 = nn.Linear(d_model * 2, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout = nn.Dropout(0.1)
+        self.dropout1 = nn.Dropout(0.1)
+        self.dropout2 = nn.Dropout(0.1)
+        self.sigmoid = nn.Sigmoid()
+        self.activation = get_activation_fn('gelu')
+        self.sal_conv = nn.Conv2d(d_model, 1, 1)
+        self.pool_ratios = pool_ratios
+
+    def forward(self, x):
+        device = x.device
+        b, c, h, w = x.size()
+        loc, glb = x.split([4, 1], dim=0)  # 4,c,h,w; 1,c,h,w
+
+        patched_glb = rearrange(glb, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2)
+
+        token_attention_map = self.sigmoid(self.sal_conv(glb))
+        token_attention_map = F.interpolate(token_attention_map, size=patches2image(loc).shape[-2:], mode='nearest')
+        loc = loc * rearrange(token_attention_map, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2)
+
+        pools = []
+        for pool_ratio in self.pool_ratios:
+            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
+            pool = F.adaptive_avg_pool2d(patched_glb, tgt_hw)
+            pools.append(rearrange(pool, 'nl c h w -> nl c (h w)'))  # nl(4),c,hw
+
+        pools = rearrange(torch.cat(pools, 2), "nl c nphw -> nl nphw 1 c")
+        loc_ = rearrange(loc, 'nl c h w -> nl (h w) 1 c')
+
+        outputs = []
+        for i, q in enumerate(loc_.unbind(dim=0)):  # traverse all local patches
+            v = pools[i]
+            k = v
+            outputs.append(self.attention[i](q, k, v)[0])
+
+        outputs = torch.cat(outputs, 1)
+        src = loc.view(4, c, -1).permute(2, 0, 1) + self.dropout1(outputs)
+        src = self.norm1(src)
+        src = src + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(src)).clone())))
+        src = self.norm2(src)
+        src = src.permute(1, 2, 0).reshape(4, c, h, w)  # freshed loc
+        glb = glb + F.interpolate(patches2image(src), size=glb.shape[-2:], mode='nearest')  # freshed glb
+
+        return torch.cat((src, glb), 0), token_attention_map
+
+
+class BEN_Base(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.backbone = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+        emb_dim = 128
+        self.sideout5 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout4 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout3 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout2 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout1 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+
+        self.output5 = make_cbr(1024, emb_dim)
+        self.output4 = make_cbr(512, emb_dim)
+        self.output3 = make_cbr(256, emb_dim)
+        self.output2 = make_cbr(128, emb_dim)
+        self.output1 = make_cbr(128, emb_dim)
+
+        self.multifieldcrossatt = MCLM(emb_dim, 1, [1, 4, 8])
+        self.conv1 = make_cbr(emb_dim, emb_dim)
+        self.conv2 = make_cbr(emb_dim, emb_dim)
+        self.conv3 = make_cbr(emb_dim, emb_dim)
+        self.conv4 = make_cbr(emb_dim, emb_dim)
+        self.dec_blk1 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk2 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk3 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk4 = MCRM(emb_dim, 1, [2, 4, 8])
+
+        self.insmask_head = nn.Sequential(
+            nn.Conv2d(emb_dim, 384, kernel_size=3, padding=1),
+            nn.InstanceNorm2d(384),
+            nn.GELU(),
+            nn.Conv2d(384, 384, kernel_size=3, padding=1),
+            nn.InstanceNorm2d(384),
+            nn.GELU(),
+            nn.Conv2d(384, emb_dim, kernel_size=3, padding=1)
+        )
+
+        self.shallow = nn.Sequential(nn.Conv2d(3, emb_dim, kernel_size=3, padding=1))
+        self.upsample1 = make_cbg(emb_dim, emb_dim)
+        self.upsample2 = make_cbg(emb_dim, emb_dim)
+        self.output = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+
+        for m in self.modules():
+            if isinstance(m, nn.GELU) or isinstance(m, nn.Dropout):
+                m.inplace = True
+
+    def forward(self, x):
+        device = x.device
+        shallow = self.shallow(x)
+        glb = rescale_to(x, scale_factor=0.5, interpolation='bilinear')
+        loc = image2patches(x)
+        input = torch.cat((loc, glb), dim=0)
+        feature = self.backbone(input)
+        e5 = self.output5(feature[4])  # (5,128,16,16)
+        e4 = self.output4(feature[3])  # (5,128,32,32)
+        e3 = self.output3(feature[2])  # (5,128,64,64)
+        e2 = self.output2(feature[1])  # (5,128,128,128)
+        e1 = self.output1(feature[0])  # (5,128,128,128)
+        loc_e5, glb_e5 = e5.split([4, 1], dim=0)
+        e5 = self.multifieldcrossatt(loc_e5, glb_e5)  # (4,128,16,16)
+
+        e4, tokenattmap4 = self.dec_blk4(e4 + resize_as(e5, e4))
+        e4 = self.conv4(e4)
+        e3, tokenattmap3 = self.dec_blk3(e3 + resize_as(e4, e3))
+        e3 = self.conv3(e3)
+        e2, tokenattmap2 = self.dec_blk2(e2 + resize_as(e3, e2))
+        e2 = self.conv2(e2)
+        e1, tokenattmap1 = self.dec_blk1(e1 + resize_as(e2, e1))
+        e1 = self.conv1(e1)
+        loc_e1, glb_e1 = e1.split([4, 1], dim=0)
+        output1_cat = patches2image(loc_e1)  # (1,128,256,256)
+        output1_cat = output1_cat + resize_as(glb_e1, output1_cat)
+        final_output = self.insmask_head(output1_cat)  # (1,128,256,256)
+        final_output = final_output + resize_as(shallow, final_output)
+        final_output = self.upsample1(rescale_to(final_output))
+        final_output = rescale_to(final_output + resize_as(shallow, final_output))
+        final_output = self.upsample2(final_output)
+        final_output = self.output(final_output)
+
+        return final_output.sigmoid()
+
+    @torch.no_grad()
+    def inference(self,image):
+        image, h, w,original_image =  rgb_loader_refiner(image)
+
+        img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device)
+
+        res = self.forward(img_tensor)
+
+        pred_array = postprocess_image(res, im_size=[w, h])
+
+        mask_image = Image.fromarray(pred_array, mode='L')
+
+        blurred_mask = mask_image.filter(ImageFilter.GaussianBlur(radius=1))
+
+        original_image_rgba = original_image.convert("RGBA")
+
+        foreground = original_image_rgba.copy()
+
+        foreground.putalpha(blurred_mask)
+
+        return blurred_mask, foreground
+
+    def loadcheckpoints(self,model_path):
+        model_dict = torch.load(model_path, map_location="cpu", weights_only=True)
+        self.load_state_dict(model_dict['model_state_dict'], strict=True)
+        del model_path
+
+
+
+
+def rgb_loader_refiner( original_image):
+        h, w = original_image.size
+        # # Apply EXIF orientation
+        image = ImageOps.exif_transpose(original_image)
+        # Convert to RGB if necessary
+        if image.mode != 'RGB':
+            image = image.convert('RGB')
+
+        # Resize the image
+        image = image.resize((1024, 1024), resample=Image.LANCZOS)
+
+        return image.convert('RGB'), h, w,original_image
+
+# Define the image transformation
+img_transform = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.ConvertImageDtype(torch.float32),
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+])
+
+def postprocess_image(result: torch.Tensor, im_size: list) -> np.ndarray:
+    result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear'), 0)
+    ma = torch.max(result)
+    mi = torch.min(result)
+    result = (result - mi) / (ma - mi)
+    im_array = (result * 255).permute(1, 2, 0).cpu().data.numpy().astype(np.uint8)
+    im_array = np.squeeze(im_array)
+    return im_array
+
+
+
+

models/RMBG/BEN2/BEN2.py

@@ -0,0 +1,1401 @@
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers  import DropPath, to_2tuple, trunc_normal_
+from PIL import Image, ImageOps
+from torchvision import transforms
+import numpy as np
+import random
+import cv2
+import os 
+import subprocess
+import time
+import tempfile
+
+
+
+
+def set_random_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+set_random_seed(9)
+
+
+torch.set_float32_matmul_precision('highest')
+
+
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        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], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_chans=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+
+        outs = [x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+
+
+        return tuple(outs)
+
+
+
+
+
+
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "gelu":
+        return F.gelu
+
+    raise RuntimeError(F"activation should be gelu, not {activation}.")
+
+
+def make_cbr(in_dim, out_dim):
+    return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU())
+
+
+def make_cbg(in_dim, out_dim):
+    return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU())
+
+
+def rescale_to(x, scale_factor: float = 2, interpolation='nearest'):
+    return F.interpolate(x, scale_factor=scale_factor, mode=interpolation)
+
+
+def resize_as(x, y, interpolation='bilinear'):
+    return F.interpolate(x, size=y.shape[-2:], mode=interpolation)
+
+
+def image2patches(x):
+    """b c (hg h) (wg w) -> (hg wg b) c h w"""
+    x = rearrange(x, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2 )
+    return x
+
+
+def patches2image(x):
+    """(hg wg b) c h w -> b c (hg h) (wg w)"""
+    x = rearrange(x, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2)
+    return x
+
+
+
+class PositionEmbeddingSine:
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+        self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32)
+
+    def __call__(self, b, h, w):
+        device = self.dim_t.device
+        mask = torch.zeros([b, h, w], dtype=torch.bool, device=device)
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(dim=1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(dim=2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+
+        return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+
+
+
+class PositionEmbeddingSine:
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+        self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32)
+
+    def __call__(self, b, h, w):
+        device = self.dim_t.device
+        mask = torch.zeros([b, h, w], dtype=torch.bool, device=device)
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(dim=1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(dim=2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+
+        return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+
+
+class MCLM(nn.Module):
+    def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]):
+        super(MCLM, self).__init__()
+        self.attention = nn.ModuleList([
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1)
+        ])
+
+        self.linear1 = nn.Linear(d_model, d_model * 2)
+        self.linear2 = nn.Linear(d_model * 2, d_model)
+        self.linear3 = nn.Linear(d_model, d_model * 2)
+        self.linear4 = nn.Linear(d_model * 2, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout = nn.Dropout(0.1)
+        self.dropout1 = nn.Dropout(0.1)
+        self.dropout2 = nn.Dropout(0.1)
+        self.activation = get_activation_fn('gelu')
+        self.pool_ratios = pool_ratios
+        self.p_poses = []
+        self.g_pos = None
+        self.positional_encoding = PositionEmbeddingSine(num_pos_feats=d_model // 2, normalize=True)
+
+    def forward(self, l, g):
+        """
+        l: 4,c,h,w
+        g: 1,c,h,w
+        """
+        self.p_poses = []
+        self.g_pos = None 
+        b, c, h, w = l.size()
+        # 4,c,h,w -> 1,c,2h,2w
+        concated_locs = rearrange(l, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2)
+
+        pools = []
+        for pool_ratio in self.pool_ratios:
+             # b,c,h,w
+            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
+            pool = F.adaptive_avg_pool2d(concated_locs, tgt_hw)
+            pools.append(rearrange(pool, 'b c h w -> (h w) b c'))
+            if self.g_pos is None:
+                pos_emb = self.positional_encoding(pool.shape[0], pool.shape[2], pool.shape[3])
+                pos_emb = rearrange(pos_emb, 'b c h w -> (h w) b c')
+                self.p_poses.append(pos_emb)
+        pools = torch.cat(pools, 0)
+        if self.g_pos is None:
+            self.p_poses = torch.cat(self.p_poses, dim=0)
+            pos_emb = self.positional_encoding(g.shape[0], g.shape[2], g.shape[3])
+            self.g_pos = rearrange(pos_emb, 'b c h w -> (h w) b c')
+
+        device = pools.device
+        self.p_poses = self.p_poses.to(device)
+        self.g_pos = self.g_pos.to(device)
+
+
+        # attention between glb (q) & multisensory concated-locs (k,v)
+        g_hw_b_c = rearrange(g, 'b c h w -> (h w) b c')
+
+
+        g_hw_b_c = g_hw_b_c + self.dropout1(self.attention[0](g_hw_b_c + self.g_pos, pools + self.p_poses, pools)[0])
+        g_hw_b_c = self.norm1(g_hw_b_c)
+        g_hw_b_c = g_hw_b_c + self.dropout2(self.linear2(self.dropout(self.activation(self.linear1(g_hw_b_c)).clone())))
+        g_hw_b_c = self.norm2(g_hw_b_c)
+
+        # attention between origin locs (q) & freashed glb (k,v)
+        l_hw_b_c = rearrange(l, "b c h w -> (h w) b c")
+        _g_hw_b_c = rearrange(g_hw_b_c, '(h w) b c -> h w b c', h=h, w=w)
+        _g_hw_b_c = rearrange(_g_hw_b_c, "(ng h) (nw w) b c -> (h w) (ng nw b) c", ng=2, nw=2)
+        outputs_re = []
+        for i, (_l, _g) in enumerate(zip(l_hw_b_c.chunk(4, dim=1), _g_hw_b_c.chunk(4, dim=1))):
+            outputs_re.append(self.attention[i + 1](_l, _g, _g)[0])  # (h w) 1 c
+        outputs_re = torch.cat(outputs_re, 1)  # (h w) 4 c
+
+        l_hw_b_c = l_hw_b_c + self.dropout1(outputs_re)
+        l_hw_b_c = self.norm1(l_hw_b_c)
+        l_hw_b_c = l_hw_b_c + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(l_hw_b_c)).clone())))
+        l_hw_b_c = self.norm2(l_hw_b_c)
+
+        l = torch.cat((l_hw_b_c, g_hw_b_c), 1)  # hw,b(5),c
+        return rearrange(l, "(h w) b c -> b c h w", h=h, w=w)  ## (5,c,h*w)
+
+
+
+
+
+
+
+
+
+class MCRM(nn.Module):
+    def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None):
+        super(MCRM, self).__init__()
+        self.attention = nn.ModuleList([
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
+            nn.MultiheadAttention(d_model, num_heads, dropout=0.1)
+        ])
+        self.linear3 = nn.Linear(d_model, d_model * 2)
+        self.linear4 = nn.Linear(d_model * 2, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout = nn.Dropout(0.1)
+        self.dropout1 = nn.Dropout(0.1)
+        self.dropout2 = nn.Dropout(0.1)
+        self.sigmoid = nn.Sigmoid()
+        self.activation = get_activation_fn('gelu')
+        self.sal_conv = nn.Conv2d(d_model, 1, 1)
+        self.pool_ratios = pool_ratios
+
+    def forward(self, x):
+        device = x.device
+        b, c, h, w = x.size()
+        loc, glb = x.split([4, 1], dim=0)  # 4,c,h,w; 1,c,h,w
+
+        patched_glb = rearrange(glb, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2)
+
+        token_attention_map = self.sigmoid(self.sal_conv(glb))
+        token_attention_map = F.interpolate(token_attention_map, size=patches2image(loc).shape[-2:], mode='nearest')
+        loc = loc * rearrange(token_attention_map, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2)
+
+        pools = []
+        for pool_ratio in self.pool_ratios:
+            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
+            pool = F.adaptive_avg_pool2d(patched_glb, tgt_hw)
+            pools.append(rearrange(pool, 'nl c h w -> nl c (h w)'))  # nl(4),c,hw
+
+        pools = rearrange(torch.cat(pools, 2), "nl c nphw -> nl nphw 1 c")
+        loc_ = rearrange(loc, 'nl c h w -> nl (h w) 1 c')
+
+        outputs = []
+        for i, q in enumerate(loc_.unbind(dim=0)):  # traverse all local patches
+            v = pools[i]
+            k = v
+            outputs.append(self.attention[i](q, k, v)[0])
+
+        outputs = torch.cat(outputs, 1)
+        src = loc.view(4, c, -1).permute(2, 0, 1) + self.dropout1(outputs)
+        src = self.norm1(src)
+        src = src + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(src)).clone())))
+        src = self.norm2(src)
+        src = src.permute(1, 2, 0).reshape(4, c, h, w)  # freshed loc
+        glb = glb + F.interpolate(patches2image(src), size=glb.shape[-2:], mode='nearest')  # freshed glb
+
+        return torch.cat((src, glb), 0), token_attention_map
+
+
+
+class BEN_Base(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.backbone = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+        emb_dim = 128
+        self.sideout5 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout4 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout3 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout2 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+        self.sideout1 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+
+        self.output5 = make_cbr(1024, emb_dim)
+        self.output4 = make_cbr(512, emb_dim)
+        self.output3 = make_cbr(256, emb_dim)
+        self.output2 = make_cbr(128, emb_dim)
+        self.output1 = make_cbr(128, emb_dim)
+
+        self.multifieldcrossatt = MCLM(emb_dim, 1, [1, 4, 8])
+        self.conv1 = make_cbr(emb_dim, emb_dim)
+        self.conv2 = make_cbr(emb_dim, emb_dim)
+        self.conv3 = make_cbr(emb_dim, emb_dim)
+        self.conv4 = make_cbr(emb_dim, emb_dim)
+        self.dec_blk1 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk2 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk3 = MCRM(emb_dim, 1, [2, 4, 8])
+        self.dec_blk4 = MCRM(emb_dim, 1, [2, 4, 8])
+
+        self.insmask_head = nn.Sequential(
+            nn.Conv2d(emb_dim, 384, kernel_size=3, padding=1),
+            nn.InstanceNorm2d(384),
+            nn.GELU(),
+            nn.Conv2d(384, 384, kernel_size=3, padding=1),
+            nn.InstanceNorm2d(384),
+            nn.GELU(),
+            nn.Conv2d(384, emb_dim, kernel_size=3, padding=1)
+        )
+
+        self.shallow = nn.Sequential(nn.Conv2d(3, emb_dim, kernel_size=3, padding=1))
+        self.upsample1 = make_cbg(emb_dim, emb_dim)
+        self.upsample2 = make_cbg(emb_dim, emb_dim)
+        self.output = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1))
+
+        for m in self.modules():
+            if isinstance(m, nn.GELU) or isinstance(m, nn.Dropout):
+                m.inplace = True
+
+    
+
+    @torch.inference_mode()
+    @torch.autocast(device_type="cuda",dtype=torch.float16)
+    def forward(self, x):
+        real_batch = x.size(0)
+
+        shallow_batch = self.shallow(x)
+        glb_batch = rescale_to(x, scale_factor=0.5, interpolation='bilinear')
+
+
+
+        final_input = None
+        for i in range(real_batch):
+            start = i * 4
+            end   = (i + 1) * 4
+            loc_batch = image2patches(x[i,:,:,:].unsqueeze(dim=0))
+            input_ = torch.cat((loc_batch, glb_batch[i,:,:,:].unsqueeze(dim=0)), dim=0)  
+            
+            
+            if final_input == None:
+                final_input= input_
+            else: final_input = torch.cat((final_input, input_), dim=0)
+
+        features = self.backbone(final_input)
+        outputs = []
+        
+        for i in range(real_batch):
+
+            start = i * 5
+            end   = (i + 1) * 5
+            
+            f4 = features[4][start:end, :, :, :]  # shape: [5, C, H, W]
+            f3 = features[3][start:end, :, :, :]
+            f2 = features[2][start:end, :, :, :]
+            f1 = features[1][start:end, :, :, :]
+            f0 = features[0][start:end, :, :, :]
+            e5 = self.output5(f4)
+            e4 = self.output4(f3)
+            e3 = self.output3(f2)
+            e2 = self.output2(f1)
+            e1 = self.output1(f0)
+            loc_e5, glb_e5 = e5.split([4, 1], dim=0)
+            e5 = self.multifieldcrossatt(loc_e5, glb_e5)  # (4,128,16,16)
+
+
+            e4, tokenattmap4 = self.dec_blk4(e4 + resize_as(e5, e4)) 
+            e4 = self.conv4(e4) 
+            e3, tokenattmap3 = self.dec_blk3(e3 + resize_as(e4, e3))
+            e3 = self.conv3(e3)
+            e2, tokenattmap2 = self.dec_blk2(e2 + resize_as(e3, e2))
+            e2 = self.conv2(e2)
+            e1, tokenattmap1 = self.dec_blk1(e1 + resize_as(e2, e1))
+            e1 = self.conv1(e1)
+
+            loc_e1, glb_e1 = e1.split([4, 1], dim=0)
+
+            output1_cat = patches2image(loc_e1)  # (1,128,256,256)
+
+            # add glb feat in
+            output1_cat = output1_cat + resize_as(glb_e1, output1_cat)
+            # merge
+            final_output = self.insmask_head(output1_cat)  # (1,128,256,256)
+            # shallow feature merge
+            shallow = shallow_batch[i,:,:,:].unsqueeze(dim=0)
+            final_output = final_output + resize_as(shallow, final_output)
+            final_output = self.upsample1(rescale_to(final_output))
+            final_output = rescale_to(final_output + resize_as(shallow, final_output))
+            final_output = self.upsample2(final_output)
+            final_output = self.output(final_output)
+            mask = final_output.sigmoid()
+            outputs.append(mask)
+
+        return torch.cat(outputs, dim=0)
+
+
+
+
+    def loadcheckpoints(self,model_path):
+        model_dict = torch.load(model_path, map_location="cpu", weights_only=True)
+        self.load_state_dict(model_dict['model_state_dict'], strict=True)
+        del model_path
+
+    def inference(self,image,refine_foreground=False):
+        
+        set_random_seed(9)
+        # image = ImageOps.exif_transpose(image)
+        if isinstance(image, Image.Image):            
+            image, h, w,original_image =  rgb_loader_refiner(image)
+            if torch.cuda.is_available():
+
+                img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device)
+            else:
+                img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device)
+
+            
+            with torch.no_grad():
+                res = self.forward(img_tensor)
+
+            # Show Results
+            if refine_foreground == True:
+
+                pred_pil = transforms.ToPILImage()(res.squeeze())
+                image_masked = refine_foreground_process(original_image, pred_pil)
+                
+                image_masked.putalpha(pred_pil.resize(original_image.size))
+                return image_masked
+
+            else:
+                alpha = postprocess_image(res, im_size=[w,h])
+                pred_pil = transforms.ToPILImage()(alpha)
+                mask = pred_pil.resize(original_image.size)
+                original_image.putalpha(mask)
+                # mask = Image.fromarray(alpha)
+
+                return original_image
+
+
+        else:
+            foregrounds = []
+            for batch in image:
+                image, h, w,original_image =  rgb_loader_refiner(batch)
+                if torch.cuda.is_available():
+
+                    img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device)
+                else:
+                    img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device)
+
+                with torch.no_grad():
+                    res = self.forward(img_tensor)
+
+                if refine_foreground == True:
+
+                    pred_pil = transforms.ToPILImage()(res.squeeze())
+                    image_masked = refine_foreground_process(original_image, pred_pil)
+                    
+                    image_masked.putalpha(pred_pil.resize(original_image.size))
+
+                    foregrounds.append(image_masked)
+                else:
+                    alpha = postprocess_image(res, im_size=[w,h])
+                    pred_pil = transforms.ToPILImage()(alpha)
+                    mask = pred_pil.resize(original_image.size)
+                    original_image.putalpha(mask)
+                    # mask = Image.fromarray(alpha)
+                    foregrounds.append(original_image)
+
+            return foregrounds
+
+
+
+
+    def segment_video(self, video_path, output_path="./", fps=0, refine_foreground=False, batch=1, print_frames_processed=True, webm = False, rgb_value= (0, 255, 0)):
+    
+        """
+        Segments the given video to extract the foreground (with alpha) from each frame
+        and saves the result as either a WebM video (with alpha channel) or MP4 (with a
+        color background).
+
+        Args:
+            video_path (str):
+                Path to the input video file.
+
+            output_path (str, optional):
+                Directory (or full path) where the output video and/or files will be saved.
+                Defaults to "./".
+
+            fps (int, optional):
+                The frames per second (FPS) to use for the output video. If 0 (default), the
+                original FPS of the input video is used. Otherwise, overrides it.
+
+            refine_foreground (bool, optional):
+                Whether to run an additional “refine foreground” process on each frame. 
+                Defaults to False.
+
+            batch (int, optional):
+                Number of frames to process at once (inference batch size). Large batch sizes
+                may require more GPU memory. Defaults to 1.
+
+            print_frames_processed (bool, optional):
+                If True (default), prints progress (how many frames have been processed) to 
+                the console.
+
+            webm (bool, optional):
+                If True (default), exports a WebM video with alpha channel (VP9 / yuva420p).
+                If False, exports an MP4 video composited over a solid color background.
+
+            rgb_value (tuple, optional):
+                The RGB background color (e.g., green screen) used to composite frames when
+                saving to MP4. Defaults to (0, 255, 0).
+
+        Returns:
+            None. Writes the output video(s) to disk in the specified format.
+        """
+
+        
+        cap = cv2.VideoCapture(video_path)
+        if not cap.isOpened():
+            raise IOError(f"Cannot open video: {video_path}")
+
+        original_fps = cap.get(cv2.CAP_PROP_FPS)
+        original_fps = 30 if original_fps == 0 else original_fps
+        fps = original_fps if fps == 0 else fps
+
+        ret, first_frame = cap.read()
+        if not ret:
+            raise ValueError("No frames found in the video.")
+        height, width = first_frame.shape[:2]
+        cap.set(cv2.CAP_PROP_POS_FRAMES, 0)
+
+        foregrounds = []
+        frame_idx = 0
+        processed_count = 0
+        batch_frames = []
+        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                if batch_frames:
+                    batch_results = self.inference(batch_frames, refine_foreground)
+                    if isinstance(batch_results, Image.Image):
+                        foregrounds.append(batch_results)
+                    else:
+                        foregrounds.extend(batch_results)
+                    if print_frames_processed:
+                        print(f"Processed frames {frame_idx-len(batch_frames)+1} to {frame_idx} of {total_frames}")
+                break
+
+            # Process every frame instead of using intervals
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            pil_frame = Image.fromarray(frame_rgb)
+            batch_frames.append(pil_frame)
+            
+            if len(batch_frames) == batch:
+                batch_results = self.inference(batch_frames, refine_foreground)
+                if isinstance(batch_results, Image.Image):
+                    foregrounds.append(batch_results)
+                else:
+                    foregrounds.extend(batch_results)
+                if print_frames_processed:
+                    print(f"Processed frames {frame_idx-batch+1} to {frame_idx} of {total_frames}")
+                batch_frames = []
+                processed_count += batch
+
+            frame_idx += 1
+
+
+        if webm:
+            alpha_webm_path = os.path.join(output_path, "foreground.webm")
+            pil_images_to_webm_alpha(foregrounds, alpha_webm_path, fps=original_fps)
+
+        else:
+            cap.release()
+            fg_output = os.path.join(output_path, 'foreground.mp4')
+            
+            pil_images_to_mp4(foregrounds, fg_output, fps=original_fps,rgb_value=rgb_value)
+            cv2.destroyAllWindows()
+            
+            try:
+                fg_audio_output = os.path.join(output_path, 'foreground_output_with_audio.mp4')
+                add_audio_to_video(fg_output, video_path, fg_audio_output)
+            except Exception as e:
+                print("No audio found in the original video")
+                print(e)
+
+
+
+
+
+def rgb_loader_refiner( original_image):
+        h, w = original_image.size
+
+        image = original_image
+        # Convert to RGB if necessary
+        if image.mode != 'RGB':
+            image = image.convert('RGB')
+
+        # Resize the image
+        image = image.resize((1024, 1024), resample=Image.LANCZOS)
+
+        return image.convert('RGB'), h, w,original_image
+
+# Define the image transformation
+img_transform = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.ConvertImageDtype(torch.float16), 
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+])
+
+img_transform32 = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.ConvertImageDtype(torch.float32), 
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+])
+
+
+
+
+
+def pil_images_to_mp4(images, output_path, fps=24, rgb_value=(0, 255, 0)):
+    """
+    Converts an array of PIL images to an MP4 video.
+    
+    Args:
+        images: List of PIL images
+        output_path: Path to save the MP4 file
+        fps: Frames per second (default: 24)
+        rgb_value: Background RGB color tuple (default: green (0, 255, 0))
+    """
+    if not images:
+        raise ValueError("No images provided to convert to MP4.")
+
+    width, height = images[0].size
+    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+    video_writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
+
+    for image in images:
+        # If image has alpha channel, composite onto the specified background color
+        if image.mode == 'RGBA':
+            # Create background image with specified RGB color
+            background = Image.new('RGB', image.size, rgb_value)
+            background = background.convert('RGBA')
+            # Composite the image onto the background
+            image = Image.alpha_composite(background, image)
+            image = image.convert('RGB')
+        else:
+            # Ensure RGB format for non-alpha images
+            image = image.convert('RGB')
+
+        # Convert to OpenCV format and write
+        open_cv_image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
+        video_writer.write(open_cv_image)
+    
+    video_writer.release()
+
+def pil_images_to_webm_alpha(images, output_path, fps=30):
+    """
+    Converts a list of PIL RGBA images to a VP9 .webm video with alpha channel.
+
+    NOTE: Not all players will display alpha in WebM. 
+          Browsers like Chrome/Firefox typically do support VP9 alpha.
+    """
+    if not images:
+        raise ValueError("No images provided for WebM with alpha.")
+
+    # Ensure output directory exists
+    os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        # Save frames as PNG (with alpha)
+        for idx, img in enumerate(images):
+            if img.mode != "RGBA":
+                img = img.convert("RGBA")
+            out_path = os.path.join(tmpdir, f"{idx:06d}.png")
+            img.save(out_path, "PNG")
+
+        # Construct ffmpeg command
+        # -c:v libvpx-vp9 => VP9 encoder
+        # -pix_fmt yuva420p => alpha-enabled pixel format
+        # -auto-alt-ref 0 => helps preserve alpha frames (libvpx quirk)
+        ffmpeg_cmd = [
+            "ffmpeg", "-y",
+            "-framerate", str(fps),
+            "-i", os.path.join(tmpdir, "%06d.png"),
+            "-c:v", "libvpx-vp9",
+            "-pix_fmt", "yuva420p",
+            "-auto-alt-ref", "0",
+            output_path
+        ]
+
+        subprocess.run(ffmpeg_cmd, check=True)
+
+    print(f"WebM with alpha saved to {output_path}")
+
+def add_audio_to_video(video_without_audio_path, original_video_path, output_path):
+    """
+    Check if the original video has an audio stream. If yes, add it. If not, skip.
+    """
+    # 1) Probe original video for audio streams
+    probe_command = [
+        'ffprobe', '-v', 'error', 
+        '-select_streams', 'a:0', 
+        '-show_entries', 'stream=index', 
+        '-of', 'csv=p=0', 
+        original_video_path
+    ]
+    result = subprocess.run(probe_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+
+    # result.stdout is empty if no audio stream found
+    if not result.stdout.strip():
+        print("No audio track found in original video, skipping audio addition.")
+        return
+    
+    print("Audio track detected; proceeding to mux audio.")
+    # 2) If audio found, run ffmpeg to add it
+    command = [
+        'ffmpeg', '-y',
+        '-i', video_without_audio_path,
+        '-i', original_video_path,
+        '-c', 'copy',
+        '-map', '0:v:0',
+        '-map', '1:a:0',  # we know there's an audio track now
+        output_path
+    ]
+    subprocess.run(command, check=True)
+    print(f"Audio added successfully => {output_path}")
+
+
+
+
+
+### Thanks to the source: https://huggingface.co/ZhengPeng7/BiRefNet/blob/main/handler.py
+def refine_foreground_process(image, mask, r=90):
+    if mask.size != image.size:
+        mask = mask.resize(image.size)
+    image = np.array(image) / 255.0
+    mask = np.array(mask) / 255.0
+    estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r)
+    image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
+    return image_masked
+
+
+def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90):
+    # Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
+    alpha = alpha[:, :, None]
+    F, blur_B = FB_blur_fusion_foreground_estimator(image, image, image, alpha, r)
+    return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0]
+
+
+def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
+    if isinstance(image, Image.Image):
+        image = np.array(image) / 255.0
+    blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
+
+    blurred_FA = cv2.blur(F * alpha, (r, r))
+    blurred_F = blurred_FA / (blurred_alpha + 1e-5)
+
+    blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
+    blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
+    F = blurred_F + alpha * \
+        (image - alpha * blurred_F - (1 - alpha) * blurred_B)
+    F = np.clip(F, 0, 1)
+    return F, blurred_B
+
+    
+
+def postprocess_image(result: torch.Tensor, im_size: list) -> np.ndarray:
+    result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear'), 0)
+    ma = torch.max(result)
+    mi = torch.min(result)
+    result = (result - mi) / (ma - mi)
+    im_array = (result * 255).permute(1, 2, 0).cpu().data.numpy().astype(np.uint8)
+    im_array = np.squeeze(im_array)
+    return im_array
+
+
+
+
+def rgb_loader_refiner( original_image):
+        h, w = original_image.size
+        # # Apply EXIF orientation
+
+        image = ImageOps.exif_transpose(original_image)
+
+        if original_image.mode != 'RGB':
+            original_image = original_image.convert('RGB')
+
+        image = original_image
+        # Convert to RGB if necessary
+
+        # Resize the image
+        image = image.resize((1024, 1024), resample=Image.LANCZOS)
+
+        return image, h, w,original_image
+
+
+

models/RMBG/BEN2/BEN2_Base.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:926144a876bda06f125555b4f5a239ece89dc6eb838a863700ca9bf192161a1c
+size 1134584206

models/RMBG/BEN2/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/BiRefNet-HR/BiRefNet_config.py

@@ -0,0 +1,11 @@
+from transformers import PretrainedConfig
+
+class BiRefNetConfig(PretrainedConfig):
+    model_type = "SegformerForSemanticSegmentation"
+    def __init__(
+        self,
+        bb_pretrained=False,
+        **kwargs
+    ):
+        self.bb_pretrained = bb_pretrained
+        super().__init__(**kwargs)

models/RMBG/BiRefNet-HR/birefnet.py

@@ -0,0 +1,2248 @@
+### config.py
+
+import os
+import math
+
+
+class Config():
+    def __init__(self) -> None:
+        # PATH settings
+        self.sys_home_dir = os.path.expanduser('~')     # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
+            'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD',     # leave DIS-VD for evaluation.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4)     # DIS needs high lr to converge faster. Adapt the lr linearly
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][6]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            'structure': 5 * 0,    # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
+        # with open(run_sh_file[0], 'r') as f:
+        #     lines = f.readlines()
+        #     self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
+        #     self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
+        # self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+
+
+### models/backbones/pvt_v2.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+# from config import Config
+
+# config = Config()
+
+class Mlp(nn.Module):
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * 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 Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+## @register_model
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+## @register_model
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+## @register_model
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+## @register_model
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+### models/backbones/swin_v1.py
+
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+# from config import Config
+
+
+# config = Config()
+
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            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], -1
+            )   # Wh*Ww, Wh*Ww, nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        # Turn int to torch.tensor for the compatiability with torch.compile in PyTorch 2.5.
+        Hp = torch.ceil(torch.tensor(H) / self.window_size).to(torch.int64) * self.window_size
+        Wp = torch.ceil(torch.tensor(W) / self.window_size).to(torch.int64) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)).to(x.dtype)
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+            
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model
+
+
+
+### models/modules/deform_conv.py
+
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+        
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+        
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+        
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x
+
+
+
+
+### utils.py
+
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)
+
+
+
+### dataset.py
+
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+### models/backbones/build_backbones.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+# from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+# from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+# from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model
+
+
+
+### models/modules/decoder_blocks.py
+
+import torch
+import torch.nn as nn
+# from models.aspp import ASPP, ASPPDeformable
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+
+### models/modules/lateral_blocks.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+
+### models/modules/aspp.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.deform_conv import DeformableConv2d
+# from config import Config
+
+
+# config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+
+### models/refinement/refiner.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.ing import *
+# from models.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs
+
+
+
+### models/stem_layer.py
+
+import torch.nn as nn
+# from utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x
+
+
+### models/birefnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from transformers import PreTrainedModel
+from einops import rearrange
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.aspp import ASPP, ASPPDeformable
+# from models.ing import *
+# from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+# from models.stem_layer import StemLayer
+from .BiRefNet_config import BiRefNetConfig
+
+
+def image2patches(image, grid_h=2, grid_w=2, patch_ref=None, transformation='b c (hg h) (wg w) -> (b hg wg) c h w'):
+    if patch_ref is not None:
+        grid_h, grid_w = image.shape[-2] // patch_ref.shape[-2], image.shape[-1] // patch_ref.shape[-1]
+    patches = rearrange(image, transformation, hg=grid_h, wg=grid_w)
+    return patches
+
+def patches2image(patches, grid_h=2, grid_w=2, patch_ref=None, transformation='(b hg wg) c h w -> b c (hg h) (wg w)'):
+    if patch_ref is not None:
+        grid_h, grid_w = patch_ref.shape[-2] // patches[0].shape[-2], patch_ref.shape[-1] // patches[0].shape[-1]
+    image = rearrange(patches, transformation, hg=grid_h, wg=grid_w)
+    return image
+
+class BiRefNet(
+    PreTrainedModel
+):
+    config_class = BiRefNetConfig
+    def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
+        super(BiRefNet, self).__init__(config)
+        bb_pretrained = config.bb_pretrained
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=x4, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p3, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p2, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision and self.training:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

models/RMBG/BiRefNet-HR/config.json

@@ -0,0 +1,20 @@
+{
+  "_name_or_path": "ZhengPeng7/BiRefNet_HR",
+  "architectures": [
+    "BiRefNet"
+  ],
+  "auto_map": {
+    "AutoConfig": "BiRefNet_config.BiRefNetConfig",
+    "AutoModelForImageSegmentation": "birefnet.BiRefNet"
+  },
+  "custom_pipelines": {
+    "image-segmentation": {
+      "pt": [
+        "AutoModelForImageSegmentation"
+      ],
+      "tf": [],
+      "type": "image"
+    }
+  },
+  "bb_pretrained": false
+}

models/RMBG/BiRefNet-HR/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/BiRefNet-HR/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d678bafec0b0019fbb073b7fd02f05ede25dc4b15254f23b2fb0be333200c0d
+size 444473596

models/RMBG/BiRefNet/BiRefNet-HR-matting.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5a4de698739ea5e0e8bbab28e1b293dde95092b87a442d566cbc585c53cef55
+size 444473596

models/RMBG/BiRefNet/BiRefNet-HR.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d678bafec0b0019fbb073b7fd02f05ede25dc4b15254f23b2fb0be333200c0d
+size 444473596

models/RMBG/BiRefNet/BiRefNet-general.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77277264c0e8c74149d3ff2fade4fd8176965b7108f3c5fc3b8c9c811edb4519
+size 884878856

models/RMBG/BiRefNet/BiRefNet-matting.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9875de5b1e6c8eb5fdaa8c727a82927ce442cdc87ba3abee6a77e6fa46c25bb
+size 884878856

models/RMBG/BiRefNet/BiRefNet-portrait.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a4eb3a5469b75f0cccaec6772c22fc30e6c12ed429c2c0ba71b43e3d8d97182
+size 884878856

models/RMBG/BiRefNet/BiRefNet_512x512.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d94ae0eefb2d2020192001e984ecd6b367478118257a3132e6a484bbf18b0f41
+size 444473596

models/RMBG/BiRefNet/BiRefNet_config.py

@@ -0,0 +1,11 @@
+from transformers import PretrainedConfig
+
+class BiRefNetConfig(PretrainedConfig):
+    model_type = "SegformerForSemanticSegmentation"
+    def __init__(
+        self,
+        bb_pretrained=False,
+        **kwargs
+    ):
+        self.bb_pretrained = bb_pretrained
+        super().__init__(**kwargs)

models/RMBG/BiRefNet/BiRefNet_lite-2K.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa2e4a5af5eb3904694feb40f2b39ec5dd7cd9110906590cfeb982f09a46021d
+size 177634392

models/RMBG/BiRefNet/BiRefNet_lite.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4417d89795250e698c3cb0ae8df15743810065f646f48a694fdfa7ca052d0815
+size 177634392

models/RMBG/BiRefNet/birefnet.py

@@ -0,0 +1,2248 @@
+### config.py
+
+import os
+import math
+
+
+class Config():
+    def __init__(self) -> None:
+        # PATH settings
+        self.sys_home_dir = os.path.expanduser('~')     # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
+            'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD',     # leave DIS-VD for evaluation.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4)     # DIS needs high lr to converge faster. Adapt the lr linearly
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][6]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            'structure': 5 * 0,    # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
+        # with open(run_sh_file[0], 'r') as f:
+        #     lines = f.readlines()
+        #     self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
+        #     self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
+        # self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+
+
+### models/backbones/pvt_v2.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+# from config import Config
+
+# config = Config()
+
+class Mlp(nn.Module):
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * 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 Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+## @register_model
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+## @register_model
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+## @register_model
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+## @register_model
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+### models/backbones/swin_v1.py
+
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+# from config import Config
+
+
+# config = Config()
+
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            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], -1
+            )   # Wh*Ww, Wh*Ww, nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        # Turn int to torch.tensor for the compatiability with torch.compile in PyTorch 2.5.
+        Hp = torch.ceil(torch.tensor(H) / self.window_size).to(torch.int64) * self.window_size
+        Wp = torch.ceil(torch.tensor(W) / self.window_size).to(torch.int64) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)).to(x.dtype)
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+            
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model
+
+
+
+### models/modules/deform_conv.py
+
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+        
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+        
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+        
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x
+
+
+
+
+### utils.py
+
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)
+
+
+
+### dataset.py
+
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+### models/backbones/build_backbones.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+# from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+# from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+# from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model
+
+
+
+### models/modules/decoder_blocks.py
+
+import torch
+import torch.nn as nn
+# from models.aspp import ASPP, ASPPDeformable
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+
+### models/modules/lateral_blocks.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+
+### models/modules/aspp.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.deform_conv import DeformableConv2d
+# from config import Config
+
+
+# config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+
+### models/refinement/refiner.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.ing import *
+# from models.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs
+
+
+
+### models/stem_layer.py
+
+import torch.nn as nn
+# from utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x
+
+
+### models/birefnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from transformers import PreTrainedModel
+from einops import rearrange
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.aspp import ASPP, ASPPDeformable
+# from models.ing import *
+# from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+# from models.stem_layer import StemLayer
+from BiRefNet_config import BiRefNetConfig
+
+
+def image2patches(image, grid_h=2, grid_w=2, patch_ref=None, transformation='b c (hg h) (wg w) -> (b hg wg) c h w'):
+    if patch_ref is not None:
+        grid_h, grid_w = image.shape[-2] // patch_ref.shape[-2], image.shape[-1] // patch_ref.shape[-1]
+    patches = rearrange(image, transformation, hg=grid_h, wg=grid_w)
+    return patches
+
+def patches2image(patches, grid_h=2, grid_w=2, patch_ref=None, transformation='(b hg wg) c h w -> b c (hg h) (wg w)'):
+    if patch_ref is not None:
+        grid_h, grid_w = patch_ref.shape[-2] // patches[0].shape[-2], patch_ref.shape[-1] // patches[0].shape[-1]
+    image = rearrange(patches, transformation, hg=grid_h, wg=grid_w)
+    return image
+
+class BiRefNet(
+    PreTrainedModel
+):
+    config_class = BiRefNetConfig
+    def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
+        super(BiRefNet, self).__init__(config)
+        bb_pretrained = config.bb_pretrained
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=x4, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p3, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p2, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision and self.training:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

models/RMBG/BiRefNet/birefnet_lite.py

@@ -0,0 +1,2248 @@
+### config.py
+
+import os
+import math
+
+
+class Config():
+    def __init__(self) -> None:
+        # PATH settings
+        self.sys_home_dir = os.path.expanduser('~')     # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
+            'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD',     # leave DIS-VD for evaluation.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4)     # DIS needs high lr to converge faster. Adapt the lr linearly
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][3]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            'structure': 5 * 0,    # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
+        # with open(run_sh_file[0], 'r') as f:
+        #     lines = f.readlines()
+        #     self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
+        #     self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
+        # self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+
+
+### models/backbones/pvt_v2.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+# from config import Config
+
+# config = Config()
+
+class Mlp(nn.Module):
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * 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 Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+## @register_model
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+## @register_model
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+## @register_model
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+## @register_model
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+### models/backbones/swin_v1.py
+
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+# from config import Config
+
+
+# config = Config()
+
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            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], -1
+            )   # Wh*Ww, Wh*Ww, nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        # Turn int to torch.tensor for the compatiability with torch.compile in PyTorch 2.5.
+        Hp = torch.ceil(torch.tensor(H) / self.window_size).to(torch.int64) * self.window_size
+        Wp = torch.ceil(torch.tensor(W) / self.window_size).to(torch.int64) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)).to(x.dtype)
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+            
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model
+
+
+
+### models/modules/deform_conv.py
+
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+        
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+        
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+        
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x
+
+
+
+
+### utils.py
+
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)
+
+
+
+### dataset.py
+
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+### models/backbones/build_backbones.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+# from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+# from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+# from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model
+
+
+
+### models/modules/decoder_blocks.py
+
+import torch
+import torch.nn as nn
+# from models.aspp import ASPP, ASPPDeformable
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+
+### models/modules/lateral_blocks.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+
+### models/modules/aspp.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.deform_conv import DeformableConv2d
+# from config import Config
+
+
+# config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+
+### models/refinement/refiner.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.ing import *
+# from models.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs
+
+
+
+### models/stem_layer.py
+
+import torch.nn as nn
+# from utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x
+
+
+### models/birefnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from transformers import PreTrainedModel
+from einops import rearrange
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.aspp import ASPP, ASPPDeformable
+# from models.ing import *
+# from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+# from models.stem_layer import StemLayer
+from BiRefNet_config import BiRefNetConfig
+
+
+def image2patches(image, grid_h=2, grid_w=2, patch_ref=None, transformation='b c (hg h) (wg w) -> (b hg wg) c h w'):
+    if patch_ref is not None:
+        grid_h, grid_w = image.shape[-2] // patch_ref.shape[-2], image.shape[-1] // patch_ref.shape[-1]
+    patches = rearrange(image, transformation, hg=grid_h, wg=grid_w)
+    return patches
+
+def patches2image(patches, grid_h=2, grid_w=2, patch_ref=None, transformation='(b hg wg) c h w -> b c (hg h) (wg w)'):
+    if patch_ref is not None:
+        grid_h, grid_w = patch_ref.shape[-2] // patches[0].shape[-2], patch_ref.shape[-1] // patches[0].shape[-1]
+    image = rearrange(patches, transformation, hg=grid_h, wg=grid_w)
+    return image
+
+class BiRefNet(
+    PreTrainedModel
+):
+    config_class = BiRefNetConfig
+    def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
+        super(BiRefNet, self).__init__(config)
+        bb_pretrained = config.bb_pretrained
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=x4, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p3, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p2, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision and self.training:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

models/RMBG/BiRefNet/config.json

@@ -0,0 +1,20 @@
+{
+  "_name_or_path": "1038lab/BiRefNet",
+  "architectures": [
+    "BiRefNet"
+  ],
+  "auto_map": {
+    "AutoConfig": "BiRefNet_config.BiRefNetConfig",
+    "AutoModelForImageSegmentation": "birefnet.BiRefNet"
+  },
+  "custom_pipelines": {
+    "image-segmentation": {
+      "pt": [
+        "AutoModelForImageSegmentation"
+      ],
+      "tf": [],
+      "type": "image"
+    }
+  },
+  "bb_pretrained": false
+}

models/RMBG/BiRefNet/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/INSPYRENET/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/INSPYRENET/inspyrenet.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a6fe2a73ab0532d6d0b8d82849a9760a226df719e3063d09b4149ece6f80fcd
+size 367520613

models/RMBG/INSPYRENET/inspyrenet.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c8ab4541f9a960b376a80362b07f1d96262e689c501365e6323faa958de47ea
+size 367120444

models/RMBG/RMBG-2.0/BiRefNet_config.py

@@ -0,0 +1,11 @@
+from transformers import PretrainedConfig
+
+class BiRefNetConfig(PretrainedConfig):
+    model_type = "SegformerForSemanticSegmentation"
+    def __init__(
+        self,
+        bb_pretrained=False,
+        **kwargs
+    ):
+        self.bb_pretrained = bb_pretrained
+        super().__init__(**kwargs)

models/RMBG/RMBG-2.0/birefnet.py

@@ -0,0 +1,2244 @@
+### config.py
+
+import os
+import math
+
+
+class Config():
+    def __init__(self) -> None:
+        # PATH settings
+        self.sys_home_dir = os.path.expanduser('~')    # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
+            'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD',     # leave DIS-VD for evaluation.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4)     # DIS needs high lr to converge faster. Adapt the lr linearly
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][6]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            'structure': 5 * 0,    # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
+        # with open(run_sh_file[0], 'r') as f:
+        #     lines = f.readlines()
+        #     self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
+        #     self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
+        # self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+
+
+### models/backbones/pvt_v2.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+# from config import Config
+
+# config = Config()
+
+class Mlp(nn.Module):
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * 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 Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+## @register_model
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+## @register_model
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+## @register_model
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+## @register_model
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+### models/backbones/swin_v1.py
+
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+# from config import Config
+
+
+# config = Config()
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    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
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        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], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            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], -1)  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            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 SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        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
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                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(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=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_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+            
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model
+
+
+
+### models/modules/deform_conv.py
+
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+        
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+        
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+        
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x
+
+
+
+
+### utils.py
+
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)
+
+
+
+### dataset.py
+
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+### models/backbones/build_backbones.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+# from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+# from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+# from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model
+
+
+
+### models/modules/decoder_blocks.py
+
+import torch
+import torch.nn as nn
+# from models.aspp import ASPP, ASPPDeformable
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+
+### models/modules/lateral_blocks.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+
+### models/modules/aspp.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.deform_conv import DeformableConv2d
+# from config import Config
+
+
+# config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+
+### models/refinement/refiner.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.ing import *
+# from models.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs
+
+
+
+### models/stem_layer.py
+
+import torch.nn as nn
+# from utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x
+
+
+### models/birefnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from transformers import PreTrainedModel
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.aspp import ASPP, ASPPDeformable
+# from models.ing import *
+# from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+# from models.stem_layer import StemLayer
+from .BiRefNet_config import BiRefNetConfig
+
+
+class BiRefNet(
+    PreTrainedModel
+):
+    config_class = BiRefNetConfig
+    def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
+        super(BiRefNet, self).__init__(config)
+        bb_pretrained = config.bb_pretrained
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def get_patches_batch(self, x, p):
+        _size_h, _size_w = p.shape[2:]
+        patches_batch = []
+        for idx in range(x.shape[0]):
+            columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1)
+            patches_x = []
+            for column_x in columns_x:
+                patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)]
+            patch_sample = torch.cat(patches_x, dim=1)
+            patches_batch.append(patch_sample)
+        return torch.cat(patches_batch, dim=0)
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, x4) if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p3) if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p2) if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

models/RMBG/RMBG-2.0/config.json

@@ -0,0 +1,20 @@
+{
+  "_name_or_path": "ZhengPeng7/BiRefNet",
+  "architectures": [
+    "BiRefNet"
+  ],
+  "auto_map": {
+    "AutoConfig": "BiRefNet_config.BiRefNetConfig",
+    "AutoModelForImageSegmentation": "birefnet.BiRefNet"
+  },
+  "custom_pipelines": {
+    "image-segmentation": {
+      "pt": [
+        "AutoModelForImageSegmentation"
+      ],
+      "tf": [],
+      "type": "image"
+    }
+  },
+  "bb_pretrained": false
+}

models/RMBG/RMBG-2.0/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/RMBG-2.0/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:566ed80c3d95f87ada6864d4cbe2290a1c5eb1c7bb0b123e984f60f76b02c3a7
+size 884878856

models/RMBG/SAM/Moonlit Serenade.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a219a66c5d8590a19a4dc7f8bc661170fd6afeab9db8e642c474bb3571ede4a
+size 4812946

models/RMBG/SAM/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/SAM/mobile_sam.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dbb90523a35330fedd7f1d3dfc66f995213d81b29a5ca8108dbcdd4e37d6c2f
+size 40728226

models/RMBG/SAM/mobile_sam.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:158f98ffe079d9a5f0c6ef6ed8a238ce2a61220754fa972d93156200d693c1bf
+size 40613136

models/RMBG/SAM/sam_hq_vit_b.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14a9d662cd6f5a9c2dba6d40ab0058d88d287e4a18fd6fdc6ad5fb1a3fdeaa57
+size 379335069

models/RMBG/SAM/sam_hq_vit_h.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7ac14a085326d9fa6199c8c698c4f0e7280afdbb974d2c4660ec60877b45e35
+size 2570940653

models/RMBG/SAM/sam_hq_vit_l.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1a6c385d62bf005ded91a54d5ec55c985cfc4103ef89c08d90f39f04934c343
+size 1254865805

models/RMBG/SAM/sam_vit_b.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec2df62732614e57411cdcf32a23ffdf28910380d03139ee0f4fcbe91eb8c912
+size 375042383

models/RMBG/SAM/sam_vit_h.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7bf3b02f3ebf1267aba913ff637d9a2d5c33d3173bb679e46d9f338c26f262e
+size 2564550879

models/RMBG/SAM/sam_vit_l.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3adcc4315b642a4d2101128f611684e8734c41232a17c648ed1693702a49a622
+size 1249524607

models/RMBG/grounding-dino/GroundingDINO_SwinB.cfg.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_B_384_22k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

models/RMBG/grounding-dino/GroundingDINO_SwinT_OGC.cfg.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_T_224_1k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

models/RMBG/grounding-dino/gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

models/RMBG/grounding-dino/groundingdino_swinb_cogcoor.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46270f7a822e6906b655b729c90613e48929d0f2bb8b9b76fd10a856f3ac6ab7
+size 938057991

models/RMBG/grounding-dino/groundingdino_swint_ogc.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b3ca2563c77c69f651d7bd133e97139c186df06231157a64c507099c52bc799
+size 693997677