training_code/MVANet/README.org

* Requirements
#+begin_src conf :tangle ./requirements.txt
  einops
  pillow
  prodigyopt
  tensorboard
  timm
  torch
  torchvision
#+end_src

* Download trained model
#+begin_src sh :shebang #!/bin/sh :results output :tangle ./download.sh
  "efficient_download.sh" \
      'https://huggingface.co/aravindhv10/Self-Correction-Human-Parsing/resolve/main/checkpoints/Model_80.pth' \
      'Model_80.pth' \
      '6ca28df33ba8476ac13be329a1b1b8b390da5d8042638fb124df3c067c2fe45bccde4366643b830066cbe0164ddbb978a1987a398b4a987f99d908903b44774f' \
      "${HOME}/GITHUB/aravind-h-v/dreambooth_experiments/cloth_segmentation/MVANet_Train/pretrained_model/Model_80.pth" \
  ;
#+end_src

* Swin code

** swin.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.import.py
  import  os
  os.environ["CUDA_VISIBLE_DEVICES"] ='0'
#+end_src

** swin.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.import.py
  import numpy as np
#+end_src

** swin.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.import.py
  import torch
  import torch.nn as nn
  import torch.nn.functional as F
  import torch.utils.checkpoint as checkpoint
#+end_src

** swin.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.import.py
  from timm.models import load_checkpoint
  from timm.models.layers import DropPath
  from timm.models.layers import to_2tuple
  from timm.models.layers import trunc_normal_

  # from mmdet.utils import get_root_logger
#+end_src

** swin.function.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.function.py
  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


  def SwinT(pretrained=True):
      model = SwinTransformer(embed_dim=96,
                              depths=[2, 2, 6, 2],
                              num_heads=[3, 6, 12, 24],
                              window_size=7)
      # if pretrained is True:
      #     model.load_state_dict(torch.load(
      #         'data/backbone_ckpt/swin_tiny_patch4_window7_224.pth',
      #         map_location='cpu')['model'],
      #                           strict=False)

      return model


  def SwinS(pretrained=True):
      model = SwinTransformer(embed_dim=96,
                              depths=[2, 2, 18, 2],
                              num_heads=[3, 6, 12, 24],
                              window_size=7)
      # if pretrained is True:
      #     model.load_state_dict(torch.load(
      #         'data/backbone_ckpt/swin_small_patch4_window7_224.pth',
      #         map_location='cpu')['model'],
      #                           strict=False)

      return model


  def SwinB(pretrained=True):
      model = SwinTransformer(embed_dim=128,
                              depths=[2, 2, 18, 2],
                              num_heads=[4, 8, 16, 32],
                              window_size=12)
      # if pretrained is True:
      #     model.load_state_dict(
      #         torch.load('./swin_base_patch4_window12_384_22kto1k.pth',
      #                    map_location='cpu')['model'],
      #         strict=False)

      return model


  def SwinL(pretrained=True):
      model = SwinTransformer(embed_dim=192,
                              depths=[2, 2, 18, 2],
                              num_heads=[6, 12, 24, 48],
                              window_size=12)
      # if pretrained is True:
      #     model.load_state_dict(torch.load(
      #         'data/backbone_ckpt/swin_large_patch4_window12_384_22kto1k.pth',
      #         map_location='cpu')['model'],
      #                           strict=False)

      return model
#+end_src

** swin.class.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.class.py
  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


  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 init_weights(self, pretrained=None):
          """Initialize the weights in backbone.

          Args:
              pretrained (str, optional): Path to pre-trained weights.
                  Defaults to None.
          """

          def _init_weights(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)

          if isinstance(pretrained, str):
              self.apply(_init_weights)
              # logger = get_root_logger()
              load_checkpoint(self, pretrained, strict=False, logger=None)
          elif pretrained is None:
              self.apply(_init_weights)
          else:
              raise TypeError('pretrained must be a str or None')

      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 train(self, mode=True):
          """Convert the model into training mode while keep layers freezed."""
          super(SwinTransformer, self).train(mode)
          self._freeze_stages()
#+end_src

* Main code

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  import os

  os.environ["CUDA_VISIBLE_DEVICES"] = '0'
  HOME_DIR = os.environ.get('HOME', '/root')
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  import sys

  sys.path.append(os.path.dirname(os.path.abspath(__file__)))
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  from datetime import datetime
  import argparse
  import numpy as np
  import random
  import math
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  import cv2
  from PIL import Image
  from PIL import ImageEnhance
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  from einops import rearrange
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  import torch
  import torch.nn as nn
  import torch.nn.functional as F
  import torch.utils.data as data

  from torch.autograd import Variable
  from torch.backends import cudnn
  from torch.cuda import amp
  from torch.utils.tensorboard import SummaryWriter

  from torchvision import transforms
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  from prodigyopt import Prodigy
#+end_src

** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
  # from model.MVANet import MVANet
  from swin import SwinB
#+end_src

** train.function.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.function.py
  def get_activation_fn(activation):
      """Return an activation function given a string"""
      if activation == "relu":
          return F.relu
      if activation == "gelu":
          return F.gelu
      if activation == "glu":
          return F.glu
      raise RuntimeError(F"activation should be relu/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.BatchNorm2d(out_dim), nn.PReLU())


  def make_cbg(in_dim, out_dim):
      return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1),
                           nn.BatchNorm2d(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


  def structure_loss(pred, mask):
      weit = 1 + 5 * torch.abs(
          F.avg_pool2d(mask, kernel_size=31, stride=1, padding=15) - mask)
      wbce = F.binary_cross_entropy_with_logits(pred, mask, reduction='none')
      wbce = (weit * wbce).sum(dim=(2, 3)) / weit.sum(dim=(2, 3))

      pred = torch.sigmoid(pred)
      inter = ((pred * mask) * weit).sum(dim=(2, 3))

      union = ((pred + mask) * weit).sum(dim=(2, 3))
      wiou = 1 - (inter + 1) / (union - inter + 1)

      return (wbce + wiou).mean()


  def clip_gradient(optimizer, grad_clip):
      for group in optimizer.param_groups:
          for param in group['params']:
              if param.grad is not None:
                  param.grad.data.clamp_(-grad_clip, grad_clip)


  def adjust_lr(optimizer, init_lr, epoch, decay_rate=0.1, decay_epoch=5):
      decay = decay_rate**(epoch // decay_epoch)
      for param_group in optimizer.param_groups:
          param_group['lr'] *= decay


  def truncated_normal_(tensor, mean=0, std=1):
      size = tensor.shape
      tmp = tensor.new_empty(size + (4, )).normal_()
      valid = (tmp < 2) & (tmp > -2)
      ind = valid.max(-1, keepdim=True)[1]
      tensor.data.copy_(tmp.gather(-1, ind).squeeze(-1))
      tensor.data.mul_(std).add_(mean)


  def init_weights(m):
      if type(m) == nn.Conv2d or type(m) == nn.ConvTranspose2d:
          nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='relu')
          #nn.init.normal_(m.weight, std=0.001)
          #nn.init.normal_(m.bias, std=0.001)
          truncated_normal_(m.bias, mean=0, std=0.001)


  def init_weights_orthogonal_normal(m):
      if type(m) == nn.Conv2d or type(m) == nn.ConvTranspose2d:
          nn.init.orthogonal_(m.weight)
          truncated_normal_(m.bias, mean=0, std=0.001)
          #nn.init.normal_(m.bias, std=0.001)


  def l2_regularisation(m):
      l2_reg = None

      for W in m.parameters():
          if l2_reg is None:
              l2_reg = W.norm(2)
          else:
              l2_reg = l2_reg + W.norm(2)
      return l2_reg


  def check_mkdir(dir_name):
      if not os.path.isdir(dir_name):
          os.makedirs(dir_name)


  # several data augumentation strategies
  def cv_random_flip(img, label):
      flip_flag = random.randint(0, 1)
      flip_flag2 = random.randint(0, 1)

      # left right flip
      if flip_flag == 1:
          img = img.transpose(Image.FLIP_LEFT_RIGHT)
          label = label.transpose(Image.FLIP_LEFT_RIGHT)

      # top bottom flip
      if flip_flag2 == 1:
          img = img.transpose(Image.FLIP_TOP_BOTTOM)
          label = label.transpose(Image.FLIP_TOP_BOTTOM)

      return img, label


  def random_crop_full(image, X, Y, TX, TY):
      image_width = image.size[0]
      image_height = image.size[1]
      final_width = image_width * TX
      final_height = image_height * TY

      start_x = (1.0 - TX) * X * image_width
      start_y = (1.0 - TY) * Y * image_height

      random_region = (start_x, start_y, start_x + final_width,
                       start_y + final_height)

      return image.crop(random_region)


  def random_crop(image, X, Y, T):
      image_width = image.size[0]
      image_height = image.size[1]
      final_width = image_width * T
      final_height = image_height * T

      start_x = (1.0 - T) * X * image_width
      start_y = (1.0 - T) * Y * image_height

      random_region = (start_x, start_y, start_x + final_width,
                       start_y + final_height)

      return image.crop(random_region)


  def garment_color_jitter(image, mask):
      image = np.array(image)
      mask = np.array(mask)
      mask = (mask > 127).astype(dtype=np.uint8)
      image = cv2.cvtColor(src=image, code=cv2.COLOR_RGB2HSV_FULL)
      image[:, :, 0] += mask * np.random.randint(0, 255)
      image = cv2.cvtColor(src=image, code=cv2.COLOR_HSV2RGB_FULL)
      image = Image.fromarray(image)
      return image


  def garment_color_jitter_rotate(image, mask, rotate_index=0, shift_amount=0):
      image = np.array(image)
      mask = np.array(mask)

      if rotate_index == 1:

          image = cv2.rotate(src=image, rotateCode=cv2.ROTATE_90_CLOCKWISE)
          mask = cv2.rotate(src=mask, rotateCode=cv2.ROTATE_90_CLOCKWISE)

      elif rotate_index == 2:

          image = cv2.rotate(src=image, rotateCode=cv2.ROTATE_180)
          mask = cv2.rotate(src=mask, rotateCode=cv2.ROTATE_180)

      elif rotate_index == 3:

          image = cv2.rotate(src=image,
                             rotateCode=cv2.ROTATE_90_COUNTERCLOCKWISE)

          mask = cv2.rotate(src=mask, rotateCode=cv2.ROTATE_90_COUNTERCLOCKWISE)

      image = cv2.cvtColor(src=image,
                           code=cv2.COLOR_RGB2HSV_FULL).astype(dtype=np.int32)
      # image[:, :, 0] += mask_tmp * shift_amount
      image[:, :, 0] += shift_amount
      image[:, :, 0] %= 255
      image = cv2.cvtColor(src=image.astype(np.uint8),
                           code=cv2.COLOR_HSV2RGB_FULL)

      image = Image.fromarray(image)
      mask = Image.fromarray(mask)

      return image, mask


  def randomCrop_Both(image, label):

      image, label = garment_color_jitter_rotate(
          image=image,
          mask=label,
          rotate_index=np.random.randint(0, 4),
          shift_amount=np.random.randint(-4, +4),
      )

      TX = (np.random.rand() * 0.6) + 0.4
      TY = (np.random.rand() * 0.6) + 0.4
      X = np.random.rand()
      Y = np.random.rand()
      return random_crop_full(image, X, Y, TX,
                              TY), random_crop_full(label, X, Y, TX, TY)


  def randomCrop_Old(image, label):

      # image, label = garment_color_jitter_rotate(
      #     image=image,
      #     mask=label,
      #     rotate_index=np.random.randint(0, 4),
      #     shift_amount=np.random.randint(0, 256))

      # image, label = garment_color_jitter_rotate(
      #     image=image,
      #     mask=label,
      #     rotate_index=np.random.randint(0, 4),
      #     shift_amount=0,
      # )

      T = (np.random.rand() * 0.6) + 0.4
      X = np.random.rand()
      Y = np.random.rand()
      return random_crop(image, X, Y, T), random_crop(label, X, Y, T)


  def randomCrop(image, label):
      return randomCrop_Both(image, label)


  def randomCrop_original(image, label):
      image_width = image.size[0]
      image_height = image.size[1]
      border = min(image_width, image_height) // 2

      crop_win_width = np.random.randint(image_width - border, image_width)
      crop_win_height = np.random.randint(image_height - border, image_height)

      random_region = ((image_width - crop_win_width) >> 1,
                       (image_height - crop_win_height) >> 1,
                       (image_width + crop_win_width) >> 1,
                       (image_height + crop_win_height) >> 1)

      return image.crop(random_region), label.crop(random_region)


  def randomRotation(image, label):
      mode = Image.BICUBIC
      if random.random() > 0.8:
          random_angle = np.random.randint(-15, 15)
          image = image.rotate(random_angle, mode)
          label = label.rotate(random_angle, mode)
      return image, label


  def colorEnhance(image):
      bright_intensity = random.randint(5, 15) / 10.0
      image = ImageEnhance.Brightness(image).enhance(bright_intensity)
      contrast_intensity = random.randint(5, 15) / 10.0
      image = ImageEnhance.Contrast(image).enhance(contrast_intensity)
      color_intensity = random.randint(0, 20) / 10.0
      image = ImageEnhance.Color(image).enhance(color_intensity)
      sharp_intensity = random.randint(0, 30) / 10.0
      image = ImageEnhance.Sharpness(image).enhance(sharp_intensity)
      return image


  def randomGaussian(image, mean=0.1, sigma=0.35):

      def gaussianNoisy(im, mean=mean, sigma=sigma):
          for _i in range(len(im)):
              im[_i] += random.gauss(mean, sigma)
          return im

      img = np.asarray(image)
      width, height = img.shape
      img = gaussianNoisy(img[:].flatten(), mean, sigma)
      img = img.reshape([width, height])
      return Image.fromarray(np.uint8(img))


  def randomPeper(img):
      img = np.array(img)
      noiseNum = int(0.0015 * img.shape[0] * img.shape[1])
      for i in range(noiseNum):

          randX = random.randint(0, img.shape[0] - 1)

          randY = random.randint(0, img.shape[1] - 1)

          if random.randint(0, 1) == 0:

              img[randX, randY] = 0

          else:

              img[randX, randY] = 255
      return Image.fromarray(img)


  # dataloader for training
  def get_loader(image_root,
                 gt_root,
                 batchsize,
                 trainsize,
                 shuffle=True,
                 num_workers=12,
                 pin_memory=False):
      print('DEBUG 6')
      dataset = DISDataset(image_root, gt_root, trainsize)
      print('DEBUG 7')
      data_loader = data.DataLoader(dataset=dataset,
                                    batch_size=batchsize,
                                    shuffle=shuffle,
                                    num_workers=num_workers,
                                    pin_memory=pin_memory)
      print('DEBUG 8')
      return data_loader
#+end_src

** train.class.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.class.py
  class AvgMeter(object):

      def __init__(self, num=40):
          self.num = num
          self.reset()

      def reset(self):
          self.val = 0
          self.avg = 0
          self.sum = 0
          self.count = 0
          self.losses = []

      def update(self, val, n=1):
          self.val = val
          self.sum += val * n
          self.count += n
          self.avg = self.sum / self.count
          self.losses.append(val)

      def show(self):
          a = len(self.losses)
          b = np.maximum(a - self.num, 0)
          c = self.losses[b:]
          #print(c)
          #d = torch.mean(torch.stack(c))
          #print(d)
          return torch.mean(torch.stack(c))


  class Running_Avg(object):

      def __init__(self, weight=0.999):
          self.weight = weight
          self.reset()

      def reset(self):
          self.n = 0
          self.val = 0

      def update(self, val, n=1):
          self.val = (self.weight * self.val) + ((1 - self.weight) * val)
          self.n = (self.weight * self.n) + ((1 - self.weight) * n)

      def show(self):
          if self.n == 0:
              return 0
          else:
              return self.val / self.n
#+end_src

** Main training dataset

*** COMMENT Original
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.class.py
  # dataset for training
  # The current loader is not using the normalized depth maps for training and test. If you use the normalized depth maps
  # (e.g., 0 represents background and 1 represents foreground.), the performance will be further improved.
  class DISDataset(data.Dataset):

      def __init__(self, image_root, gt_root, trainsize):
          self.trainsize = trainsize
          self.images = [
              image_root + f for f in os.listdir(image_root)
              if f.endswith('.jpg') or f.endswith('.png') or f.endswith('tif')
          ]
          self.gts = [
              gt_root + f for f in os.listdir(gt_root)
              if f.endswith('.jpg') or f.endswith('.png') or f.endswith('tif')
          ]
          self.images = sorted(self.images)
          self.gts = sorted(self.gts)
          self.filter_files()
          self.size = len(self.images)
          self.img_transform = transforms.Compose([
              transforms.Resize((self.trainsize, self.trainsize)),
              transforms.ToTensor(),
              transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
          ])
          self.gt_transform = transforms.Compose([
              transforms.Resize((self.trainsize, self.trainsize)),
              transforms.ToTensor()
          ])

      def __getitem__(self, index):
          image = self.rgb_loader(self.images[index])
          gt = self.binary_loader(self.gts[index])
          image, gt = cv_random_flip(image, gt)
          image, gt = randomCrop(image, gt)
          image, gt = randomRotation(image, gt)
          image = colorEnhance(image)
          image = self.img_transform(image)
          gt = self.gt_transform(gt)

          return image, gt

      def filter_files(self):
          assert len(self.images) == len(self.gts) and len(self.gts) == len(
              self.images)
          images = []
          gts = []
          for img_path, gt_path in zip(self.images, self.gts):
              img = Image.open(img_path)
              gt = Image.open(gt_path)
              if img.size == gt.size:
                  images.append(img_path)
                  gts.append(gt_path)
          self.images = images
          self.gts = gts

      def rgb_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('RGB')

      def binary_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('L')

      def resize(self, img, gt):
          assert img.size == gt.size
          w, h = img.size
          if h < self.trainsize or w < self.trainsize:
              h = max(h, self.trainsize)
              w = max(w, self.trainsize)
              return img.resize((w, h), Image.BILINEAR), gt.resize((w, h),
                                                                   Image.NEAREST)
          else:
              return img, gt

      def __len__(self):
          return self.size
#+end_src

*** Changed
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.class.py
  # dataset for training
  # The current loader is not using the normalized depth maps for training and test. If you use the normalized depth maps
  # (e.g., 0 represents background and 1 represents foreground.), the performance will be further improved.
  class DISDataset(data.Dataset):

      def __init__(self, image_root, gt_root, trainsize):
          self.trainsize = trainsize
          end_pattern = '_segm.png'
          files = list(f for f in os.listdir(gt_root) if f.endswith(end_pattern))
          files.sort()

          self.gts = list(gt_root + f for f in files)

          self.images = list(image_root + f[0:-len(end_pattern)] + '.jpg'
                             for f in files)

          self.size = len(self.images)

          self.img_transform = transforms.Compose([
              transforms.Resize((self.trainsize, self.trainsize)),
              transforms.ToTensor(),
              transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
          ])

          self.gt_transform = transforms.Compose([
              transforms.Resize((self.trainsize, self.trainsize)),
              transforms.ToTensor()
          ])

      def __getitem__(self, index):
          image = self.rgb_loader(self.images[index])
          gt = self.binary_loader(self.gts[index])
          image, gt = cv_random_flip(image, gt)
          image, gt = randomCrop(image, gt)
          image, gt = randomRotation(image, gt)
          image = colorEnhance(image)
          image = self.img_transform(image)
          gt = self.gt_transform(gt)

          return image, gt

      def filter_files(self):
          assert len(self.images) == len(self.gts) and len(self.gts) == len(
              self.images)
          images = []
          gts = []
          for img_path, gt_path in zip(self.images, self.gts):
              img = Image.open(img_path)
              gt = Image.open(gt_path)
              if img.size == gt.size:
                  images.append(img_path)
                  gts.append(gt_path)
          self.images = images
          self.gts = gts

      def rgb_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('RGB')

      def binary_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('L')

      def resize(self, img, gt):
          assert img.size == gt.size
          w, h = img.size
          if h < self.trainsize or w < self.trainsize:
              h = max(h, self.trainsize)
              w = max(w, self.trainsize)
              return img.resize((w, h), Image.BILINEAR), gt.resize((w, h),
                                                                   Image.NEAREST)
          else:
              return img, gt

      def __len__(self):
          return self.size
#+end_src

** train.class.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.class.py
  # test dataset and loader
  class test_dataset:

      def __init__(self, image_root, depth_root, testsize):
          self.testsize = testsize
          self.images = [
              image_root + f for f in os.listdir(image_root)
              if f.endswith('.jpg')
          ]
          self.depths = [
              depth_root + f for f in os.listdir(depth_root)
              if f.endswith('.bmp') or f.endswith('.png')
          ]
          self.images = sorted(self.images)
          self.depths = sorted(self.depths)
          self.transform = transforms.Compose([
              transforms.Resize((self.testsize, self.testsize)),
              transforms.ToTensor(),
              transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
          ])
          # self.gt_transform = transforms.Compose([
          #     transforms.Resize((self.trainsize, self.trainsize)),
          #     transforms.ToTensor()])
          self.depths_transform = transforms.Compose([
              transforms.Resize((self.testsize, self.testsize)),
              transforms.ToTensor()
          ])
          self.size = len(self.images)
          self.index = 0

      def load_data(self):
          image = self.rgb_loader(self.images[self.index])
          HH = image.size[0]
          WW = image.size[1]
          image = self.transform(image).unsqueeze(0)
          depth = self.rgb_loader(self.depths[self.index])
          depth = self.depths_transform(depth).unsqueeze(0)

          name = self.images[self.index].split('/')[-1]
          # image_for_post=self.rgb_loader(self.images[self.index])
          # image_for_post=image_for_post.resize(gt.size)
          if name.endswith('.jpg'):
              name = name.split('.jpg')[0] + '.png'
          self.index += 1
          self.index = self.index % self.size
          return image, depth, HH, WW, name

      def rgb_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('RGB')

      def binary_loader(self, path):
          with open(path, 'rb') as f:
              img = Image.open(f)
              return img.convert('L')

      def __len__(self):
          return self.size


  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,
                                    device='cuda')

      def __call__(self, b, h, w):
          mask = torch.zeros([b, h, w], dtype=torch.bool, device='cuda')
          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).cuda()
              x_embed = ((x_embed - 0.5) / (x_embed[:, :, -1:] + eps) *
                         self.scale).cuda()

          dim_t = self.temperature**(2 * (self.dim_t // 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('relu')
          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')

          # 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 inf_MCLM(nn.Module):

      def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]):
          super(inf_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('relu')
          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)
          self.p_poses = []
          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')

          # 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('relu')
          self.sal_conv = nn.Conv2d(d_model, 1, 1)
          self.pool_ratios = pool_ratios
          self.positional_encoding = PositionEmbeddingSine(
              num_pos_feats=d_model // 2, normalize=True)

      def forward(self, x):
          b, c, h, w = x.size()
          loc, glb = x.split([4, 1], dim=0)  # 4,c,h,w; 1,c,h,w
          # b(4),c,h,w
          patched_glb = rearrange(glb,
                                  'b c (hg h) (wg w) -> (hg wg b) c h w',
                                  hg=2,
                                  wg=2)

          # generate token attention map
          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
          # nl(4),c,nphw -> nl(4),nphw,1,c
          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
              # np*hw,1,c
              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 inf_MCRM(nn.Module):

      def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None):
          super(inf_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('relu')
          self.sal_conv = nn.Conv2d(d_model, 1, 1)
          self.pool_ratios = pool_ratios
          self.positional_encoding = PositionEmbeddingSine(
              num_pos_feats=d_model // 2, normalize=True)

      def forward(self, x):
          b, c, h, w = x.size()
          loc, glb = x.split([4, 1], dim=0)  # 4,c,h,w; 1,c,h,w
          # b(4),c,h,w
          patched_glb = rearrange(glb,
                                  'b c (hg h) (wg w) -> (hg wg b) c h w',
                                  hg=2,
                                  wg=2)

          # generate token attention map
          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
          # nl(4),c,nphw -> nl(4),nphw,1,c
          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
              # np*hw,1,c
              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)


  # model for single-scale training
  class MVANet(nn.Module):

      def __init__(self):
          super().__init__()
          self.backbone = SwinB(pretrained=True)
          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.BatchNorm2d(384), nn.PReLU(),
              nn.Conv2d(384, 384, kernel_size=3, padding=1), nn.BatchNorm2d(384),
              nn.PReLU(), 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.ReLU) or isinstance(m, nn.Dropout):
                  m.inplace = True

      def forward(self, x):
          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)
          # 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
          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)
          ####
          sideout5 = self.sideout5(e5).cuda()
          sideout4 = self.sideout4(e4)
          sideout3 = self.sideout3(e3)
          sideout2 = self.sideout2(e2)
          sideout1 = self.sideout1(e1)
          #######glb_sideouts ######
          glb5 = self.sideout5(glb_e5)
          glb4 = sideout4[-1, :, :, :].unsqueeze(0)
          glb3 = sideout3[-1, :, :, :].unsqueeze(0)
          glb2 = sideout2[-1, :, :, :].unsqueeze(0)
          glb1 = sideout1[-1, :, :, :].unsqueeze(0)
          ####### concat 4 to 1 #######
          sideout1 = patches2image(sideout1[:-1]).cuda()
          sideout2 = patches2image(
              sideout2[:-1]).cuda()  ####(5,c,h,w) -> (1 c 2h,2w)
          sideout3 = patches2image(sideout3[:-1]).cuda()
          sideout4 = patches2image(sideout4[:-1]).cuda()
          sideout5 = patches2image(sideout5[:-1]).cuda()
          if self.training:
              return sideout5, sideout4, sideout3, sideout2, sideout1, final_output, glb5, glb4, glb3, glb2, glb1, tokenattmap4, tokenattmap3, tokenattmap2, tokenattmap1
          else:
              return final_output


  # model for multi-scale testing
  class inf_MVANet(nn.Module):

      def __init__(self):
          super().__init__()
          self.backbone = SwinB(pretrained=True)

          emb_dim = 128
          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 = inf_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 = inf_MCRM(emb_dim, 1, [2, 4, 8])
          self.dec_blk2 = inf_MCRM(emb_dim, 1, [2, 4, 8])
          self.dec_blk3 = inf_MCRM(emb_dim, 1, [2, 4, 8])
          self.dec_blk4 = inf_MCRM(emb_dim, 1, [2, 4, 8])

          self.insmask_head = nn.Sequential(
              nn.Conv2d(emb_dim, 384, kernel_size=3, padding=1),
              nn.BatchNorm2d(384), nn.PReLU(),
              nn.Conv2d(384, 384, kernel_size=3, padding=1), nn.BatchNorm2d(384),
              nn.PReLU(), 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.ReLU) or isinstance(m, nn.Dropout):
                  m.inplace = True

      def forward(self, x):
          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])
          e4 = self.output4(feature[3])
          e3 = self.output3(feature[2])
          e2 = self.output2(feature[1])
          e1 = self.output1(feature[0])
          print(e5.shape)
          loc_e5, glb_e5 = e5.split([4, 1], dim=0)
          e5_cat = self.multifieldcrossatt(loc_e5, glb_e5)

          e4 = self.conv4(self.dec_blk4(e4 + resize_as(e5_cat, e4)))
          e3 = self.conv3(self.dec_blk3(e3 + resize_as(e4, e3)))
          e2 = self.conv2(self.dec_blk2(e2 + resize_as(e3, e2)))
          e1 = self.conv1(self.dec_blk1(e1 + resize_as(e2, e1)))
          loc_e1, glb_e1 = e1.split([4, 1], dim=0)
          # after decoder, concat loc features to a whole one, and merge
          output1_cat = patches2image(loc_e1)
          # add glb feat in
          output1_cat = output1_cat + resize_as(glb_e1, output1_cat)
          # merge
          final_output = self.insmask_head(output1_cat)
          # shallow feature merge
          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
#+end_src

** train.execute.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.execute.py
  writer = SummaryWriter()

  cudnn.benchmark = True

  parser = argparse.ArgumentParser()
  parser.add_argument('--epoch', type=int, default=80, help='epoch number')
  parser.add_argument('--lr_gen', type=float, default=1e-5, help='learning rate')
  parser.add_argument('--batchsize',
                      type=int,
                      default=1,
                      help='training batch size')
  parser.add_argument('--trainsize',
                      type=int,
                      default=1024,
                      help='training dataset size')
  parser.add_argument('--decay_rate',
                      type=float,
                      default=0.9,
                      help='decay rate of learning rate')
  parser.add_argument('--decay_epoch',
                      type=int,
                      default=80,
                      help='every n epochs decay learning rate')

  opt = parser.parse_args()
  print('Generator Learning Rate: {}'.format(opt.lr_gen))
  # build models
  if hasattr(torch.cuda, 'empty_cache'):
      torch.cuda.empty_cache()
  generator = MVANet()
  generator.cuda()
  print('DEBUG 3')

  pretrained_dict = torch.load(
      HOME_DIR +
      '/GITHUB/aravind-h-v/dreambooth_experiments/cloth_segmentation/MVANet_Train/pretrained_model/Model_80.pth',
      map_location='cuda')

  model_dict = generator.state_dict()
  pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
  model_dict.update(pretrained_dict)
  generator.load_state_dict(model_dict)

  generator_params = generator.parameters()
  # generator_optimizer = torch.optim.Adam(generator_params, opt.lr_gen)
  generator_optimizer = Prodigy(generator_params, lr=1., weight_decay=0.01)

  print('DEBUG 4')

  image_root = './data/image/'
  gt_root = './data/mask/'

  train_loader = get_loader(image_root,
                            gt_root,
                            batchsize=opt.batchsize,
                            trainsize=opt.trainsize)

  print('DEBUG 5')

  total_step = len(train_loader)
  to_pil = transforms.ToPILImage()
  ## define loss
  print('DEBUG 2')

  CE = torch.nn.BCELoss()
  mse_loss = torch.nn.MSELoss(size_average=True, reduce=True)
  size_rates = [1]
  criterion = nn.BCEWithLogitsLoss().cuda()
  criterion_mae = nn.L1Loss().cuda()
  criterion_mse = nn.MSELoss().cuda()
  use_fp16 = True
  scaler = amp.GradScaler(enabled=use_fp16)
  print('DEBUG 1')

  for epoch in range(1, opt.epoch + 1):
      torch.cuda.empty_cache()
      generator.train()
      # loss_record = AvgMeter()
      loss_record = Running_Avg()
      print('Generator Learning Rate: {}'.format(
          generator_optimizer.param_groups[0]['lr']))
      for i, pack in enumerate(train_loader, start=1):
          torch.cuda.empty_cache()
          for rate in size_rates:
              torch.cuda.empty_cache()
              generator_optimizer.zero_grad()
              images, gts = pack
              images = Variable(images)
              gts = Variable(gts)
              images = images.cuda()
              gts = gts.cuda()
              trainsize = int(round(opt.trainsize * rate / 32) * 32)
              if rate != 1:
                  images = F.upsample(images,
                                      size=(trainsize, trainsize),
                                      mode='bilinear',
                                      align_corners=True)
                  gts = F.upsample(gts,
                                   size=(trainsize, trainsize),
                                   mode='bilinear',
                                   align_corners=True)

              b, c, h, w = gts.size()
              target_1 = F.upsample(gts, size=h // 4, mode='nearest')
              target_2 = F.upsample(gts, size=h // 8, mode='nearest').cuda()
              target_3 = F.upsample(gts, size=h // 16, mode='nearest').cuda()
              target_4 = F.upsample(gts, size=h // 32, mode='nearest').cuda()
              target_5 = F.upsample(gts, size=h // 64, mode='nearest').cuda()

              with amp.autocast(enabled=use_fp16):
                  sideout5, sideout4, sideout3, sideout2, sideout1, final, glb5, glb4, glb3, glb2, glb1, tokenattmap4, tokenattmap3, tokenattmap2, tokenattmap1 = generator.forward(
                      images)
                  loss1 = structure_loss(sideout5, target_4)
                  loss2 = structure_loss(sideout4, target_3)
                  loss3 = structure_loss(sideout3, target_2)
                  loss4 = structure_loss(sideout2, target_1)
                  loss5 = structure_loss(sideout1, target_1)
                  loss6 = structure_loss(final, gts)
                  loss7 = structure_loss(glb5, target_5)
                  loss8 = structure_loss(glb4, target_4)
                  loss9 = structure_loss(glb3, target_3)
                  loss10 = structure_loss(glb2, target_2)
                  loss11 = structure_loss(glb1, target_2)
                  loss12 = structure_loss(tokenattmap4, target_3)
                  loss13 = structure_loss(tokenattmap3, target_2)
                  loss14 = structure_loss(tokenattmap2, target_1)
                  loss15 = structure_loss(tokenattmap1, target_1)
                  loss = loss1 + loss2 + loss3 + loss4 + loss5 + loss6 + 0.3 * (
                      loss7 + loss8 + loss9 + loss10 +
                      loss11) + 0.3 * (loss12 + loss13 + loss14 + loss15)
                  Loss_loc = loss1 + loss2 + loss3 + loss4 + loss5 + loss6
                  Loss_glb = loss7 + loss8 + loss9 + loss10 + loss11
                  Loss_map = loss12 + loss13 + loss14 + loss15
                  writer.add_scalar('loss', loss.item(),
                                    epoch * len(train_loader) + i)

              generator_optimizer.zero_grad()
              scaler.scale(loss).backward()
              scaler.step(generator_optimizer)
              scaler.update()

              if rate == 1:
                  loss_record.update(loss.data, opt.batchsize)

          if i % 10 == 0 or i == total_step:
              print(
                  '{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], gen Loss: {:.4f}'
                  .format(datetime.now(), epoch, opt.epoch, i, total_step,
                          loss_record.show()))

          if i % 8000 == 0 or i == total_step:
              save_path = './saved_model/'
              if not os.path.exists(save_path):
                  os.mkdir(save_path)
              torch.save(
                  generator.state_dict(),
                  save_path + 'Model' + '_%d' % epoch + '_%d' % i + '.pth')

      # adjust_lr(generator_optimizer, opt.lr_gen, epoch, opt.decay_rate,
      #           opt.decay_epoch)
      # save checkpoints every 20 epochs
      # if epoch % 20 == 0:
      if True:

          save_path = './saved_model/'
          if not os.path.exists(save_path):
              os.mkdir(save_path)

          save_path = './saved_model/MVANet/'
          if not os.path.exists(save_path):
              os.mkdir(save_path)

          torch.save(generator.state_dict(),
                     save_path + 'Model' + '_%d' % epoch + '.pth')
#+end_src

* SAMPLE

** train

*** train.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.import.py
#+end_src

*** train.function.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.function.py
#+end_src

*** train.class.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.class.py
#+end_src

*** train.execute.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./train.execute.py
#+end_src

** swin

*** swin.import.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.import.py
#+end_src

*** swin.function.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.function.py
#+end_src

*** swin.class.py
#+begin_src python :shebang #!/usr/bin/python3 :results output :tangle ./swin.class.py
#+end_src

* UNIFY
#+begin_src sh :shebang #!/bin/sh :results output :tangle ./train.unify.sh
  . "${HOME}/dbnew.sh"

  echo '#!/usr/bin/python3' > './train.py' 

  cat \
          './train.import.py' \
          './train.function.py' \
          './train.class.py' \
          './train.execute.py' \
      | expand | yapf3 \
      | grep -v '^#!/usr/bin/python3$' \
      >> './train.py' \
  ;

  echo '#!/usr/bin/python3' > './swin.py' 

  cat \
          './swin.import.py' \
          './swin.function.py' \
          './swin.class.py' \
      | expand | yapf3 \
      | grep -v '^#!/usr/bin/python3$' \
      >> './swin.py' \
  ;

  rm -vf -- \
      './swin.class.py' \
      './swin.function.py' \
      './swin.import.py' \
      './train.class.py' \
      './train.execute.py' \
      './train.function.py' \
      './train.import.py' \
      './train.unify.sh' \
  ;
#+end_src

* Run
#+begin_src sh :shebang #!/bin/sh :results output :tangle ./run.sh
  . "${HOME}/dbnew.sh"

  cd "$('dirname' '--' "${0}")"

  pip3 install -r './requirements.txt'

  python3 ./train.py --batchsize 4
#+end_src

* WORK SPACE

** ELISP
#+begin_src elisp
  (save-buffer)
  (org-babel-tangle)
  (shell-command "./train.unify.sh")
#+end_src

#+RESULTS:
: 0

** SHELL
#+begin_src sh :shebang #!/bin/sh :results output 
  realpath .
  cd /home/asd/GITHUB/aravind-h-v/dreambooth_experiments/cloth_segmentation/MVANet_Train
#+end_src