Add IPAD model implementation

IPAD/__init__.py

@@ -0,0 +1,5 @@
+# IPAD Model Package
+from .model.video_swin_transformer import VST
+from .model.memory_module import MemModule
+
+__all__ = ['VST', 'MemModule']

IPAD/model/VST_block.py

@@ -0,0 +1,568 @@
+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, trunc_normal_
+
+from functools import reduce, lru_cache
+from operator import mul
+from einops import rearrange
+
+import logging
+
+
+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, D, H, W, C)
+        window_size (tuple[int]): window size
+    Returns:
+        windows: (B*num_windows, window_size*window_size, C)
+    """
+    B, D, H, W, C = x.shape
+    x = x.view(B, D // window_size[0], window_size[0], H // window_size[1], window_size[1], W // window_size[2], window_size[2], C)
+    windows = x.permute(0, 1, 3, 5, 2, 4, 6, 7).contiguous().view(-1, reduce(mul, window_size), C)
+    return windows
+
+
+def window_reverse(windows, window_size, B, D, H, W):
+    """
+    Args:
+        windows: (B*num_windows, window_size, window_size, C)
+        window_size (tuple[int]): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, D, H, W, C)
+    """
+    x = windows.view(B, D // window_size[0], H // window_size[1], W // window_size[2], window_size[0], window_size[1], window_size[2], -1)
+    x = x.permute(0, 1, 4, 2, 5, 3, 6, 7).contiguous().view(B, D, H, W, -1)
+    return x
+
+
+
+
+def get_window_size(x_size, window_size, shift_size=None):
+    use_window_size = list(window_size)
+    if shift_size is not None:
+        use_shift_size = list(shift_size)
+    for i in range(len(x_size)):
+        if x_size[i] <= window_size[i]:
+            use_window_size[i] = x_size[i]
+            if shift_size is not None:
+                use_shift_size[i] = 0
+
+    if shift_size is None:
+        return tuple(use_window_size)
+    else:
+        return tuple(use_window_size), tuple(use_shift_size)
+
+
+class WindowAttention3D(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 temporal length, 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=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wd, 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) * (2 * window_size[2] - 1), num_heads))  # 2*Wd-1 * 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_d = torch.arange(self.window_size[0])
+        coords_h = torch.arange(self.window_size[1])
+        coords_w = torch.arange(self.window_size[2])
+        coords = torch.stack(torch.meshgrid(coords_d, coords_h, coords_w))  # 3, Wd, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 3, Wd*Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 3, Wd*Wh*Ww, Wd*Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wd*Wh*Ww, Wd*Wh*Ww, 3
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 2] += self.window_size[2] - 1
+
+        relative_coords[:, :, 0] *= (2 * self.window_size[1] - 1) * (2 * self.window_size[2] - 1)
+        relative_coords[:, :, 1] *= (2 * self.window_size[2] - 1)
+        relative_position_index = relative_coords.sum(-1)  # Wd*Wh*Ww, Wd*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, N, N) 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]  # B_, nH, N, C
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index[:N, :N].reshape(-1)].reshape(
+            N, N, -1)  # Wd*Wh*Ww,Wd*Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wd*Wh*Ww, Wd*Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0) # B_, nH, N, N
+
+        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 SwinTransformerBlock3D(nn.Module):
+    """ Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (tuple[int]): Window size.
+        shift_size (tuple[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=(2,7,7), shift_size=(0,0,0),
+                 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, use_checkpoint=False):
+        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
+        self.use_checkpoint=use_checkpoint
+
+        assert 0 <= self.shift_size[0] < self.window_size[0], "shift_size must in 0-window_size"
+        assert 0 <= self.shift_size[1] < self.window_size[1], "shift_size must in 0-window_size"
+        assert 0 <= self.shift_size[2] < self.window_size[2], "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention3D(
+            dim, window_size=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)
+
+    def forward_part1(self, x, mask_matrix):
+        B, D, H, W, C = x.shape
+        window_size, shift_size = get_window_size((D, H, W), self.window_size, self.shift_size)
+
+        x = self.norm1(x)
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = pad_d0 = 0
+        pad_d1 = (window_size[0] - D % window_size[0]) % window_size[0]
+        pad_b = (window_size[1] - H % window_size[1]) % window_size[1]
+        pad_r = (window_size[2] - W % window_size[2]) % window_size[2]
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b, pad_d0, pad_d1))
+        _, Dp, Hp, Wp, _ = x.shape
+        # cyclic shift
+        if any(i > 0 for i in shift_size):
+            shifted_x = torch.roll(x, shifts=(-shift_size[0], -shift_size[1], -shift_size[2]), dims=(1, 2, 3))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+        # partition windows
+        x_windows = window_partition(shifted_x, window_size)  # B*nW, Wd*Wh*Ww, C
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # B*nW, Wd*Wh*Ww, C
+        # merge windows
+        attn_windows = attn_windows.view(-1, *(window_size+(C,)))
+        shifted_x = window_reverse(attn_windows, window_size, B, Dp, Hp, Wp)  # B D' H' W' C
+        # reverse cyclic shift
+        if any(i > 0 for i in shift_size):
+            x = torch.roll(shifted_x, shifts=(shift_size[0], shift_size[1], shift_size[2]), dims=(1, 2, 3))
+        else:
+            x = shifted_x
+
+        if pad_d1 >0 or pad_r > 0 or pad_b > 0:
+            x = x[:, :D, :H, :W, :].contiguous()
+        return x
+
+    def forward_part2(self, x):
+        return self.drop_path(self.mlp(self.norm2(x)))
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, D, H, W, C).
+            mask_matrix: Attention mask for cyclic shift.
+        """
+
+        shortcut = x
+        if self.use_checkpoint:
+            x = checkpoint.checkpoint(self.forward_part1, x, mask_matrix)
+        else:
+            x = self.forward_part1(x, mask_matrix)
+        x = shortcut + self.drop_path(x)
+
+        if self.use_checkpoint:
+            x = x + checkpoint.checkpoint(self.forward_part2, x)
+        else:
+            x = x + self.forward_part2(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):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, D, H, W, C).
+        """
+        B, D, H, W, C = x.shape
+
+        # 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 D H/2 W/2 C
+        x1 = x[:, :, 1::2, 0::2, :]  # B D H/2 W/2 C
+        x2 = x[:, :, 0::2, 1::2, :]  # B D H/2 W/2 C
+        x3 = x[:, :, 1::2, 1::2, :]  # B D H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B D H/2 W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+# cache each stage results
+@lru_cache()
+def compute_mask(D, H, W, window_size, shift_size, device):
+    img_mask = torch.zeros((1, D, H, W, 1), device=device)  # 1 Dp Hp Wp 1
+    cnt = 0
+    for d in slice(-window_size[0]), slice(-window_size[0], -shift_size[0]), slice(-shift_size[0],None):
+        for h in slice(-window_size[1]), slice(-window_size[1], -shift_size[1]), slice(-shift_size[1],None):
+            for w in slice(-window_size[2]), slice(-window_size[2], -shift_size[2]), slice(-shift_size[2],None):
+                img_mask[:, d, h, w, :] = cnt
+                cnt += 1
+    mask_windows = window_partition(img_mask, window_size)  # nW, ws[0]*ws[1]*ws[2], 1
+    mask_windows = mask_windows.squeeze(-1)  # nW, ws[0]*ws[1]*ws[2]
+    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))
+    return attn_mask
+
+
+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 (tuple[int]): Local window size. Default: (1,7,7).
+        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
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=(1,7,7),
+                 mlp_ratio=4.,
+                 qkv_bias=False,
+                 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 = tuple(i // 2 for i in window_size)
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock3D(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=(0,0,0) if (i % 2 == 0) else self.shift_size,
+                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,
+                use_checkpoint=use_checkpoint,
+            )
+            for i in range(depth)])
+        
+        self.downsample = downsample
+        if self.downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+
+    def forward(self, x):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, C, D, H, W).
+        """
+        # calculate attention mask for SW-MSA
+        B, C, D, H, W = x.shape
+        window_size, shift_size = get_window_size((D,H,W), self.window_size, self.shift_size)
+        x = rearrange(x, 'b c d h w -> b d h w c')
+        Dp = int(np.ceil(D / window_size[0])) * window_size[0]
+        Hp = int(np.ceil(H / window_size[1])) * window_size[1]
+        Wp = int(np.ceil(W / window_size[2])) * window_size[2]
+        attn_mask = compute_mask(Dp, Hp, Wp, window_size, shift_size, x.device)
+        for blk in self.blocks:
+            x = blk(x, attn_mask)
+        x = x.view(B, D, H, W, -1)
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+        x = rearrange(x, 'b d h w c -> b c d h w')
+        return x
+
+
+class PatchEmbed3D(nn.Module):
+    """ Video to Patch Embedding.
+    Args:
+        patch_size (int): Patch token size. Default: (2,4,4).
+        in_chans (int): Number of input video 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=(2,4,4), in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        self.patch_size = patch_size
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv3d(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
+        _, _, D, H, W = x.size()
+        if W % self.patch_size[2] != 0:
+            x = F.pad(x, (0, self.patch_size[2] - W % self.patch_size[2]))
+        if H % self.patch_size[1] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[1] - H % self.patch_size[1]))
+        if D % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, 0, 0, self.patch_size[0] - D % self.patch_size[0]))
+
+        x = self.proj(x)  # B C D Wh Ww
+        if self.norm is not None:
+            D, Wh, Ww = x.size(2), x.size(3), x.size(4)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, D, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer3D(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        patch_size (int | tuple(int)): Patch size. Default: (4,4,4).
+        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: Truee
+        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: Normalization layer. Default: nn.LayerNorm.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: False.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+    """
+
+    def __init__(self,
+                 pretrained=None,
+                 pretrained2d=True,
+                 patch_size=(4,4,4),
+                 in_chans=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=(2,7,7),
+                 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,
+                 patch_norm=False,
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrained = pretrained
+        self.pretrained2d = pretrained2d
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.patch_norm = patch_norm
+        self.frozen_stages = frozen_stages
+        self.window_size = window_size
+        self.patch_size = patch_size
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed3D(
+            patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        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)
+
+        self.num_features = int(embed_dim * 2**(self.num_layers-1))
+
+        # add a norm layer for each output
+        self.norm = norm_layer(self.num_features)
+
+        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:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages):
+                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)
+
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            
+            x = layer(x.contiguous())
+        x = rearrange(x, 'n c d h w -> n d h w c')
+        x = self.norm(x)
+        x = rearrange(x, 'n d h w c -> n c d h w')
+        return x
+
+    # def train(self, mode=True):
+    #     """Convert the model into training mode while keep layers freezed."""
+    #     super(SwinTransformer3D, self).train(mode)
+    #     self._freeze_stages()

IPAD/model/__init__.py

@@ -0,0 +1,5 @@
+from __future__ import absolute_import, print_function
+
+from model.memory_module import *
+from model.memae_3dconv import *
+from model.entropy_loss import *

IPAD/model/autoencoder.py

@@ -0,0 +1,24 @@
+import torch
+from .reconstruction_model import Reconstruction3DEncoder, Reconstruction3DDecoder
+
+class convAE(torch.nn.Module):
+    def __init__(self):  # for reconstruction
+        super(convAE, self).__init__()
+
+        self.reconstruction = True
+
+        # self.encoder = Reconstruction3DEncoder(chnum_in=1)  # black and white
+        # self.decoder = Reconstruction3DDecoder(chnum_in=1)  # black and white
+        self.encoder = Reconstruction3DEncoder(chnum_in=3)  # RGB
+        self.decoder = Reconstruction3DDecoder(chnum_in=3)  # RGB
+
+    def forward(self, x):
+        # print(x.shape)
+        fea = self.encoder(x)
+        # print(fea.shape)
+        output = self.decoder(fea.clone())
+        # print(output.shape)
+
+        return output
+
+

IPAD/model/entropy_loss.py

@@ -0,0 +1,50 @@
+from __future__ import absolute_import, print_function
+import torch
+from torch import nn
+
+
+def feature_map_permute(input):
+    s = input.data.shape
+    l = len(s)
+
+    # permute feature channel to the last:
+    # NxCxDxHxW --> NxDxHxW x C
+    if l == 2:
+        x = input # NxC
+    elif l == 3:
+        x = input.permute(0, 2, 1)
+    elif l == 4:
+        x = input.permute(0, 2, 3, 1)
+    elif l == 5:
+        x = input.permute(0, 2, 3, 4, 1)
+    else:
+        x = []
+        print('wrong feature map size')
+    x = x.contiguous()
+    # NxDxHxW x C --> (NxDxHxW) x C
+    x = x.view(-1, s[1])
+    return x
+
+class EntropyLoss(nn.Module):
+    def __init__(self, eps = 1e-12):
+        super(EntropyLoss, self).__init__()
+        self.eps = eps
+
+    def forward(self, x):
+        b = x * torch.log(x + self.eps)
+        b = -1.0 * b.sum(dim=1)
+        b = b.mean()
+        return b
+
+class EntropyLossEncap(nn.Module):
+    def __init__(self, eps = 1e-12):
+        super(EntropyLossEncap, self).__init__()
+        self.eps = eps
+        self.entropy_loss = EntropyLoss(eps)
+
+    def forward(self, input):
+        score = feature_map_permute(input)
+        ent_loss_val = self.entropy_loss(score)
+        return ent_loss_val
+
+

IPAD/model/memae_3dconv.py

@@ -0,0 +1,53 @@
+from __future__ import absolute_import, print_function
+import torch
+from torch import nn
+
+from model import MemModule
+
+class AutoEncoderCov3DMem(nn.Module):
+    def __init__(self, chnum_in, mem_dim, shrink_thres=0.0025):
+        super(AutoEncoderCov3DMem, self).__init__()
+        print('AutoEncoderCov3DMem')
+        self.chnum_in = chnum_in
+        feature_num = 128
+        feature_num_2 = 96
+        feature_num_x2 = 256
+        self.encoder = nn.Sequential(
+            nn.Conv3d(self.chnum_in, feature_num_2, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num_2, feature_num, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num_x2, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True)
+        )
+        self.mem_rep = MemModule(mem_dim=mem_dim, fea_dim=feature_num_x2, shrink_thres =shrink_thres)
+        self.decoder = nn.Sequential(
+            nn.ConvTranspose3d(feature_num_x2, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_x2, feature_num, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num, feature_num_2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_2, self.chnum_in, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1),
+                               output_padding=(0, 1, 1))
+        )
+
+    def forward(self, x):
+        f = self.encoder(x)
+        res_mem = self.mem_rep(f)
+        f = res_mem['output']
+        att = res_mem['att']
+        output = self.decoder(f)
+        return {'output': output, 'att': att}

IPAD/model/memory_module.py

@@ -0,0 +1,112 @@
+from __future__ import absolute_import, print_function
+import torch
+from torch import nn
+import math
+from torch.nn.parameter import Parameter
+from torch.nn import functional as F
+import numpy as np
+
+#
+class MemoryUnit(nn.Module):
+    def __init__(self, mem_dim, fea_dim, shrink_thres=0.0025):
+        super(MemoryUnit, self).__init__()
+        self.mem_dim = mem_dim
+        self.fea_dim = fea_dim
+        self.weight = Parameter(torch.Tensor(self.mem_dim, self.fea_dim))  # M x C
+        self.bias = None
+        self.shrink_thres= shrink_thres
+        # self.hard_sparse_shrink_opt = nn.Hardshrink(lambd=shrink_thres)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        stdv = 1. / math.sqrt(self.weight.size(1))
+        self.weight.data.uniform_(-stdv, stdv)
+        if self.bias is not None:
+            self.bias.data.uniform_(-stdv, stdv)
+
+    def forward(self, input, period_score):
+        # print(input.shape)
+        score,indices = torch.max(period_score,dim=1)
+        indices = (torch.floor((indices/126)*self.mem_dim).cpu().numpy()).astype(int)
+        # # print(indices)
+        att_weight = F.linear(input, self.weight)  # Fea x Mem^T, (TxC) x (CxM) = TxM
+        a = score[i]
+        att_weight[:,indices[i]-7:indices[i]+8]=att_weight[:,indices[i]-7:indices[i]+8]+att_weight[:,indices[i]-7:indices[i]+8].clone()*score[i]
+        att_weight = F.softmax(att_weight, dim=1)  # TxM
+        # print(att_weight.shape)
+        # print(period_score.shape)
+        # ReLU based shrinkage, hard shrinkage for positive value
+        if(self.shrink_thres>0):
+            att_weight = hard_shrink_relu(att_weight, lambd=self.shrink_thres)
+            # att_weight = F.softshrink(att_weight, lambd=self.shrink_thres)
+            # normalize???
+            att_weight = F.normalize(att_weight, p=1, dim=1)
+            # att_weight = F.softmax(att_weight, dim=1)
+            # att_weight = self.hard_sparse_shrink_opt(att_weight)
+        
+        mem_trans = self.weight.permute(1, 0)  # Mem^T, MxC
+        output = F.linear(att_weight, mem_trans)  # AttWeight x Mem^T^T = AW x Mem, (TxM) x (MxC) = TxC
+        return {'output': output, 'att': att_weight}  # output, att_weight
+
+    def extra_repr(self):
+        return 'mem_dim={}, fea_dim={}'.format(
+            self.mem_dim, self.fea_dim is not None
+        )
+
+
+# NxCxHxW -> (NxHxW)xC -> addressing Mem, (NxHxW)xC -> NxCxHxW
+class MemModule(nn.Module):
+    def __init__(self, mem_dim, fea_dim, shrink_thres=0.0025, device='cuda'):
+        super(MemModule, self).__init__()
+        self.mem_dim = mem_dim
+        self.fea_dim = fea_dim
+        self.shrink_thres = shrink_thres
+        self.memory = MemoryUnit(self.mem_dim, self.fea_dim, self.shrink_thres)
+
+    def forward(self, input, period_score):
+        s = input.data.shape
+        l = len(s)# 5
+        if l == 3:
+            x = input.permute(0, 2, 1)
+        elif l == 4:
+            x = input.permute(0, 2, 3, 1)
+        elif l == 5:
+            x = input.permute(0, 2, 3, 4, 1)
+        else:
+            x = []
+            print('wrong feature map size')
+        x = x.contiguous()
+        x = x.view(-1, s[1])
+        #
+        y_and = self.memory(x,period_score)
+        #
+        y = y_and['output']
+        att = y_and['att']
+
+        if l == 3:
+            y = y.view(s[0], s[2], s[1])
+            y = y.permute(0, 2, 1)
+            att = att.view(s[0], s[2], self.mem_dim)
+            att = att.permute(0, 2, 1)
+        elif l == 4:
+            y = y.view(s[0], s[2], s[3], s[1])
+            y = y.permute(0, 3, 1, 2)
+            att = att.view(s[0], s[2], s[3], self.mem_dim)
+            att = att.permute(0, 3, 1, 2)
+        elif l == 5:
+            y = y.view(s[0], s[2], s[3], s[4], s[1])
+            y = y.permute(0, 4, 1, 2, 3)
+            att = att.view(s[0], s[2], s[3], s[4], self.mem_dim)
+            att = att.permute(0, 4, 1, 2, 3)
+        else:
+            y = x
+            att = att
+            print('wrong feature map size')
+        return {'output': y, 'att': att}
+
+# relu based hard shrinkage function, only works for positive values
+def hard_shrink_relu(input, lambd=0, epsilon=1e-12):
+    output = (F.relu(input-lambd) * input) / (torch.abs(input - lambd) + epsilon)
+    return output
+

IPAD/model/pseudoanomaly_utils.py

@@ -0,0 +1,298 @@
+import random
+import torch
+import numpy as np
+import torch.nn.functional as F
+import torchvision.transforms as transforms
+import copy
+
+def create_pseudoanomaly_cifar_smooth(img, cifar_img, max_size, h, w, dataset, max_move=0):
+    assert 0 <= max_size <= 1
+
+    pil_img = transforms.ToPILImage()(cifar_img)
+    pil_img = transforms.Grayscale(num_output_channels=1)(pil_img)
+    cifar_img = transforms.ToTensor()(pil_img)
+
+    cifar_img = transforms.Normalize(mean=[0.5], std=[0.5])(cifar_img)
+
+    cifar_patch = F.interpolate(cifar_img.unsqueeze(0), size=(h, w), mode='bilinear', align_corners=True)
+
+    x_mu, y_mu = random.randint(0, w), random.randint(0, h)
+    x_sigma = max(10, int(np.random.uniform(high=max_size) * w))
+    y_sigma = max(10, int(np.random.uniform(high=max_size) * h))
+    if max_move == 0:
+        mask = torch.tensor(_get_gaussian_mask(x_mu, y_mu, x_sigma, y_sigma, h, w)).to(img.device).float()
+        img = mask * cifar_patch.to(img.device) + (1-mask) * img
+    else:
+        mask = []
+        for frame_idx in range(img.size(1)):
+            delta_x = np.random.randint(-max_move, max_move + 1)
+            delta_y = np.random.randint(-max_move, max_move + 1)
+            mask_ = torch.tensor(_get_gaussian_mask(x_mu, y_mu, x_sigma, y_sigma, h, w)).to(img.device).float()
+
+            img[:, frame_idx] = mask_ * cifar_patch.to(img.device) + (1-mask_) * img[:, frame_idx]
+            mask.append(mask_)
+
+            x_mu = min(max(0, x_mu + delta_x), w)
+            y_mu = min(max(0, y_mu + delta_y), h)
+
+        mask = torch.stack(mask, dim=0)
+
+    return img, mask
+
+
+def create_pseudoanomaly_cifar_smoothborder(img, cifar_img, max_size, h, w, dataset, max_move=0):
+    assert 0 <= max_size <= 1
+
+    pil_img = transforms.ToPILImage()(cifar_img)
+    pil_img = transforms.Grayscale(num_output_channels=1)(pil_img)
+    cifar_img = transforms.ToTensor()(pil_img)
+
+
+    cifar_img = transforms.Normalize(mean=[0.5], std=[0.5])(cifar_img)
+
+    cifar_patch = F.interpolate(cifar_img.unsqueeze(0), size=(h, w), mode='bilinear', align_corners=True)
+
+    cx, cy = np.random.randint(w), np.random.randint(h)
+
+    cut_w= max(10, int(np.random.uniform(high=max_size) * w))
+    cut_h = max(10, int(np.random.uniform(high=max_size) * h))
+    if max_move == 0:
+        mask = torch.tensor(_get_smoothborder_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+        img = mask * cifar_patch.to(img.device) + (1-mask) * img
+
+    else:
+        mask = []
+        for frame_idx in range(img.size(1)):
+            delta_x = np.random.randint(-max_move, max_move + 1)
+            delta_y = np.random.randint(-max_move, max_move + 1)
+            mask_ = torch.tensor(_get_smoothborder_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+
+            img[:, frame_idx] = mask_ * cifar_patch.to(img.device) + (1 - mask_) * img[:, frame_idx]
+            mask.append(mask_)
+
+            cx = min(max(0, cx + delta_x), w)
+            cy = min(max(0, cy + delta_y), h)
+
+        mask = torch.stack(mask, dim=0)
+
+    return img, mask
+
+
+
+def create_pseudoanomaly_cifar_cutmix(img, cifar_img, max_size, h, w, dataset, max_move=0):
+    assert 0 <= max_size <= 1
+
+    pil_img = transforms.ToPILImage()(cifar_img)
+    pil_img = transforms.Grayscale(num_output_channels=1)(pil_img)
+    cifar_img = transforms.ToTensor()(pil_img)
+
+    cifar_img = transforms.Normalize(mean=[0.5], std=[0.5])(cifar_img)
+
+    cifar_patch = F.interpolate(cifar_img.unsqueeze(0), size=(h, w), mode='bilinear', align_corners=True)
+
+    cx, cy = np.random.randint(w), np.random.randint(h)
+
+    cut_w= max(10, int(np.random.uniform(high=max_size) * w))
+    cut_h = max(10, int(np.random.uniform(high=max_size) * h))
+    if max_move == 0:
+        mask = torch.tensor(_get_cutmix_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+        img = mask * cifar_patch.to(img.device) + (1-mask) * img
+
+    else:
+        mask = []
+        for frame_idx in range(img.size(1)):
+            delta_x = np.random.randint(-max_move, max_move + 1)
+            delta_y = np.random.randint(-max_move, max_move + 1)
+            mask_ = torch.tensor(_get_cutmix_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+
+            img[:, frame_idx] = mask_ * cifar_patch.to(img.device) + (1 - mask_) * img[:, frame_idx]
+            mask.append(mask_)
+
+            cx = min(max(0, cx + delta_x), w)
+            cy = min(max(0, cy + delta_y), h)
+
+        mask = torch.stack(mask, dim=0)
+
+    return img, mask
+
+
+
+def create_pseudoanomaly_cifar_mixupcutmix(img, cifar_img, max_size, h, w, dataset, max_move=0):
+    assert 0 <= max_size <= 1
+
+    pil_img = transforms.ToPILImage()(cifar_img)
+    pil_img = transforms.Grayscale(num_output_channels=1)(pil_img)
+    cifar_img = transforms.ToTensor()(pil_img)
+
+    cifar_img = transforms.Normalize(mean=[0.5], std=[0.5])(cifar_img)
+
+    cifar_patch = F.interpolate(cifar_img.unsqueeze(0), size=(h, w), mode='bilinear', align_corners=True)
+
+    cx, cy = np.random.randint(w), np.random.randint(h)
+
+    cut_w= max(10, int(np.random.uniform(high=max_size) * w))
+    cut_h = max(10, int(np.random.uniform(high=max_size) * h))
+    if max_move == 0:
+        mask = torch.tensor(_get_cutmix_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+        img = mask * 0.5 * cifar_patch.to(img.device) + mask * 0.5 * img + (1-mask) * img
+
+    else:
+        mask = []
+        for frame_idx in range(img.size(1)):
+            delta_x = np.random.randint(-max_move, max_move + 1)
+            delta_y = np.random.randint(-max_move, max_move + 1)
+            mask_ = torch.tensor(_get_cutmix_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+
+            img[:, frame_idx] = mask_ * 0.5 * cifar_patch.to(img.device) + mask_ * 0.5 * img[:, frame_idx] + (1 - mask_) * img[:, frame_idx]
+            mask.append(mask_)
+
+            cx = min(max(0, cx + delta_x), w)
+            cy = min(max(0, cy + delta_y), h)
+
+        mask = torch.stack(mask, dim=0)
+
+    return img, mask
+
+
+
+def create_pseudoanomaly_seq_smoothborder(img, seq, max_size, h, w, dataset, max_move=0):
+    assert 0 <= max_size <= 1
+
+    cx, cy = np.random.randint(w), np.random.randint(h)
+
+    cut_w= max(10, int(np.random.uniform(high=max_size) * w))
+    cut_h = max(10, int(np.random.uniform(high=max_size) * h))
+    if max_move == 0:
+        mask = torch.tensor(_get_smoothborder_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+        img = mask * seq.to(img.device) + (1-mask) * img
+    else:
+        mask = []
+        for frame_idx in range(img.size(1)):
+            delta_x = np.random.randint(-max_move, max_move + 1)
+            delta_y = np.random.randint(-max_move, max_move + 1)
+            mask_ = torch.tensor(_get_smoothborder_mask(cx, cy, cut_h, cut_w, h, w)).to(img.device).float()
+
+            img[:, frame_idx] = mask_ * seq[:, frame_idx].to(img.device) + (1 - mask_) * img[:, frame_idx]
+            mask.append(mask_)
+
+            cx = min(max(0, cx + delta_x), w)
+            cy = min(max(0, cy + delta_y), h)
+
+        mask = torch.stack(mask, dim=0)
+
+    return img, mask
+
+
+
+
+
+def _get_gaussian_mask(x_mu, y_mu, x_sigma, y_sigma, h, w):
+    x, y = np.arange(w), np.arange(h)
+
+    # x_mu, y_mu = random.randint(0, w), random.randint(0, h)
+    # x_sigma = max(10, int(np.random.uniform(high=max_size) * w))
+    # y_sigma = max(10, int(np.random.uniform(high=max_size) * h))
+
+    gx = np.exp(-(x - x_mu) ** 2 / (2 * x_sigma ** 2))
+    gy = np.exp(-(y - y_mu) ** 2 / (2 * y_sigma ** 2))
+    g = np.outer(gx, gy)
+    # g /= np.sum(g)  # normalize, if you want that
+
+    # sum_g = np.sum(g)
+    # lam = sum_g / (w * h)
+    # print(lam)
+
+    # plt.imshow(g, interpolation="nearest", origin="lower")
+    # plt.show()
+    # g = np.dstack([g, g, g])
+
+    return g
+
+# a = _get_gaussian_mask(0.5, 256, 256)
+
+def _get_smoothborder_mask(cx, cy, Cut_h, Cut_w, h, w):
+    lam = np.random.beta(1, 1)
+    percentage = 0.1
+    cut_rat = np.sqrt(1. - lam)
+
+    # Cut_w = min(np.int(max_size*w), max(2, np.int(w * cut_rat)))
+    # Cut_h = min(np.int(max_size*h), max(2, np.int(h * cut_rat)))
+
+    # cx, cy = np.random.randint(w), np.random.randint(h)
+
+    bbx1 = np.clip(cx - Cut_w // 2, 0, w)  # top left x
+    bby1 = np.clip(cy - Cut_h // 2, 0, h)  # top left y
+    bbx2 = np.clip(cx + Cut_w // 2, 0, w)  # bottom right x
+    bby2 = np.clip(cy + Cut_h // 2, 0, h)  # bottom right y
+
+    img = np.zeros((w, h))
+    img2, img3 = np.ones_like(img), np.ones_like(img)
+    img[bbx1:bbx2, bby1:bby2] = img2[bbx1:bbx2, bby1:bby2]
+
+    lo = bbx1 - (Cut_w // 2) * percentage  # left side: beginning linear from 0
+    li = bbx1  # + (Cut_w // 2) * percentage  # left side: end of linear at 1
+    ri = bbx2  # - (Cut_w // 2) * percentage  # right : start linear from 1
+    ro = bbx2 + (Cut_w // 2) * percentage  # right: end linear at 0
+
+    to = bby1 - (Cut_h // 2) * percentage  # top: start linear from 0
+    ti = bby1  # + (Cut_h // 2) * percentage  # top: end linear at 1
+    bi = bby2  # - (Cut_h // 2) * percentage  # bottom: start linear from 1
+    bo = bby2 + (Cut_h // 2) * percentage  # bottom: end linear at 0
+
+    # glx = lambda x: ((x - lo) / (li - lo))
+    # grx = lambda x: (-(x - ro) / (ro - ri))
+    # gtx = lambda x: ((x - to) / (ti - to))
+    # gbx = lambda x: (-(x - bo) / (bo - bi))
+
+    for i in range(w):
+        for j in range(h):
+            if i < cx:
+                img2[j][i] = ((i - lo) / (li - lo))  # linear going up
+            else:
+                img2[j][i] = (-(i - ro) / (ro - ri))  # linear going down
+            if j < cy:
+                img3[j][i] = ((j - to) / (ti - to))
+            else:
+                img3[j][i] = (-(j - bo) / (bo - bi))
+
+    img2[img2 < 0] = 0
+    img2[img2 > 1] = 1
+
+    img3[img3 < 0] = 0
+    img3[img3 > 1] = 1
+
+    # plt.figure()
+    # plt.subplot(131)
+    # plt.imshow(img2)
+    # # plt.show()
+    # plt.subplot(132)
+    # plt.imshow(img3)
+    # # plt.show()
+    img4 = np.multiply(img2, img3)
+    # sum_img4 = np.sum(img4)
+    # lam = sum_img4 / (w * h)
+
+    # plt.subplot(133)
+    # plt.imshow(img4)
+    # plt.show()
+    return img4  #, lam
+
+# a = _get_smoothborder_mask(0.5, 256, 256)
+
+
+def _get_cutmix_mask(cx, cy, Cut_h, Cut_w, h, w):
+    lam = np.random.beta(1, 1)
+
+    bbx1 = np.clip(cx - Cut_w // 2, 0, w)  # top left x
+    bby1 = np.clip(cy - Cut_h // 2, 0, h)  # top left y
+    bbx2 = np.clip(bbx1 + Cut_w, 0, w)  # bottom right x
+    bby2 = np.clip(bby1 + Cut_h, 0, h)  # bottom right y
+
+    img = np.zeros((w, h))
+    img2 = np.ones_like(img)
+    img[bby1:bby2, bbx1:bbx2] = img2[bby1:bby2, bbx1:bbx2]
+
+    return img  #, lam
+
+# a = _get_cutmix_mask(100, 100, 15, 30, 256, 256)
+# a = _get_smoothborder_mask(100, 100, 15, 30, 256, 256)

IPAD/model/reconstruction_model.py

@@ -0,0 +1,104 @@
+import torch.nn as nn
+from functools import reduce
+from operator import mul
+import torch
+
+class Reconstruction3DEncoder(nn.Module):
+    def __init__(self, chnum_in):
+        super(Reconstruction3DEncoder, self).__init__()
+
+        # Dong Gong's paper code
+        self.chnum_in = chnum_in
+        feature_num = 128
+        feature_num_2 = 96
+        feature_num_x2 = 256
+        self.encoder = nn.Sequential(
+            nn.Conv3d(self.chnum_in, feature_num_2, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num_2, feature_num, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Conv3d(feature_num_x2, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True)
+        )
+
+    def forward(self, x):
+        x = self.encoder(x)
+        return x
+
+
+class Reconstruction3DDecoder(nn.Module):
+    def __init__(self, chnum_in):
+        super(Reconstruction3DDecoder, self).__init__()
+
+        # Dong Gong's paper code + Tanh
+        self.chnum_in = chnum_in
+        feature_num = 128
+        feature_num_2 = 96
+        feature_num_x2 = 256
+        
+        self.decoder = nn.Sequential(
+            nn.ConvTranspose3d(feature_num_x2, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_x2, feature_num, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num, feature_num_2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_2, self.chnum_in, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1),
+                               output_padding=(0, 1, 1)),
+            nn.Tanh()
+        )
+
+    def forward(self, x):
+        x = self.decoder(x)
+        return x
+
+
+class VST3DDecoder(nn.Module):
+    def __init__(self, chnum_out):
+        super(VST3DDecoder, self).__init__()
+
+        # Dong Gong's paper code + Tanh
+        self.chnum_out = chnum_out
+        feature_num = 128
+        feature_num_2 = 96
+        feature_num_x2 = 256
+        feature_num_in = 768
+        self.transformer_decoder = nn.Sequential(
+            # (4,768,4,8,8)
+            nn.ConvTranspose3d(feature_num_in, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            # (4,256,4,16,16)
+            nn.ConvTranspose3d(feature_num_x2, feature_num_x2, (3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1),
+                               output_padding=(1, 1, 1)),
+            nn.BatchNorm3d(feature_num_x2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_x2, feature_num, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1),
+                               output_padding=(0, 1, 1)),
+            nn.BatchNorm3d(feature_num),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num, feature_num_2, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1),
+                               output_padding=(0, 1, 1)),
+            nn.BatchNorm3d(feature_num_2),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.ConvTranspose3d(feature_num_2, self.chnum_out, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1),
+                               output_padding=(0, 1, 1)),
+            nn.Tanh()
+        )
+
+    def forward(self, x):
+        x = self.transformer_decoder(x)
+        return x

IPAD/model/utils.py

@@ -0,0 +1,142 @@
+import numpy as np
+from collections import OrderedDict
+import os
+import glob
+import cv2
+import torch.utils.data as data
+import random
+from PIL import Image
+
+
+rng = np.random.RandomState(2020)
+
+def np_load_frame(filename, resize_height, resize_width, grayscale=False):
+    grayscale = False
+    """
+    Load image path and convert it to numpy.ndarray. Notes that the color channels are BGR and the color space
+    is normalized from [0, 255] to [-1, 1].
+
+    :param filename: the full path of image
+    :param resize_height: resized height
+    :param resize_width: resized width
+    :return: numpy.ndarray
+    """
+    if grayscale:
+        image_decoded = cv2.imread(filename, cv2.IMREAD_GRAYSCALE)
+    else:
+        image_decoded = cv2.imread(filename)
+    image_resized = cv2.resize(image_decoded, (resize_width, resize_height))
+    # image_resized = np.copy(image_decoded)
+    image_resized = image_resized.astype(dtype=np.float32)
+    image_resized = (image_resized / 127.5) - 1.0
+    return image_resized
+
+
+
+class Reconstruction3DDataLoader(data.Dataset):
+    def __init__(self, video_folder, transform, resize_height, resize_width, num_frames=16,
+                 img_extension='.jpg', dataset='ped2', jump=[2], hold=[2], return_normal_seq=False):
+        self.dir = video_folder
+        self.transform = transform
+        self.videos = OrderedDict()
+        self._resize_height = resize_height
+        self._resize_width = resize_width
+        self._num_frames = num_frames
+
+        self.extension = img_extension
+        self.dataset = dataset
+
+        self.setup()
+        self.samples, self.background_models = self.get_all_samples()
+
+        self.jump = jump
+        self.hold = hold
+        self.return_normal_seq = return_normal_seq  # for fast and slow moving
+
+    def setup(self):
+        videos = glob.glob(os.path.join(self.dir, '*/'))
+        for video in sorted(videos):
+            print(video)
+            video_name = video.split('/')[-2]
+            self.videos[video_name] = {}
+            self.videos[video_name]['path'] = video
+            self.videos[video_name]['frame'] = glob.glob(os.path.join(video, '*' + self.extension))
+            self.videos[video_name]['frame'].sort()
+            self.videos[video_name]['length'] = len(self.videos[video_name]['frame'])
+
+    def get_all_samples(self):
+        frames = []
+        background_models = []
+        videos = glob.glob(os.path.join(self.dir, '*/'))
+        for video in sorted(videos):
+            video_name = video.split('/')[-2]
+
+            for i in range(len(self.videos[video_name]['frame']) - self._num_frames + 1):
+                frames.append(self.videos[video_name]['frame'][i])
+                # background_models.append(bg_filename)
+
+        return frames, background_models
+
+    def __getitem__(self, index):
+        # index = 8
+        video_name = self.samples[index].split('/')[-2]
+        if self.dataset == 'shanghai' and 'training' in self.samples[index]:
+            frame_name = int(self.samples[index].split('/')[-1].split('.')[-2]) - 1
+        else:
+            frame_name = int(self.samples[index].split('/')[-1].split('.')[-2])
+
+        batch = []
+        for i in range(self._num_frames):
+            image = np_load_frame(self.videos[video_name]['frame'][frame_name + i], self._resize_height,
+                                  self._resize_width, grayscale=True)
+            if self.transform is not None:
+                batch.append(self.transform(image))
+        # batch:len=16 ,batch[0]:torch(3,256,256)
+        img = OrderedDict()
+        img['batch'] = np.stack(batch, axis=1)
+        img['index'] = frame_name*200//len(self.videos[video_name]['frame'])
+        # return np.stack(batch, axis=1)
+        return img
+
+    def __len__(self):
+        return len(self.samples)
+
+
+class Reconstruction3DDataLoaderJump(Reconstruction3DDataLoader):
+    def __getitem__(self, index):
+        # index = 8
+        video_name = self.samples[index].split('/')[-2]
+        if self.dataset == 'shanghai' and 'training' in self.samples[index]:  # bcos my shanghai's start from 1
+            frame_name = int(self.samples[index].split('/')[-1].split('.')[-2]) - 1
+        else:
+            frame_name = int(self.samples[index].split('/')[-1].split('.')[-2])
+
+        batch = []
+        normal_batch = []
+        jump = random.choice(self.jump)
+
+        retry = 0
+        while len(self.videos[video_name]['frame']) < frame_name + (self._num_frames-1) * jump and retry < 10:
+            # reselect the frame_name
+            frame_name = np.random.randint(len(self.videos[video_name]['frame']))
+            retry += 1
+
+        for i in range(self._num_frames):
+            image = np_load_frame(self.videos[video_name]['frame'][min(frame_name + i*jump, len(self.videos[video_name]['frame'])-1)], self._resize_height,
+                                  self._resize_width, grayscale=True)
+
+            if self.transform is not None:
+                batch.append(self.transform(image))
+
+        if self.return_normal_seq:
+            for i in range(self._num_frames):
+                image = np_load_frame(self.videos[video_name]['frame'][min(frame_name + i, len(self.videos[video_name]['frame'])-1)], self._resize_height,
+                                      self._resize_width, grayscale=True)
+
+                if self.transform is not None:
+                    normal_batch.append(self.transform(image))
+            return np.stack(batch, axis=1), np.stack(normal_batch, axis=1)
+
+        else:
+            return np.stack(batch, axis=1), normal_batch
+

IPAD/model/video_swin_transformer.py

@@ -0,0 +1,48 @@
+import torch
+from .reconstruction_model import Reconstruction3DEncoder, Reconstruction3DDecoder, VST3DDecoder
+from .VST_block import SwinTransformer3D
+from einops import rearrange
+from model import MemModule
+import torch.nn as nn
+from torch.nn import functional as F
+
+class VST(torch.nn.Module):
+    def __init__(self, mem_dim=2000, shrink_thres=0.0025):  # for reconstruction
+        super(VST, self).__init__()
+        self.reconstruction = True
+        # self.chnum_in = chnum_in
+
+        # self.encoder = Reconstruction3DEncoder(chnum_in=1)  # black and white
+        # self.decoder = Reconstruction3DDecoder(chnum_in=1)  # black and white
+        self.transformer_encoder = SwinTransformer3D()
+        self.mem_rep = MemModule(mem_dim=mem_dim, fea_dim=768, shrink_thres=shrink_thres)
+        self.period = nn.Sequential(
+            nn.Conv3d(768, 768, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(768),
+            nn.LeakyReLU(0.2, inplace=True),
+            # (batch_size,256,4,4,4)
+            nn.Flatten(1),
+            nn.Linear(768*4*4*4,4096),
+            nn.ReLU(),
+            nn.Linear(4096,2048),
+            nn.ReLU(),
+            nn.Linear(2048,200),
+        )
+        self.transformer_decoder = VST3DDecoder(chnum_out=3)
+        # self.encoder = Reconstruction3DEncoder(chnum_in=3)  # RGB
+        # self.decoder = Reconstruction3DDecoder(chnum_in=3)  # RGB
+
+    def forward(self, x):
+        
+        feature = self.transformer_encoder(x)
+        #feature (batch_size,768,4,8,8)
+        recon_index = self.period(feature)
+        # print(recon_index[0])
+        res_mem = self.mem_rep(feature, recon_index)
+        feature = res_mem['output']
+        att = res_mem['att']
+        output = self.transformer_decoder(feature.clone())
+
+        return {'output': output, 'att': att, 'recon_index': recon_index}
+
+