Refactor swin transfomer so later we can reuse component for 3d version (#6088) (#6100)

* Use List[int] instead of int for window_size and shift_size

* Make PatchMerging and SwinTransformerBlock able to handle 2d and 3d cases

* Separate patch embedding from SwinTransformer and enable to get model without head by specifying num_heads=None

* Dont use if before padding so it is fx friendly

* Put the handling on window_size edge cases on separate function and wrap with torch.fx.wrap so it is excluded from tracing

* Update the weight url to the converted weight with new structure

* Update the accuracy of swin_transformer

* Change assert to Exception and nit

* Make num_classes optional

* Add typing output for _fix_window_and_shift_size function

* init head to None to make it jit scriptable

* Revert the change to make num_classes optional

* Revert unneccesarry changes that might be risky

* Remove self.head declaration

Author: YosuaMichael

torchvision/models/swin_transformer.py

@@ -39,18 +39,23 @@ def __init__(self, dim: int, norm_layer: Callable[..., nn.Module] = nn.LayerNorm
         self.norm = norm_layer(4 * dim)
 
     def forward(self, x: Tensor):
-        B, H, W, C = x.shape
-
-        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
+        """
+        Args:
+            x (Tensor): input tensor with expected layout of [..., H, W, C]
+        Returns:
+            Tensor with layout of [..., H/2, W/2, 2*C]
+        """
+        H, W, _ = x.shape[-3:]
+        x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[..., 0::2, 0::2, :]  # ... H/2 W/2 C
+        x1 = x[..., 1::2, 0::2, :]  # ... H/2 W/2 C
+        x2 = x[..., 0::2, 1::2, :]  # ... H/2 W/2 C
+        x3 = x[..., 1::2, 1::2, :]  # ... H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # ... H/2 W/2 4*C
 
         x = self.norm(x)
-        x = self.reduction(x)
-        x = x.view(B, H // 2, W // 2, 2 * C)
+        x = self.reduction(x)  # ... H/2 W/2 2*C
         return x
 
 
@@ -59,9 +64,9 @@ def shifted_window_attention(
     qkv_weight: Tensor,
     proj_weight: Tensor,
     relative_position_bias: Tensor,
-    window_size: int,
+    window_size: List[int],
     num_heads: int,
-    shift_size: int = 0,
+    shift_size: List[int],
     attention_dropout: float = 0.0,
     dropout: float = 0.0,
     qkv_bias: Optional[Tensor] = None,
@@ -75,9 +80,9 @@ def shifted_window_attention(
         qkv_weight (Tensor[in_dim, out_dim]): The weight tensor of query, key, value.
         proj_weight (Tensor[out_dim, out_dim]): The weight tensor of projection.
         relative_position_bias (Tensor): The learned relative position bias added to attention.
-        window_size (int): Window size.
+        window_size (List[int]): Window size.
         num_heads (int): Number of attention heads.
-        shift_size (int): Shift size for shifted window attention. Default: 0.
+        shift_size (List[int]): Shift size for shifted window attention.
         attention_dropout (float): Dropout ratio of attention weight. Default: 0.0.
         dropout (float): Dropout ratio of output. Default: 0.0.
         qkv_bias (Tensor[out_dim], optional): The bias tensor of query, key, value. Default: None.
@@ -87,23 +92,25 @@ def shifted_window_attention(
     """
     B, H, W, C = input.shape
     # pad feature maps to multiples of window size
-    pad_r = (window_size - W % window_size) % window_size
-    pad_b = (window_size - H % window_size) % window_size
+    pad_r = (window_size[1] - W % window_size[1]) % window_size[1]
+    pad_b = (window_size[0] - H % window_size[0]) % window_size[0]
     x = F.pad(input, (0, 0, 0, pad_r, 0, pad_b))
     _, pad_H, pad_W, _ = x.shape
 
-    # If window size is larger than feature size, there is no need to shift window.
-    if window_size == min(pad_H, pad_W):
-        shift_size = 0
+    # If window size is larger than feature size, there is no need to shift window
+    if window_size[0] >= pad_H:
+        shift_size[0] = 0
+    if window_size[1] >= pad_W:
+        shift_size[1] = 0
 
     # cyclic shift
-    if shift_size > 0:
-        x = torch.roll(x, shifts=(-shift_size, -shift_size), dims=(1, 2))
+    if sum(shift_size) > 0:
+        x = torch.roll(x, shifts=(-shift_size[0], -shift_size[1]), dims=(1, 2))
 
     # partition windows
-    num_windows = (pad_H // window_size) * (pad_W // window_size)
-    x = x.view(B, pad_H // window_size, window_size, pad_W // window_size, window_size, C)
-    x = x.permute(0, 1, 3, 2, 4, 5).reshape(B * num_windows, window_size * window_size, C)  # B*nW, Ws*Ws, C
+    num_windows = (pad_H // window_size[0]) * (pad_W // window_size[1])
+    x = x.view(B, pad_H // window_size[0], window_size[0], pad_W // window_size[1], window_size[1], C)
+    x = x.permute(0, 1, 3, 2, 4, 5).reshape(B * num_windows, window_size[0] * window_size[1], C)  # B*nW, Ws*Ws, C
 
     # multi-head attention
     qkv = F.linear(x, qkv_weight, qkv_bias)
@@ -114,17 +121,18 @@ def shifted_window_attention(
     # add relative position bias
     attn = attn + relative_position_bias
 
-    if shift_size > 0:
+    if sum(shift_size) > 0:
         # generate attention mask
         attn_mask = x.new_zeros((pad_H, pad_W))
-        slices = ((0, -window_size), (-window_size, -shift_size), (-shift_size, None))
+        h_slices = ((0, -window_size[0]), (-window_size[0], -shift_size[0]), (-shift_size[0], None))
+        w_slices = ((0, -window_size[1]), (-window_size[1], -shift_size[1]), (-shift_size[1], None))
         count = 0
-        for h in slices:
-            for w in slices:
+        for h in h_slices:
+            for w in w_slices:
                 attn_mask[h[0] : h[1], w[0] : w[1]] = count
                 count += 1
-        attn_mask = attn_mask.view(pad_H // window_size, window_size, pad_W // window_size, window_size)
-        attn_mask = attn_mask.permute(0, 2, 1, 3).reshape(num_windows, window_size * window_size)
+        attn_mask = attn_mask.view(pad_H // window_size[0], window_size[0], pad_W // window_size[1], window_size[1])
+        attn_mask = attn_mask.permute(0, 2, 1, 3).reshape(num_windows, window_size[0] * window_size[1])
         attn_mask = attn_mask.unsqueeze(1) - attn_mask.unsqueeze(2)
         attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
         attn = attn.view(x.size(0) // num_windows, num_windows, num_heads, x.size(1), x.size(1))
@@ -139,12 +147,12 @@ def shifted_window_attention(
     x = F.dropout(x, p=dropout)
 
     # reverse windows
-    x = x.view(B, pad_H // window_size, pad_W // window_size, window_size, window_size, C)
+    x = x.view(B, pad_H // window_size[0], pad_W // window_size[1], window_size[0], window_size[1], C)
     x = x.permute(0, 1, 3, 2, 4, 5).reshape(B, pad_H, pad_W, C)
 
     # reverse cyclic shift
-    if shift_size > 0:
-        x = torch.roll(x, shifts=(shift_size, shift_size), dims=(1, 2))
+    if sum(shift_size) > 0:
+        x = torch.roll(x, shifts=(shift_size[0], shift_size[1]), dims=(1, 2))
 
     # unpad features
     x = x[:, :H, :W, :].contiguous()
@@ -162,15 +170,17 @@ class ShiftedWindowAttention(nn.Module):
     def __init__(
         self,
         dim: int,
-        window_size: int,
-        shift_size: int,
+        window_size: List[int],
+        shift_size: List[int],
         num_heads: int,
         qkv_bias: bool = True,
         proj_bias: bool = True,
         attention_dropout: float = 0.0,
         dropout: float = 0.0,
     ):
         super().__init__()
+        if len(window_size) != 2 or len(shift_size) != 2:
+            raise ValueError("window_size and shift_size must be of length 2")
         self.window_size = window_size
         self.shift_size = shift_size
         self.num_heads = num_heads
@@ -182,29 +192,35 @@ def __init__(
 
         # define a parameter table of relative position bias
         self.relative_position_bias_table = nn.Parameter(
-            torch.zeros((2 * window_size - 1) * (2 * window_size - 1), num_heads)
+            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)
-        coords_w = torch.arange(self.window_size)
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
         coords = torch.stack(torch.meshgrid(coords_h, coords_w, indexing="ij"))  # 2, Wh, Ww
         coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
         relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
         relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
-        relative_coords[:, :, 0] += self.window_size - 1  # shift to start from 0
-        relative_coords[:, :, 1] += self.window_size - 1
-        relative_coords[:, :, 0] *= 2 * self.window_size - 1
+        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).view(-1)  # Wh*Ww*Wh*Ww
         self.register_buffer("relative_position_index", relative_position_index)
 
         nn.init.trunc_normal_(self.relative_position_bias_table, std=0.02)
 
     def forward(self, x: Tensor):
+        """
+        Args:
+            x (Tensor): Tensor with layout of [B, H, W, C]
+        Returns:
+            Tensor with same layout as input, i.e. [B, H, W, C]
+        """
+
+        N = self.window_size[0] * self.window_size[1]
         relative_position_bias = self.relative_position_bias_table[self.relative_position_index]  # type: ignore[index]
-        relative_position_bias = relative_position_bias.view(
-            self.window_size * self.window_size, self.window_size * self.window_size, -1
-        )
+        relative_position_bias = relative_position_bias.view(N, N, -1)
         relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous().unsqueeze(0)
 
         return shifted_window_attention(
@@ -228,31 +244,33 @@ class SwinTransformerBlock(nn.Module):
     Args:
         dim (int): Number of input channels.
         num_heads (int): Number of attention heads.
-        window_size (int): Window size. Default: 7.
-        shift_size (int): Shift size for shifted window attention. Default: 0.
+        window_size (List[int]): Window size.
+        shift_size (List[int]): Shift size for shifted window attention.
         mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.0.
         dropout (float): Dropout rate. Default: 0.0.
         attention_dropout (float): Attention dropout rate. Default: 0.0.
         stochastic_depth_prob: (float): Stochastic depth rate. Default: 0.0.
         norm_layer (nn.Module): Normalization layer.  Default: nn.LayerNorm.
+        attn_layer (nn.Module): Attention layer. Default: ShiftedWindowAttention
     """
 
     def __init__(
         self,
         dim: int,
         num_heads: int,
-        window_size: int = 7,
-        shift_size: int = 0,
+        window_size: List[int],
+        shift_size: List[int],
         mlp_ratio: float = 4.0,
         dropout: float = 0.0,
         attention_dropout: float = 0.0,
         stochastic_depth_prob: float = 0.0,
         norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
+        attn_layer: Callable[..., nn.Module] = ShiftedWindowAttention,
     ):
         super().__init__()
 
         self.norm1 = norm_layer(dim)
-        self.attn = ShiftedWindowAttention(
+        self.attn = attn_layer(
             dim,
             window_size,
             shift_size,
@@ -281,11 +299,11 @@ class SwinTransformer(nn.Module):
     Implements Swin Transformer from the `"Swin Transformer: Hierarchical Vision Transformer using
     Shifted Windows" <https://arxiv.org/pdf/2103.14030>`_ paper.
     Args:
-        patch_size (int): Patch size.
+        patch_size (List[int]): Patch size.
         embed_dim (int): Patch embedding dimension.
         depths (List(int)): Depth of each Swin Transformer layer.
         num_heads (List(int)): Number of attention heads in different layers.
-        window_size (int): Window size. Default: 7.
+        window_size (List[int]): Window size.
         mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.0.
         dropout (float): Dropout rate. Default: 0.0.
         attention_dropout (float): Attention dropout rate. Default: 0.0.
@@ -297,11 +315,11 @@ class SwinTransformer(nn.Module):
 
     def __init__(
         self,
-        patch_size: int,
+        patch_size: List[int],
         embed_dim: int,
         depths: List[int],
         num_heads: List[int],
-        window_size: int = 7,
+        window_size: List[int],
         mlp_ratio: float = 4.0,
         dropout: float = 0.0,
         attention_dropout: float = 0.0,
@@ -324,7 +342,9 @@ def __init__(
         # split image into non-overlapping patches
         layers.append(
             nn.Sequential(
-                nn.Conv2d(3, embed_dim, kernel_size=patch_size, stride=patch_size),
+                nn.Conv2d(
+                    3, embed_dim, kernel_size=(patch_size[0], patch_size[1]), stride=(patch_size[0], patch_size[1])
+                ),
                 Permute([0, 2, 3, 1]),
                 norm_layer(embed_dim),
             )
@@ -344,7 +364,7 @@ def __init__(
                         dim,
                         num_heads[i_stage],
                         window_size=window_size,
-                        shift_size=0 if i_layer % 2 == 0 else window_size // 2,
+                        shift_size=[0 if i_layer % 2 == 0 else w // 2 for w in window_size],
                         mlp_ratio=mlp_ratio,
                         dropout=dropout,
                         attention_dropout=attention_dropout,
@@ -381,11 +401,11 @@ def forward(self, x):
 
 
 def _swin_transformer(
-    patch_size: int,
+    patch_size: List[int],
     embed_dim: int,
     depths: List[int],
     num_heads: List[int],
-    window_size: int,
+    window_size: List[int],
     stochastic_depth_prob: float,
     weights: Optional[WeightsEnum],
     progress: bool,
@@ -508,11 +528,11 @@ def swin_t(*, weights: Optional[Swin_T_Weights] = None, progress: bool = True, *
     weights = Swin_T_Weights.verify(weights)
 
     return _swin_transformer(
-        patch_size=4,
+        patch_size=[4, 4],
         embed_dim=96,
         depths=[2, 2, 6, 2],
         num_heads=[3, 6, 12, 24],
-        window_size=7,
+        window_size=[7, 7],
         stochastic_depth_prob=0.2,
         weights=weights,
         progress=progress,
@@ -544,11 +564,11 @@ def swin_s(*, weights: Optional[Swin_S_Weights] = None, progress: bool = True, *
     weights = Swin_S_Weights.verify(weights)
 
     return _swin_transformer(
-        patch_size=4,
+        patch_size=[4, 4],
         embed_dim=96,
         depths=[2, 2, 18, 2],
         num_heads=[3, 6, 12, 24],
-        window_size=7,
+        window_size=[7, 7],
         stochastic_depth_prob=0.3,
         weights=weights,
         progress=progress,
@@ -580,11 +600,11 @@ def swin_b(*, weights: Optional[Swin_B_Weights] = None, progress: bool = True, *
     weights = Swin_B_Weights.verify(weights)
 
     return _swin_transformer(
-        patch_size=4,
+        patch_size=[4, 4],
         embed_dim=128,
         depths=[2, 2, 18, 2],
         num_heads=[4, 8, 16, 32],
-        window_size=7,
+        window_size=[7, 7],
         stochastic_depth_prob=0.5,
         weights=weights,
         progress=progress,