Add include_fc and use_combined_linear argument in the SABlock (#…

…7996)

Fixes #7991
Fixes #7992

### Description
Add `include_fc` and `use_combined_linear` argument in the `SABlock`.


### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [ ] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: YunLiu <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Kerfoot <[email protected]>

monai/networks/blocks/crossattention.py

@@ -59,13 +59,12 @@ def __init__(
             causal (bool, optional): whether to use causal attention.
             sequence_length (int, optional): if causal is True, it is necessary to specify the sequence length.
             rel_pos_embedding (str, optional): Add relative positional embeddings to the attention map. For now only
-            "decomposed" is supported (see https://arxiv.org/abs/2112.01526). 2D and 3D are supported.
+                "decomposed" is supported (see https://arxiv.org/abs/2112.01526). 2D and 3D are supported.
             input_size (tuple(spatial_dim), optional): Input resolution for calculating the relative positional
-            parameter size.
+                parameter size.
             attention_dtype: cast attention operations to this dtype.
-            use_flash_attention: if True, use Pytorch's inbuilt
-            flash attention for a memory efficient attention mechanism (see
-            https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
+            use_flash_attention: if True, use Pytorch's inbuilt flash attention for a memory efficient attention mechanism
+                (see https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
         """
 
         super().__init__()
@@ -109,7 +108,7 @@ def __init__(
         self.to_v = nn.Linear(self.context_input_size, inner_size, bias=qkv_bias)
         self.input_rearrange = Rearrange("b h (l d) -> b l h d", l=num_heads)
 
-        self.out_rearrange = Rearrange("b h l d -> b l (h d)")
+        self.out_rearrange = Rearrange("b l h d -> b h (l d)")
         self.drop_output = nn.Dropout(dropout_rate)
         self.drop_weights = nn.Dropout(dropout_rate)
         self.dropout_rate = dropout_rate
@@ -152,31 +151,20 @@ def forward(self, x: torch.Tensor, context: Optional[torch.Tensor] = None):
         # calculate query, key, values for all heads in batch and move head forward to be the batch dim
         b, t, c = x.size()  # batch size, sequence length, embedding dimensionality (hidden_size)
 
-        q = self.to_q(x)
+        q = self.input_rearrange(self.to_q(x))
         kv = context if context is not None else x
         _, kv_t, _ = kv.size()
-        k = self.to_k(kv)
-        v = self.to_v(kv)
+        k = self.input_rearrange(self.to_k(kv))
+        v = self.input_rearrange(self.to_v(kv))
 
         if self.attention_dtype is not None:
             q = q.to(self.attention_dtype)
             k = k.to(self.attention_dtype)
 
-        q = q.view(b, t, self.num_heads, c // self.num_heads).transpose(1, 2)  # (b, nh, t,  hs) #
-        k = k.view(b, kv_t, self.num_heads, c // self.num_heads).transpose(1, 2)  # (b, nh, kv_t, hs)
-        v = v.view(b, kv_t, self.num_heads, c // self.num_heads).transpose(1, 2)  # (b, nh, kv_t, hs)
-
         if self.use_flash_attention:
             x = torch.nn.functional.scaled_dot_product_attention(
-                query=q.transpose(1, 2),
-                key=k.transpose(1, 2),
-                value=v.transpose(1, 2),
-                scale=self.scale,
-                dropout_p=self.dropout_rate,
-                is_causal=self.causal,
-            ).transpose(
-                1, 2
-            )  # Back to (b, nh, t, hs)
+                query=q, key=k, value=v, scale=self.scale, dropout_p=self.dropout_rate, is_causal=self.causal
+            )
         else:
             att_mat = torch.einsum("blxd,blyd->blxy", q, k) * self.scale
             # apply relative positional embedding if defined
@@ -195,6 +183,7 @@ def forward(self, x: torch.Tensor, context: Optional[torch.Tensor] = None):
 
             att_mat = self.drop_weights(att_mat)
             x = torch.einsum("bhxy,bhyd->bhxd", att_mat, v)
+
         x = self.out_rearrange(x)
         x = self.out_proj(x)
         x = self.drop_output(x)

monai/networks/blocks/selfattention.py

@@ -11,7 +11,7 @@
 
 from __future__ import annotations
 
-from typing import Optional, Tuple
+from typing import Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -40,9 +40,11 @@ def __init__(
         hidden_input_size: int | None = None,
         causal: bool = False,
         sequence_length: int | None = None,
-        rel_pos_embedding: Optional[str] = None,
-        input_size: Optional[Tuple] = None,
-        attention_dtype: Optional[torch.dtype] = None,
+        rel_pos_embedding: str | None = None,
+        input_size: Tuple | None = None,
+        attention_dtype: torch.dtype | None = None,
+        include_fc: bool = True,
+        use_combined_linear: bool = True,
         use_flash_attention: bool = False,
     ) -> None:
         """
@@ -61,9 +63,10 @@ def __init__(
             input_size (tuple(spatial_dim), optional): Input resolution for calculating the relative
                 positional parameter size.
             attention_dtype: cast attention operations to this dtype.
-            use_flash_attention: if True, use Pytorch's inbuilt
-            flash attention for a memory efficient attention mechanism (see
-            https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
+            include_fc: whether to include the final linear layer. Default to True.
+            use_combined_linear: whether to use a single linear layer for qkv projection, default to True.
+            use_flash_attention: if True, use Pytorch's inbuilt flash attention for a memory efficient attention mechanism
+                (see https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
 
         """
 
@@ -105,9 +108,22 @@ def __init__(
         self.hidden_input_size = hidden_input_size if hidden_input_size else hidden_size
         self.out_proj = nn.Linear(self.inner_dim, self.hidden_input_size)
 
-        self.qkv = nn.Linear(self.hidden_input_size, self.inner_dim * 3, bias=qkv_bias)
-        self.input_rearrange = Rearrange("b h (qkv l d) -> qkv b l h d", qkv=3, l=num_heads)
-        self.out_rearrange = Rearrange("b h l d -> b l (h d)")
+        self.qkv: Union[nn.Linear, nn.Identity]
+        self.to_q: Union[nn.Linear, nn.Identity]
+        self.to_k: Union[nn.Linear, nn.Identity]
+        self.to_v: Union[nn.Linear, nn.Identity]
+
+        if use_combined_linear:
+            self.qkv = nn.Linear(self.hidden_input_size, self.inner_dim * 3, bias=qkv_bias)
+            self.to_q = self.to_k = self.to_v = nn.Identity()  # add to enable torchscript
+            self.input_rearrange = Rearrange("b h (qkv l d) -> qkv b l h d", qkv=3, l=num_heads)
+        else:
+            self.to_q = nn.Linear(self.hidden_input_size, self.inner_dim, bias=qkv_bias)
+            self.to_k = nn.Linear(self.hidden_input_size, self.inner_dim, bias=qkv_bias)
+            self.to_v = nn.Linear(self.hidden_input_size, self.inner_dim, bias=qkv_bias)
+            self.qkv = nn.Identity()  # add to enable torchscript
+            self.input_rearrange = Rearrange("b h (l d) -> b l h d", l=num_heads)
+        self.out_rearrange = Rearrange("b l h d -> b h (l d)")
         self.drop_output = nn.Dropout(dropout_rate)
         self.drop_weights = nn.Dropout(dropout_rate)
         self.dropout_rate = dropout_rate
@@ -117,6 +133,8 @@ def __init__(
         self.attention_dtype = attention_dtype
         self.causal = causal
         self.sequence_length = sequence_length
+        self.include_fc = include_fc
+        self.use_combined_linear = use_combined_linear
         self.use_flash_attention = use_flash_attention
 
         if causal and sequence_length is not None:
@@ -144,22 +162,22 @@ def forward(self, x):
         Return:
             torch.Tensor: B x (s_dim_1 * ... * s_dim_n) x C
         """
-        output = self.input_rearrange(self.qkv(x))
-        q, k, v = output[0], output[1], output[2]
+        if self.use_combined_linear:
+            output = self.input_rearrange(self.qkv(x))
+            q, k, v = output[0], output[1], output[2]
+        else:
+            q = self.input_rearrange(self.to_q(x))
+            k = self.input_rearrange(self.to_k(x))
+            v = self.input_rearrange(self.to_v(x))
 
         if self.attention_dtype is not None:
             q = q.to(self.attention_dtype)
             k = k.to(self.attention_dtype)
 
         if self.use_flash_attention:
             x = F.scaled_dot_product_attention(
-                query=q.transpose(1, 2),
-                key=k.transpose(1, 2),
-                value=v.transpose(1, 2),
-                scale=self.scale,
-                dropout_p=self.dropout_rate,
-                is_causal=self.causal,
-            ).transpose(1, 2)
+                query=q, key=k, value=v, scale=self.scale, dropout_p=self.dropout_rate, is_causal=self.causal
+            )
         else:
             att_mat = torch.einsum("blxd,blyd->blxy", q, k) * self.scale
 
@@ -179,7 +197,9 @@ def forward(self, x):
 
             att_mat = self.drop_weights(att_mat)
             x = torch.einsum("bhxy,bhyd->bhxd", att_mat, v)
+
         x = self.out_rearrange(x)
-        x = self.out_proj(x)
+        if self.include_fc:
+            x = self.out_proj(x)
         x = self.drop_output(x)
         return x

monai/networks/blocks/spatialattention.py

@@ -32,8 +32,13 @@ class SpatialAttentionBlock(nn.Module):
         spatial_dims: number of spatial dimensions, could be 1, 2, or 3.
         num_channels: number of input channels. Must be divisible by num_head_channels.
         num_head_channels: number of channels per head.
+        norm_num_groups: Number of groups for the group norm layer.
+        norm_eps: Epsilon for the normalization.
         attention_dtype: cast attention operations to this dtype.
-        use_flash_attention: if True, use flash attention for a memory efficient attention mechanism.
+        include_fc: whether to include the final linear layer. Default to True.
+        use_combined_linear: whether to use a single linear layer for qkv projection, default to False.
+        use_flash_attention: if True, use Pytorch's inbuilt flash attention for a memory efficient attention mechanism
+            (see https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
 
     """
 
@@ -45,6 +50,8 @@ def __init__(
         norm_num_groups: int = 32,
         norm_eps: float = 1e-6,
         attention_dtype: Optional[torch.dtype] = None,
+        include_fc: bool = True,
+        use_combined_linear: bool = False,
         use_flash_attention: bool = False,
     ) -> None:
         super().__init__()
@@ -60,6 +67,8 @@ def __init__(
             num_heads=num_heads,
             qkv_bias=True,
             attention_dtype=attention_dtype,
+            include_fc=include_fc,
+            use_combined_linear=use_combined_linear,
             use_flash_attention=use_flash_attention,
         )
 

monai/networks/blocks/transformerblock.py

@@ -37,6 +37,8 @@ def __init__(
         sequence_length: int | None = None,
         with_cross_attention: bool = False,
         use_flash_attention: bool = False,
+        include_fc: bool = True,
+        use_combined_linear: bool = True,
     ) -> None:
         """
         Args:
@@ -47,7 +49,9 @@ def __init__(
             qkv_bias(bool, optional): apply bias term for the qkv linear layer. Defaults to False.
             save_attn (bool, optional): to make accessible the attention matrix. Defaults to False.
             use_flash_attention: if True, use Pytorch's inbuilt flash attention for a memory efficient attention mechanism
-             (see https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
+                (see https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html).
+            include_fc: whether to include the final linear layer. Default to True.
+            use_combined_linear: whether to use a single linear layer for qkv projection, default to True.
 
         """
 
@@ -69,6 +73,8 @@ def __init__(
             save_attn=save_attn,
             causal=causal,
             sequence_length=sequence_length,
+            include_fc=include_fc,
+            use_combined_linear=use_combined_linear,
             use_flash_attention=use_flash_attention,
         )
         self.norm2 = nn.LayerNorm(hidden_size)