feat(SABlock): access to the attn matrix

monai/networks/blocks/selfattention.py

@@ -25,13 +25,21 @@ class SABlock(nn.Module):
     An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>"
     """
 
-    def __init__(self, hidden_size: int, num_heads: int, dropout_rate: float = 0.0, qkv_bias: bool = False) -> None:
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        dropout_rate: float = 0.0,
+        qkv_bias: bool = False,
+        save_attn: bool = False,
+    ) -> None:
         """
         Args:
-            hidden_size: dimension of hidden layer.
-            num_heads: number of attention heads.
-            dropout_rate: faction of the input units to drop.
-            qkv_bias: bias term for the qkv linear layer.
+            hidden_size (int): dimension of hidden layer.
+            num_heads (int): number of attention heads.
+            dropout_rate (float, optional): faction of the input units to drop. Defaults to 0.0.
+            qkv_bias (bool, optional): bias term for the qkv linear layer. Defaults to False.
+            save_attn (bool, optional): to make accessible the attention matrix. Defaults to False.
 
         """
 
@@ -52,11 +60,16 @@ def __init__(self, hidden_size: int, num_heads: int, dropout_rate: float = 0.0,
         self.drop_weights = nn.Dropout(dropout_rate)
         self.head_dim = hidden_size // num_heads
         self.scale = self.head_dim**-0.5
+        self.save_attn = save_attn
 
     def forward(self, x):
         output = self.input_rearrange(self.qkv(x))
         q, k, v = output[0], output[1], output[2]
         att_mat = (torch.einsum("blxd,blyd->blxy", q, k) * self.scale).softmax(dim=-1)
+        if self.save_attn:
+            # no gradients and new tensor;
+            # https://pytorch.org/docs/stable/generated/torch.Tensor.detach.html
+            self.att_mat = att_mat.detach()
         att_mat = self.drop_weights(att_mat)
         x = torch.einsum("bhxy,bhyd->bhxd", att_mat, v)
         x = self.out_rearrange(x)

monai/networks/blocks/transformerblock.py

@@ -24,15 +24,22 @@ class TransformerBlock(nn.Module):
     """
 
     def __init__(
-        self, hidden_size: int, mlp_dim: int, num_heads: int, dropout_rate: float = 0.0, qkv_bias: bool = False
+        self,
+        hidden_size: int,
+        mlp_dim: int,
+        num_heads: int,
+        dropout_rate: float = 0.0,
+        qkv_bias: bool = False,
+        save_attn: bool = False,
     ) -> None:
         """
         Args:
             hidden_size: dimension of hidden layer.
             mlp_dim: dimension of feedforward layer.
             num_heads: number of attention heads.
             dropout_rate: faction of the input units to drop.
-            qkv_bias: apply bias term for the qkv linear layer
+            qkv_bias: apply bias term for the qkv linear layer.
+            save_attn: to make accessible the attention matrix post training.
 
         """
 
@@ -46,7 +53,7 @@ def __init__(
 
         self.mlp = MLPBlock(hidden_size, mlp_dim, dropout_rate)
         self.norm1 = nn.LayerNorm(hidden_size)
-        self.attn = SABlock(hidden_size, num_heads, dropout_rate, qkv_bias)
+        self.attn = SABlock(hidden_size, num_heads, dropout_rate, qkv_bias, save_attn)
         self.norm2 = nn.LayerNorm(hidden_size)
 
     def forward(self, x):

tests/test_selfattention.py

@@ -52,6 +52,26 @@ def test_ill_arg(self):
         with self.assertRaises(ValueError):
             SABlock(hidden_size=620, num_heads=8, dropout_rate=0.4)
 
+    def test_access_attn_matrix(self):
+        # input format
+        hidden_size = 128
+        num_heads = 2
+        dropout_rate = 0
+        input_shape = (2, 256, hidden_size)
+
+        # be able to access the matrix
+        no_matrix_acess_blk = SABlock(hidden_size=hidden_size, num_heads=num_heads, dropout_rate=dropout_rate)
+        with self.assertRaises(AttributeError):
+            no_matrix_acess_blk(torch.randn(input_shape))
+            no_matrix_acess_blk.att_mat
+
+        # be not able to acess the attention matrix
+        matrix_acess_blk = SABlock(
+            hidden_size=hidden_size, num_heads=num_heads, dropout_rate=dropout_rate, save_attn=True
+        )
+        matrix_acess_blk(torch.randn(input_shape))
+        assert matrix_acess_blk.att_mat.shape == (input_shape[0], input_shape[0], input_shape[1], input_shape[1])
+
 
 if __name__ == "__main__":
     unittest.main()