Add dimensionality of heads argument to SABlock

monai/networks/blocks/selfattention.py

@@ -32,6 +32,7 @@ def __init__(
  dropout_rate: float = 0.0,
  qkv_bias: bool = False,
  save_attn: bool = False,
+ dim_head: int | None = None,
  ) -> None:
  """
  Args:
@@ -40,6 +41,7 @@ def __init__(
  dropout_rate (float, optional): fraction 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.
+ dim_head (int, optional): dimension of each head. Defaults to hidden_size // num_heads.
 
  """
 
@@ -52,14 +54,16 @@ def __init__(
  raise ValueError("hidden size should be divisible by num_heads.")
 
  self.num_heads = num_heads
- self.out_proj = nn.Linear(hidden_size, hidden_size)
- self.qkv = nn.Linear(hidden_size, hidden_size * 3, bias=qkv_bias)
+ self.dim_head = hidden_size // num_heads if dim_head is None else dim_head
+ self.inner_dim = self.dim_head * num_heads
+
+ self.out_proj = nn.Linear(self.inner_dim, hidden_size)
+ self.qkv = nn.Linear(hidden_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.drop_output = nn.Dropout(dropout_rate)
  self.drop_weights = nn.Dropout(dropout_rate)
- self.head_dim = hidden_size // num_heads
- self.scale = self.head_dim**-0.5
+ self.scale = self.dim_head**-0.5
  self.save_attn = save_attn
  self.att_mat = torch.Tensor()
 

tests/test_selfattention.py

@@ -74,6 +74,40 @@ def test_access_attn_matrix(self):
  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])
 
+ def test_number_of_parameters(self):
+
+ def count_sablock_params(*args, **kwargs):
+ """Count the number of parameters in a SABlock."""
+ sablock = SABlock(*args, **kwargs)
+ return sum([x.numel() for x in sablock.parameters() if x.requires_grad])
+
+ hidden_size = 128
+ num_heads = 8
+ default_dim_head = hidden_size // num_heads
+
+ # Default dim_head is hidden_size // num_heads
+ nparams_default = count_sablock_params(hidden_size=hidden_size, num_heads=num_heads)
+ nparams_like_default = count_sablock_params(
+ hidden_size=hidden_size, num_heads=num_heads, dim_head=default_dim_head
+ )
+ self.assertEqual(nparams_default, nparams_like_default)
+
+ # Increasing dim_head should increase the number of parameters
+ nparams_custom_large = count_sablock_params(
+ hidden_size=hidden_size, num_heads=num_heads, dim_head=default_dim_head * 2
+ )
+ self.assertGreater(nparams_custom_large, nparams_default)
+
+ # Decreasing dim_head should decrease the number of parameters
+ nparams_custom_small = count_sablock_params(
+ hidden_size=hidden_size, num_heads=num_heads, dim_head=default_dim_head // 2
+ )
+ self.assertGreater(nparams_default, nparams_custom_small)
+
+ # Increasing the number of heads with the default behaviour should not change the number of params.
+ nparams_default_more_heads = count_sablock_params(hidden_size=hidden_size, num_heads=num_heads * 2)
+ self.assertEqual(nparams_default, nparams_default_more_heads)
+
 
 if __name__ == "__main__":
  unittest.main()