Working feedforward of full transformer arch in notebook

3 files changed, 6 insertions, 4 deletions

diff --git a/text_recognizer/networks/transformer/attention.py b/text_recognizer/networks/transformer/attention.py
index a3b53f0..7bafc58 100644
--- a/text_recognizer/networks/transformer/attention.py
+++ b/text_recognizer/networks/transformer/attention.py
@@ -91,12 +91,12 @@ class Attention(nn.Module):
     def forward(
         self,
         x: Tensor,
-        context: Optional[Tensor],
-        mask: Optional[Tensor],
-        context_mask: Optional[Tensor],
+        context: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        context_mask: Optional[Tensor] = None,
         rotary_pos_emb: Optional[Tensor] = None,
     ) -> Tuple[Tensor, Tensor]:
-        b, n, _, device = x.shape, x.device
+        b, n, _, device = *x.shape, x.device
         q, k, v = self.qkv_fn(x)
         q, k = (
             self._apply_rotary_emb(q, k, rotary_pos_emb)
diff --git a/text_recognizer/networks/transformer/nystromer/attention.py b/text_recognizer/networks/transformer/nystromer/attention.py
index 5ab19cf..695a0d7 100644
--- a/text_recognizer/networks/transformer/nystromer/attention.py
+++ b/text_recognizer/networks/transformer/nystromer/attention.py
@@ -47,6 +47,7 @@ class NystromAttention(nn.Module):
         dropout_rate: float = 0.0,
     ):
         super().__init__()
+        self.dim = dim
         self.residual = None
         self.eps = eps
         self.num_heads = num_heads
diff --git a/text_recognizer/networks/transformer/nystromer/nystromer.py b/text_recognizer/networks/transformer/nystromer/nystromer.py
index 799a811..2113f1f 100644
--- a/text_recognizer/networks/transformer/nystromer/nystromer.py
+++ b/text_recognizer/networks/transformer/nystromer/nystromer.py
@@ -29,6 +29,7 @@ class Nystromer(nn.Module):
         glu: bool = True,
     ) -> None:
         super().__init__()
+        self.dim = dim
         self.layers = nn.ModuleList(
             [
                 nn.ModuleList(