fixing tests

metnet/layers/MultiheadSelfAttention2D.py

@@ -15,8 +15,8 @@ class MultiheadSelfAttention2D(nn.Module):
     def __init__(
         self,
         in_channels: int,
-        attention_channels: int,
-        num_heads: int,
+        attention_channels: int = 64,
+        num_heads: int = 16,
         attn_drop: float = 0.0,
         proj_drop: float = 0.0,
         use_normalised_qk: bool = True,
@@ -31,9 +31,9 @@ def __init__(
             Number of channels in the input image.
         attention_channels : int
             Number of channels used for attention computations.
-            It should be divisible by num_heads.
+            It should be divisible by num_heads, by default 64
         num_heads : int
-            Number of attention heads.
+            Number of attention heads, by default 16
         attn_drop : float, optional
             attention dropout rate, by default 0.0
         proj_drop : float, optional

tests/test_layers.py

@@ -1,8 +1,7 @@
 from metnet.layers.StochasticDepth import StochasticDepth
 from metnet.layers.SqueezeExcitation import SqueezeExcite
 from metnet.layers.MBConv import MBConv
-from metnet.layers.PartitionAttention import BlockAttention, GridAttention
-from metnet.layers.RelativeSelfAttention import RelativeSelfAttention
+from metnet.layers.MultiheadSelfAttention2D import MultiheadSelfAttention2D
 import torch
 
 
@@ -32,10 +31,10 @@ def test_mbconv():
     assert test_tensor.shape == mb_conv(test_tensor).shape
 
 
-def test_relative_attention():
+def test_multiheaded_self_attention_2D():
     n, c, h, w = 1, 3, 16, 16
     test_tensor = torch.rand(n, c, h, w)
-    rel_self_attention = RelativeSelfAttention(c)
+    rel_self_attention = MultiheadSelfAttention2D(c)
     assert test_tensor.shape == rel_self_attention(test_tensor).shape