update float32 to float64

MI/Algorithm/conformer_2class.py

@@ -82,7 +82,7 @@ class MultiHeadAttention(nn.Module):
         values = rearrange(self.values(x), "b n (h d) -> b h n d", h=self.num_heads)
         energy = torch.einsum('bhqd, bhkd -> bhqk', queries, keys)  
         if mask is not None:
-            fill_value = torch.finfo(torch.float32).min
+            fill_value = torch.finfo(torch.float64).min
             energy.mask_fill(~mask, fill_value)
 
         scaling = self.emb_size ** (1 / 2)