Fix typo and format BasicTransformerBlock attributes

examples/controlnet/train_controlnet.py

@@ -581,7 +581,7 @@ def make_train_dataset(args, tokenizer, accelerator):
 
     if args.conditioning_image_column is None:
         conditioning_image_column = column_names[2]
-        logger.info(f"conditioning image column defaulting to {caption_column}")
+        logger.info(f"conditioning image column defaulting to {conditioning_image_column}")
     else:
         conditioning_image_column = args.conditioning_image_column
         if conditioning_image_column not in column_names:

src/diffusers/models/attention.py

@@ -225,6 +225,12 @@ def __init__(
             )
 
         # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
         self.attn1 = Attention(
             query_dim=dim,
             heads=num_attention_heads,
@@ -235,10 +241,16 @@ def __init__(
             upcast_attention=upcast_attention,
         )
 
-        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
-
         # 2. Cross-Attn
         if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+            )
             self.attn2 = Attention(
                 query_dim=dim,
                 cross_attention_dim=cross_attention_dim if not double_self_attention else None,
@@ -248,30 +260,13 @@ def __init__(
                 bias=attention_bias,
                 upcast_attention=upcast_attention,
             )  # is self-attn if encoder_hidden_states is none
-        else:
-            self.attn2 = None
-
-        if self.use_ada_layer_norm:
-            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
-        elif self.use_ada_layer_norm_zero:
-            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
-        else:
-            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
-
-        if cross_attention_dim is not None or double_self_attention:
-            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
-            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
-            # the second cross attention block.
-            self.norm2 = (
-                AdaLayerNorm(dim, num_embeds_ada_norm)
-                if self.use_ada_layer_norm
-                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
-            )
         else:
             self.norm2 = None
+            self.attn2 = None
 
         # 3. Feed-forward
         self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
 
     def forward(
         self,