use functional interface for softmax in attention (#14198)

* use functional interface instead of instantiating module and immediately calling it

* fix torch.nn.functional to nn.functional. Thank you Stas!

examples/research_projects/movement-pruning/emmental/modeling_bert_masked.py

@@ -152,7 +152,7 @@ class BertSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/albert/modeling_albert.py

@@ -348,7 +348,7 @@ class AlbertAttention(nn.Module):
                 attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/beit/modeling_beit.py

@@ -244,7 +244,7 @@ class BeitSelfAttention(nn.Module):
             attention_scores = attention_scores + relative_position_bias
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/bert/modeling_bert.py

@@ -327,7 +327,7 @@ class BertSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/canine/modeling_canine.py

@@ -477,7 +477,7 @@ class CanineSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/deit/modeling_deit.py

@@ -159,7 +159,7 @@ class DeiTSelfAttention(nn.Module):
         attention_scores = attention_scores / math.sqrt(self.attention_head_size)
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/distilbert/modeling_distilbert.py

@@ -206,7 +206,7 @@ class MultiHeadSelfAttention(nn.Module):
         mask = (mask == 0).view(mask_reshp).expand_as(scores)  # (bs, n_heads, q_length, k_length)
         scores = scores.masked_fill(mask, -float("inf"))  # (bs, n_heads, q_length, k_length)
 
-        weights = nn.Softmax(dim=-1)(scores)  # (bs, n_heads, q_length, k_length)
+        weights = nn.functional.softmax(scores, dim=-1)  # (bs, n_heads, q_length, k_length)
         weights = self.dropout(weights)  # (bs, n_heads, q_length, k_length)
 
         # Mask heads if we want to

src/transformers/models/electra/modeling_electra.py

@@ -319,7 +319,7 @@ class ElectraSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/gpt2/modeling_gpt2.py

@@ -209,7 +209,7 @@ class GPT2Attention(nn.Module):
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
 
-        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
 
         # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
         attn_weights = attn_weights.type(value.dtype)
@@ -260,7 +260,7 @@ class GPT2Attention(nn.Module):
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
 
-        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
 
         # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
         if attn_weights.dtype != torch.float32:

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -199,7 +199,7 @@ class GPTNeoSelfAttention(nn.Module):
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
 
-        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
         attn_weights = attn_weights.to(value.dtype)
         attn_weights = self.attn_dropout(attn_weights)
 

src/transformers/models/gptj/modeling_gptj.py

@@ -151,7 +151,7 @@ class GPTJAttention(nn.Module):
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
 
-        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
         attn_weights = attn_weights.to(value.dtype)
         attn_weights = self.attn_dropout(attn_weights)
 

src/transformers/models/ibert/quant_modules.py

@@ -409,7 +409,7 @@ class IntSoftmax(nn.Module):
 
     def forward(self, x, scaling_factor):
         if not self.quant_mode:
-            return nn.Softmax(dim=-1)(x), None
+            return nn.functional.softmax(x, dim=-1), None
 
         x_int = x / scaling_factor
 

src/transformers/models/layoutlm/modeling_layoutlm.py

@@ -235,7 +235,7 @@ class LayoutLMSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/luke/modeling_luke.py

@@ -399,7 +399,7 @@ class LukeSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/lxmert/modeling_lxmert.py

@@ -356,7 +356,7 @@ class LxmertAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/megatron_bert/modeling_megatron_bert.py

@@ -298,7 +298,7 @@ class MegatronBertSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/mobilebert/modeling_mobilebert.py

@@ -264,7 +264,7 @@ class MobileBertSelfAttention(nn.Module):
             # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
             attention_scores = attention_scores + attention_mask
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.
         attention_probs = self.dropout(attention_probs)

src/transformers/models/mpnet/modeling_mpnet.py

@@ -184,7 +184,7 @@ class MPNetSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         attention_probs = self.dropout(attention_probs)
 

src/transformers/models/openai/modeling_openai.py

@@ -185,7 +185,7 @@ class Attention(nn.Module):
             # Apply the attention mask
             w = w + attention_mask
 
-        w = nn.Softmax(dim=-1)(w)
+        w = nn.functional.softmax(w, dim=-1)
         w = self.attn_dropout(w)
 
         # Mask heads if we want to

src/transformers/models/rembert/modeling_rembert.py

@@ -290,7 +290,7 @@ class RemBertSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/roberta/modeling_roberta.py

@@ -262,7 +262,7 @@ class RobertaSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/roformer/modeling_roformer.py

@@ -301,7 +301,7 @@ class RoFormerSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/segformer/modeling_segformer.py

@@ -165,7 +165,7 @@ class SegformerEfficientSelfAttention(nn.Module):
         attention_scores = attention_scores / math.sqrt(self.attention_head_size)
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/splinter/modeling_splinter.py

@@ -202,7 +202,7 @@ class SplinterSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/tapas/modeling_tapas.py

@@ -419,7 +419,7 @@ class TapasSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/visual_bert/modeling_visual_bert.py

@@ -244,7 +244,7 @@ class VisualBertSelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

src/transformers/models/vit/modeling_vit.py

@@ -193,7 +193,7 @@ class ViTSelfAttention(nn.Module):
         attention_scores = attention_scores / math.sqrt(self.attention_head_size)
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py

@@ -304,7 +304,7 @@ class {{cookiecutter.camelcase_modelname}}SelfAttention(nn.Module):
             attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.