Use scaled_dot_product_attention in Wav2vec2/HuBERT's SelfAttention (p…

…ytorch#3253) (pytorch#3261)

Summary:
Replace the attention computation with `torch.nn.functional.scaled_dot_product_attention` to improve running efficiency.

Pull Request resolved: pytorch#3253

Reviewed By: mthrok

Differential Revision: D44800353

Pulled By: nateanl

fbshipit-source-id: 41550d868c809099aadbe812b0ebe2c38121efb8

torchaudio/models/wav2vec2/components.py

@@ -262,7 +262,7 @@ def __init__(
 
         self.embed_dim = embed_dim
         self.num_heads = num_heads
-        self.dropout = torch.nn.Dropout(dropout)
+        self.dropout = dropout
         self.head_dim = head_dim
 
         self.scaling = self.head_dim**-0.5
@@ -304,25 +304,14 @@ def forward(
 
         shape = (batch_size, length, self.num_heads, self.head_dim)
         q = self.q_proj(x).view(*shape).transpose(2, 1)  # B, nH, L, Hd
-        k = self.k_proj(x).view(*shape).permute(0, 2, 3, 1)  # B, nH, Hd, L
+        k = self.k_proj(x).view(*shape).transpose(2, 1)  # B, nH, L, Hd
         v = self.v_proj(x).view(*shape).transpose(2, 1)  # B, nH, L, Hd
-
-        # scale down q to avoid value overflow.
-        weights = (self.scaling * q) @ k  # B, nH, L, L
-        if attention_mask is not None:
-            weights += attention_mask
-        # subtracting a constant value from the tensor won't change the output of softmax.
-        # apply the subtraction to avoid value overflow in torch.nn.functional.softmax.
-        # for more details, please see Equation 7 in https://arxiv.org/abs/2112.08778
-        weights = weights - weights.max(dim=-1, keepdim=True)[0]
-
-        weights = torch.nn.functional.softmax(weights, dim=-1)
-        weights = self.dropout(weights)
-
-        output = weights @ v  # B, nH, L, Hd
-        output = output.transpose(2, 1).reshape(batch_size, length, embed_dim)
-
-        output = self.out_proj(output)
+        dropout = self.dropout if self.training else 0.0
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            q, k, v, attn_mask=attention_mask, dropout_p=dropout, is_causal=False
+        )
+        attn_output = attn_output.transpose(1, 2).reshape(batch_size, -1, self.num_heads * self.head_dim)
+        output = self.out_proj(attn_output)
         return output, None  # Necessary for compatibility with WavLMSelAttention