Merge branch 'scaled_adam_exp236' into scaled_adam_exp242

egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py

@@ -331,6 +331,8 @@ class ZipformerEncoderLayer(nn.Module):
             d_model, attention_dim, nhead, pos_dim, dropout=0.0,
         )
 
+        self.pooling = PoolingModule(d_model)
+
         self.feed_forward1 = FeedforwardModule(d_model,
                                                feedforward_dim,
                                                dropout)
@@ -372,9 +374,9 @@ class ZipformerEncoderLayer(nn.Module):
             # ensure we get grads if self.bypass_scale becomes out of range
             return self.bypass_scale
         # hardcode warmup period for bypass scale
-        warmup_period = 4000.0
-        initial_clamp_min = 0.5
-        final_clamp_min = 0.2
+        warmup_period = 20000.0
+        initial_clamp_min = 0.75
+        final_clamp_min = 0.25
         if self.batch_count > warmup_period:
             clamp_min = final_clamp_min
         else:
@@ -382,6 +384,21 @@ class ZipformerEncoderLayer(nn.Module):
                          (self.batch_count / warmup_period) * (initial_clamp_min - final_clamp_min))
         return self.bypass_scale.clamp(min=clamp_min, max=1.0)
 
+    def get_dynamic_dropout_rate(self):
+        # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this
+        # starts at 0.2 and rapidly decreases to 0.  Its purpose is to keep the training stable
+        # at the beginning, by making the network focus on the feedforward modules.
+        if torch.jit.is_scripting() or not self.training:
+            return 0.0
+        warmup_period = 2000.0
+        initial_dropout_rate = 0.2
+        final_dropout_rate = 0.0
+        if self.batch_count > warmup_period:
+            return final_dropout_rate
+        else:
+            return (initial_dropout_rate -
+                    (initial_dropout_rate * final_dropout_rate) * (self.batch_count / warmup_period))
+
     def forward(
         self,
         src: Tensor,
@@ -411,24 +428,37 @@ class ZipformerEncoderLayer(nn.Module):
         # macaron style feed forward module
         src = src + self.feed_forward1(src)
 
+        # dropout rate for submodules that interact with time.
+        dynamic_dropout = self.get_dynamic_dropout_rate()
+
+        # pooling module
+        if torch.jit.is_scripting() or random.random() > dynamic_dropout:
+            src = src + self.pooling(src,
+                                     key_padding_mask=src_key_padding_mask)
+
         # multi-headed self-attention module
-        src_att, attn_weights = self.self_attn(
-            src,
-            pos_emb=pos_emb,
-            attn_mask=src_mask,
-            key_padding_mask=src_key_padding_mask,
-        )
-        src = src + src_att
+        use_self_attn = (random.random() > dynamic_dropout)
+        if torch.jit.is_scripting() or use_self_attn:
+            src_att, attn_weights = self.self_attn(
+                src,
+                pos_emb=pos_emb,
+                attn_mask=src_mask,
+                key_padding_mask=src_key_padding_mask,
+            )
+            src = src + src_att
 
         # convolution module
-        src = src + self.conv_module1(src, src_key_padding_mask=src_key_padding_mask)
+        if torch.jit.is_scripting() or random.random() > dynamic_dropout:
+            src = src + self.conv_module1(src, src_key_padding_mask=src_key_padding_mask)
 
 
         src = src + self.feed_forward2(src)
 
-        src = src + self.self_attn.forward2(src, attn_weights)
+        if torch.jit.is_scripting() or use_self_attn:
+            src = src + self.self_attn.forward2(src, attn_weights)
 
-        src = src + self.conv_module2(src, src_key_padding_mask=src_key_padding_mask)
+        if torch.jit.is_scripting() or random.random() > dynamic_dropout:
+            src = src + self.conv_module2(src, src_key_padding_mask=src_key_padding_mask)
 
         src = src + self.feed_forward3(src)
 
@@ -1397,6 +1427,43 @@ class RelPositionMultiheadAttention(nn.Module):
                 logging.info(f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}")
 
 
+
+
+class PoolingModule(nn.Module):
+    """
+    Averages the input over the time dimension and project with a square matrix.
+    """
+    def __init__(self,
+                 d_model: int):
+        super().__init__()
+        self.proj = ScaledLinear(d_model, d_model,
+                                 initial_scale=0.1, bias=False)
+
+    def forward(self,
+                x: Tensor,
+                key_padding_mask):
+        """
+        Args:
+           x: a Tensor of shape (T, N, C)
+          key_padding_mask: a Tensor of bool, of shape (N, T), with True in masked
+            positions.
+        Returns:
+           a Tensor of shape (1, N, C)
+        """
+        if key_padding_mask is not None:
+            pooling_mask = key_padding_mask.logical_not().to(x.dtype)  # (N, T)
+            pooling_mask = (pooling_mask / pooling_mask.sum(dim=1, keepdim=True))
+            pooling_mask = pooling_mask.transpose(0, 1).contiguous().unsqueeze(-1)
+            # now pooling_mask: (T, N, 1)
+        else:
+            num_frames = x.shape[0]
+            pooling_mask = 1.0 / num_frames
+
+        x = (x * pooling_mask).sum(dim=0, keepdim=True)
+        x = self.proj(x)
+        return x
+
+
 class FeedforwardModule(nn.Module):
     """Feedforward module in Zipformer model.
     """