Move reshape after softmax calculations, to avoid reshaping back to t…

…he original shape where needed.

PiperOrigin-RevId: 686542189

lingvo/core/attention.py

@@ -510,7 +510,6 @@ def AttenProbs(inputs: py_utils.NestedMap) -> tf.Tensor:
       """Generates probs."""
       source_batch = py_utils.GetShape(inputs.source_vecs)[2]
       multiplier = py_utils.GetShape(inputs.query_vec)[1]
-      target_batch = multiplier * source_batch
 
       # Shape of summed is [sl, tb/sb, sb, hidden_dim].
       summed = tf.tanh(inputs.source_vecs + inputs.query_vec)
@@ -568,13 +567,10 @@ def AttenProbs(inputs: py_utils.NestedMap) -> tf.Tensor:
             tf.logging.warning(
                 'packed_input is False but query_segment_id is passed.'
             )
-        source_padding = tf.reshape(source_padding, [-1, target_batch])
-        source_padding = tf.transpose(source_padding)
+        source_padding = tf.transpose(source_padding, [1, 2, 0])
 
-      # [source_length, target_batch]
-      logits = tf.reshape(logits, [-1, target_batch])
-      # [target_batch, source_length]
-      logits = tf.transpose(logits)
+      # [multiplier, source_batch, source_length]
+      logits = tf.transpose(logits, [1, 2, 0])
       # take the softmax to compute the probabilities.
       probs = self._PaddedSoftmax(logits, source_padding)
       return probs
@@ -638,31 +634,32 @@ def Atten(
         args.source_segment_id = source_segment_id
       if query_segment_id is not None:
         args.query_segment_id = query_segment_id
-      # probs is of shape [target_batch, source_length]
+      # probs is of shape [multiplier, source_batch, source_length]
       probs = py_utils.CallDefun(AttenProbs, args)
-      probs = py_utils.HasShape(probs, [target_batch, source_length])
+      probs = py_utils.HasShape(
+          probs, [multiplier, source_batch, source_length]
+      )
 
       # Apply dropout to weights if applicable.
       if not self.do_eval:
         probs = _ApplyAttentionDropout(p, probs)
 
-      # Reshape probs to be of shape
-      # [target_batch/source_batch, source_batch, source_length]
-      probs_reshaped = tf.reshape(probs, [multiplier, source_batch, -1])
+      # Shape of returned probs is [target_batch, source_length]
+      probs_returned = tf.reshape(probs, [target_batch, source_length])
       # Transpose probs to be of shape
       # [source_batch, target_batch/source_batch, source_length]
-      probs_reshaped = tf.transpose(probs_reshaped, [1, 0, 2])
+      probs = tf.transpose(probs, [1, 0, 2])
       # Batched matmul
       # [source_batch, target_batch/source_batch, source_length] *
       # [source_batch, source_length, context_dim] =
       # [source_batch, target_batch/source_batch, context_dim]
-      summed = tf.matmul(probs_reshaped, source_contexts)
+      summed = tf.matmul(probs, source_contexts)
 
       # summed is of shape
       # [target_batch/source_batch, source_batch, context_dim]
       summed = tf.transpose(summed, [1, 0, 2])
 
-      return tf.reshape(summed, [target_batch, -1]), probs
+      return tf.reshape(summed, [target_batch, -1]), probs_returned
 
     # The source batch size equals to the target batch size.
     def AttenSameBatchSize(
@@ -1036,7 +1033,8 @@ def AttenProbs(
           * query_segment_id:        [target_batch].
 
       Returns:
-        logits [target_batch, source_time].
+        logits: [n, source_batch, time]
+          where n = target_batch // source_batch
       """
       source_vecs = inputs.source_vecs
 
@@ -1121,14 +1119,9 @@ def AttenProbs(
                 'packed_input is False but query_segment_id is passed.'
             )
           source_padding = tf.transpose(source_padding, [2, 0, 1])
-        source_padding = tf.reshape(source_padding, [target_batch, -1])
 
       # => [n, source_batch, time]
       logits = tf.transpose(logits, [2, 0, 1])
-
-      # => [n * source_batch, time].
-      # This makes logits store content in the same order as query_vec.
-      logits = tf.reshape(logits, [target_batch, -1])
       if p.atten_logit_cap is not None and p.atten_logit_cap > 0:
         logits = py_utils.MaybeSoftCapLogits(logits, p.atten_logit_cap)
       probs = self._PaddedSoftmax(logits, source_padding)
@@ -1193,11 +1186,12 @@ def Atten(
         args.source_segment_id = source_segment_id
       if query_segment_id is not None:
         args.query_segment_id = query_segment_id
-      returned_probs = py_utils.CallDefun(AttenProbs, args)
-      returned_probs = py_utils.HasShape(returned_probs, [target_batch, time])
+      # [n, source_batch, time] where n = target_batch // source_batch
+      probs = py_utils.CallDefun(AttenProbs, args)
+      probs = py_utils.HasShape(probs, [-1, source_batch, time])
+      # [target_batch, time]
+      returned_probs = tf.reshape(probs, [-1, time])
 
-      # => [n, source_batch, time] where n = target_batch // source_batch
-      probs = tf.reshape(returned_probs, [-1, source_batch, time])
       # => [source_batch, n, time].
       probs = tf.transpose(probs, [1, 0, 2])
 
@@ -2455,36 +2449,34 @@ def Atten(
           ),
       )
       # Take out the padding states.
-      # _source_padding is of shape [sl, sb].
-      # reshaped to [sl, 1,  sb].
+      # [sl, 1, sb]
       source_padding = tf.expand_dims(source_padding, 1)
+      # [sl, tb/sb, sb]
       per_step_source_padding = tf.reshape(
           tf.transpose(per_step_source_padding), [-1, multiplier, sb]
       )
       if source_padding.dtype != tf.bool:
         source_padding = source_padding > 0
       if per_step_source_padding.dtype != tf.bool:
         per_step_source_padding = per_step_source_padding > 0
+      # [sl, tb/sb, sb]
       source_padding = tf.logical_or(source_padding, per_step_source_padding)
 
-      # Reshape logits to a matrix of shape [tb, sl] and takes the
-      # softmax to compute the probabilities.
-      logits = tf.transpose(tf.reshape(logits, [-1, tb]))
-      source_padding = tf.transpose(tf.reshape(source_padding, [-1, tb]))
+      # [tb/sb, sb, sl]
+      logits = tf.transpose(logits, [1, 2, 0])
+      source_padding = tf.transpose(source_padding, [1, 2, 0])
+      # Take the softmax to compute the probabilities.
       probs = self._PaddedSoftmax(logits, source_padding)
-      # Reshape probs to be of shape [tb/sb, sb, sl].
-      probs_reshaped = tf.reshape(probs, [multiplier, sb, -1])
+      # [tb, sl]
+      probs_returned = tf.reshape(probs, [tb, -1])
       # Transpose probs to be of shape [sb, tb/sb, sl]
-      probs_reshaped = tf.transpose(probs_reshaped, [1, 0, 2])
+      probs = tf.transpose(probs, [1, 0, 2])
       # [sb, tb/sb, sl] * [sb, sl, context_dim] = [sb, tb/sb, context_dim]
-      summed = tf.matmul(
-          tf.cast(probs_reshaped, source_contexts.dtype),
-          source_contexts,
-      )
+      summed = tf.matmul(tf.cast(probs, source_contexts.dtype), source_contexts)
       summed = self.QAct('atten_context', summed)
       # summed is of shape [tb/sb, sb, context_dim]
       summed = tf.transpose(summed, [1, 0, 2])
-      return tf.reshape(summed, [tb, -1]), probs
+      return tf.reshape(summed, [tb, -1]), probs_returned
 
     def AttenSameBatchSize(
         hidden_var,