Dim broadcasting only allowed for auto-generated dims (#1388)

Auto-generated dims are those via the legacy n_out
or all other internal dims.
Any dim tags created explicitly by the user
should never be treated as broadcast dims.

See also #666.

returnn/tensor/_dim_extra.py

@@ -1234,7 +1234,13 @@ def is_equal(
             if other_kind == DimTypes.Feature or not other_kind:
                 other_kind = DimTypes.Spatial
         if self.dimension != other.dimension:
-            if broadcast_matches and (self.dimension == 1 or other.dimension == 1):
+            if broadcast_matches and (
+                # Only auto-generated dim tags are allowed to be treated as broadcastable.
+                # This was another suggestion from here: https://github.com/rwth-i6/returnn/issues/666
+                # It was not implemented like this because the auto_generated flag was only introduced later.
+                (self.dimension == 1 and self.auto_generated)
+                or (other.dimension == 1 and other.auto_generated)
+            ):
                 pass  # pass on
             else:
                 return False

returnn/tensor/_tensor_extra.py

@@ -977,7 +977,13 @@ def copy_add_dim_by_tag(self, dim_tag, unbroadcast=False, axis=None) -> _t.Tenso
                 if dim_tag and dim_tag.dimension == 1 and dim_tag.batch == batch_info:
                     pass  # keep it
                 else:
-                    dim_tag = Dim(kind=Dim.Types.Batch, description="batch-broadcast", dimension=1, batch=batch_info)
+                    dim_tag = Dim(
+                        kind=Dim.Types.Batch,
+                        description="batch-broadcast",
+                        dimension=1,
+                        batch=batch_info,
+                        auto_generated=True,
+                    )
                 return self.copy_add_batch_dim(batch_dim_axis=axis, batch=batch_info, dim_tag=dim_tag)
 
         data_opts = self.get_kwargs()

tests/test_TFUtil.py

@@ -6,11 +6,11 @@
 import tensorflow as tf
 from returnn.tf.util.basic import *
 from returnn.tf.util.data import SpatialDim, FeatureDim
-from returnn.tf.network import ExternData
 import returnn.tf.compat as tf_compat
 from nose.tools import assert_equal, assert_not_equal, assert_is_instance, assert_is, assert_in, assert_true
 from numpy.testing.utils import assert_almost_equal, assert_allclose
 from pprint import pprint
+import contextlib
 import unittest
 import numpy.testing
 from returnn.util import better_exchook
@@ -762,6 +762,35 @@ def test_Data_get_common_data_broadcast_multiple_dim_tags():
     assert out.dim_tags == (batch_dim, time_dim, input_dim, feat_dim)
 
 
+@contextlib.contextmanager
+def set_behavior_version(version: int):
+    """
+    This is a context manager which sets the behavior version to the given value.
+    """
+    from returnn.util.basic import BehaviorVersion
+
+    # noinspection PyProtectedMember
+    old = BehaviorVersion._behavior_version
+    try:
+        BehaviorVersion._behavior_version = version
+        yield
+    finally:
+        BehaviorVersion._behavior_version = old
+
+
+def test_Data_get_common_data_no_broadcast_for_explicit():
+    from returnn.tf.util.data import batch_dim
+
+    time_dim = SpatialDim("time")
+    input_dim = FeatureDim("input", 3)
+    feat_dim = FeatureDim("feat", 1)
+    a = Data("a", dim_tags=[batch_dim, time_dim, input_dim])
+    b = Data("b", dim_tags=[feat_dim])
+    with set_behavior_version(0):
+        out = Data.get_common_data([a, b], allow_broadcast_all_sources=True)
+        assert out.dim_tags == (batch_dim, time_dim, input_dim, feat_dim)
+
+
 def test_Data_get_common_data_extra2_static_spatial():
     d1 = Data(name="t", shape=(None, 32, 32, 128), dtype="float32", auto_create_placeholders=True)
     d2 = Data(name="r", shape=(None, 32, 32, 128), dtype="float32", auto_create_placeholders=True)