returnn 1.20240925.224259__tar.gz → 1.20240926.154934__tar.gz

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240925.224259
+Version: 1.20240926.154934
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240926.154934/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240926.154934'
+long_version = '1.20240926.154934+git.bd0d963'

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/returnn/datasets/audio.py

@@ -63,7 +63,9 @@ class OggZipDataset(CachedDataset2):
         :param str|list[str]|((str)->str)|None targets_post_process: :func:`get_post_processor_function`,
             applied on orth
         :param bool use_cache_manager: uses :func:`returnn.util.basic.cf`
-        :param str|None segment_file: .txt or .gz text file containing sequence tags that will be used as whitelist
+        :param str|None segment_file: .txt or .gz text file containing sequence tags that will be used as whitelist.
+            Note: This is somewhat deprecated, as we also support ``seq_list_filter_file`` (via the base class),
+            which does the same but more universally.
         :param bool zip_audio_files_have_name_as_prefix:
         :param float|int|None fixed_random_subset:
           Value in [0,1] to specify the fraction, or integer >=1 which specifies number of seqs.

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/returnn/frontend/run_ctx.py

@@ -275,8 +275,8 @@ class RunCtx:
         if dims is None and expected_output:
             dims = expected_output.dims
         if dims is not None and expected_output:
-            assert (
-                expected_output.dims == dims
+            assert expected_output.dims == tuple(
+                dims
             ), f"mark_as_output: {name!r} dims mismatch from expected output, given {dims}, expected {expected_output}"
 
         if not isinstance(tensor, Tensor):

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/returnn/tensor/_dim_extra.py

@@ -1643,25 +1643,14 @@ class _DimMixin:
         if self._extra:
             self._extra.derived_from_op = None
             self._extra.derived_from_tag = None
-        if self_same_as is not self:
-            assert not self_same_as.same_as
-            if self_same_as is other_same_base:
-                return
-            with util.guard_infinite_recursion(_d.Dim.declare_same_as, self_same_as, other_same_base):
-                self_same_as.declare_same_as(other_same_base)
-            if (self.dyn_size_ext is None or not self._validate_in_current_graph()) and self_same_as.dyn_size_ext:
-                self.dyn_size_ext = self_same_as.get_dyn_size_ext_for_batch_ctx(
-                    self.batch, self.control_flow_ctx, template_only=True
-                )
-        else:
-            if self._extra:
-                for dim_ in self._extra.same_for_batch_ctx.values():
+        if self._extra:
+            for dim_ in self._extra.same_for_batch_ctx.values():
+                # noinspection PyProtectedMember
+                if dim_._extra:
                     # noinspection PyProtectedMember
-                    if dim_._extra:
-                        # noinspection PyProtectedMember
-                        dim_._extra.derived_from_op = None
-                        # noinspection PyProtectedMember
-                        dim_._extra.derived_from_tag = None
+                    dim_._extra.derived_from_op = None
+                    # noinspection PyProtectedMember
+                    dim_._extra.derived_from_tag = None
         # Now merge existing variants. But only if not derived via op, because in that case, we can (and should!)
         # automatically infer it. Note that we only got here when the other is not the same dim, so it means that
         # the other is really different, the sizes are potentially different, but we want to overtake the other.
@@ -1723,6 +1712,13 @@ class _DimMixin:
                 other_same_base.batch, other_same_base.control_flow_ctx, template_only=True
             )
             other_same_base._maybe_update()
+        if (
+            self.dyn_size_ext
+            and self.dyn_size_ext.raw_tensor is None
+            and other_same_base.dyn_size_ext.raw_tensor is not None
+        ):
+            self.dyn_size_ext = other_same_base.dyn_size_ext.copy()
+            self._maybe_update()
         if self.is_dim_known() and other.is_dim_known():
             assert self.dimension == other.dimension
         elif self.is_dim_known() and not other.is_dim_known():
@@ -1762,6 +1758,13 @@ class _DimMixin:
             if self_ is not self:
                 self.control_flow_ctx = self_.control_flow_ctx  # might be different
                 self.dyn_size_ext = self_.dyn_size_ext  # might be unset
+        if self_same_as is not self:
+            assert not self_same_as.same_as
+            assert self_same_as is not other_same_base
+            with util.guard_infinite_recursion(_d.Dim.declare_same_as, self_same_as, other_same_base):
+                # Do this at the end, when this is not our own base anymore,
+                # because otherwise any reset() might incorrectly reset the new other base.
+                self_same_as.declare_same_as(other_same_base)
 
     def _merge_same_for_batch_ctx_dict(self: _d.Dim, other: _d.Dim):
         """
@@ -2301,7 +2304,6 @@ class _DimMixin:
         value = {"dim": self.dimension}
         if self.kind is not None:
             value["kind"] = self.kind.name
-        assert self.derived_from_op is None  # not handled yet for hashing...
         return value
 
 

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240925.224259
+Version: 1.20240926.154934
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240925.224259 → returnn-1.20240926.154934}/tests/test_rf_base.py

@@ -263,6 +263,30 @@ def test_dim_value():
     run_model(extern_data, lambda *, epoch, step: _Net(), _forward_step)
 
 
+def test_dim_size_after_declare_same_as():
+    time_dim = Dim(Tensor("time", [batch_dim], dtype="int32"))
+    extern_data = TensorDict(
+        {
+            "data": Tensor("data", [batch_dim, time_dim], dtype="float32"),
+        }
+    )
+    out_spatial_dim = Dim(None, name="out_spatial")
+
+    # noinspection PyShadowingNames
+    def _forward_step(**_kwargs):
+        assert time_dim.dyn_size_ext.raw_tensor is not None
+        enc_spatial_dim = time_dim * 2
+        assert enc_spatial_dim.dyn_size_ext.raw_tensor is not None
+        out_spatial_dim.declare_same_as(enc_spatial_dim)
+        assert enc_spatial_dim.dyn_size_ext.raw_tensor is not None
+        assert out_spatial_dim.dyn_size_ext.raw_tensor is not None
+        out_spatial_dim.dyn_size_ext.mark_as_default_output(shape=[batch_dim])
+
+    # Running it twice can trigger some weird behavior in Dim.declare_same_as.
+    run_model(extern_data, lambda **_kwargs: rf.Module(), _forward_step, test_tensorflow=False)
+    run_model(extern_data, lambda **_kwargs: rf.Module(), _forward_step, test_tensorflow=False)
+
+
 def test_dim_mask():
     time_dim = Dim(Tensor("time", [batch_dim], dtype="int32"))
     in_dim = Dim(7, name="in")

returnn-1.20240925.224259/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240925.224259'
-long_version = '1.20240925.224259+git.f199967'