returnn 1.20250225.145151__tar.gz → 1.20250226.104737__tar.gz

{returnn-1.20250225.145151/returnn.egg-info → returnn-1.20250226.104737}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250225.145151
+Version: 1.20250226.104737
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250226.104737/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250226.104737'
+long_version = '1.20250226.104737+git.5a134ff'

{returnn-1.20250225.145151 → returnn-1.20250226.104737}/returnn/frontend/_backend.py

@@ -391,17 +391,17 @@ class Backend(Generic[T]):
         source: Tensor,
         *,
         dims: Sequence[Dim],
-        out_dim: Optional[Dim] = None,
-    ) -> Tuple[Tensor, Dim]:
+        out_dim: Dim,
+    ) -> Tensor:
         """
         Merges a list of axes into a single one. (Flatten the dims.)
         E.g. input is (batch, width, height, dim) and dims=(width,height), then we get (batch, width*height, dim).
         Or input is (batch, time, height, dim) and axes=(height,dim), then we get (batch, time, height*dim).
 
         :param source:
-        :param dims:
-        :param out_dim:
-        :return: tensor, out_dim
+        :param dims: list of dims to merge. len(dims) >= 2
+        :param out_dim: resulting merged dim
+        :return: tensor
         """
         raise NotImplementedError
 
@@ -939,8 +939,8 @@ class Backend(Generic[T]):
         raise NotImplementedError
 
     @staticmethod
-    def flip(source: Tensor, *, axis: Dim) -> Tensor:
-        """flip"""
+    def flip_no_mask(source: Tensor, *, axis: Dim) -> Tensor:
+        """flip, ignoring masking"""
         raise NotImplementedError
 
     @staticmethod

{returnn-1.20250225.145151 → returnn-1.20250226.104737}/returnn/frontend/array_.py

@@ -182,8 +182,29 @@ def merge_dims(
         else:
             out_dim = Dim(1, name="ext")
         return rf.expand_dim(source, out_dim), out_dim
+    if len(dims) == 1:
+        if out_dim is None or out_dim == dims[0]:
+            return source, dims[0]
+        return rf.replace_dim(source, in_dim=dims[0], out_dim=out_dim)
+    if out_dim is None:
+        out_dim = dims[0]
+        reset_dyn_size = False
+        for d in dims[1:]:
+            reset_dyn_size |= d.need_masking() and out_dim.capacity != 1
+            out_dim = out_dim * d
+        if reset_dyn_size:
+            # The dynamic sizes as calculated via dim math would not correctly describe how the tensor looks like.
+            # This would then potentially discard some of the data in the tensor in subsequent operations,
+            # when masking is applied.
+            # Thus, discard the dynamic sizes, and just treat it as a flat dim with scalar dynamic size.
+            # https://github.com/rwth-i6/returnn/issues/1694
+            out_dim_size = dims[0].get_dim_value_tensor()
+            for d in dims[1:]:
+                out_dim_size *= d.get_dim_value_tensor()
+            assert isinstance(out_dim_size, Tensor) and out_dim_size.dims == ()  # scalar
+            out_dim.dyn_size_ext = out_dim_size
     # noinspection PyProtectedMember
-    return source._raw_backend.merge_dims(source, dims=dims, out_dim=out_dim)
+    return source._raw_backend.merge_dims(source, dims=dims, out_dim=out_dim), out_dim
 
 
 def split_dims(
@@ -1066,7 +1087,7 @@ def reverse_sequence(tensor: Tensor, *, axis: Dim, handle_dynamic_dims: bool = T
     """
     if not handle_dynamic_dims or not axis.need_masking():
         # noinspection PyProtectedMember
-        return tensor._raw_backend.flip(tensor, axis=axis)
+        return tensor._raw_backend.flip_no_mask(tensor, axis=axis)
     indices = rf.combine_bc(axis.get_size_tensor(), "-", rf.range_over_dim(axis)) - 1
     return rf.gather(tensor, indices=indices, axis=axis, clip_to_valid=True)
 

{returnn-1.20250225.145151 → returnn-1.20250226.104737}/returnn/tf/frontend_layers/_backend.py

@@ -241,8 +241,8 @@ class ReturnnLayersBackend(Backend[Layer]):
         source: Tensor,
         *,
         dims: Sequence[Dim],
-        out_dim: Optional[Dim] = None,
-    ) -> Tuple[Tensor, Dim]:
+        out_dim: Dim,
+    ) -> Tensor:
         """
         Merges a list of axes into a single one. (Flatten the dims.)
         E.g. input is (batch, width, height, dim) and dims=(width,height), then we get (batch, width*height, dim).
@@ -251,18 +251,14 @@ class ReturnnLayersBackend(Backend[Layer]):
         :param source:
         :param dims:
         :param out_dim:
-        :return: tensor, out_dim
+        :return: tensor
         """
         if not isinstance(source, Tensor):
             raise TypeError(f"merge_dims: unexpected type for source {source!r}, need tensor")
-        if out_dim is None:
-            out_dim = dims[0]
-            for d in dims[1:]:
-                out_dim = out_dim * d
         layer = rfl.make_layer(
             {"class": "merge_dims", "from": source, "axes": dims, "out_dim": out_dim}, name="merge_dims"
         )
-        return layer, out_dim
+        return layer
 
     @staticmethod
     def split_dims(
@@ -692,7 +688,7 @@ class ReturnnLayersBackend(Backend[Layer]):
         )
 
     @staticmethod
-    def flip(source: Tensor, *, axis: Dim) -> Tensor:
+    def flip_no_mask(source: Tensor, *, axis: Dim) -> Tensor:
         """flip"""
         return rfl.make_layer(
             {"class": "slice", "from": source, "axis": axis, "out_dim": axis, "slice_step": -1}, name="flip"

{returnn-1.20250225.145151 → returnn-1.20250226.104737}/returnn/torch/frontend/_backend.py

@@ -262,8 +262,8 @@ class TorchBackend(Backend[torch.Tensor]):
         source: Tensor,
         *,
         dims: Sequence[Dim],
-        out_dim: Optional[Dim] = None,
-    ) -> Tuple[Tensor, Dim]:
+        out_dim: Dim,
+    ) -> Tensor:
         """
         Merges a list of axes into a single one. (Flatten the dims.)
         E.g. input is (batch, width, height, dim) and dims=(width,height), then we get (batch, width*height, dim).
@@ -272,18 +272,12 @@ class TorchBackend(Backend[torch.Tensor]):
         :param source:
         :param dims:
         :param out_dim:
-        :return: tensor, out_dim
+        :return: tensor
         """
-        assert dims
-        if len(dims) == 1:
-            return source, dims[0]
+        assert len(dims) >= 2
         first_axis = min(source.dims.index(d) for d in dims)
         pre_dims = source.dims[:first_axis]
         post_dims = [d for d in source.dims if d not in dims and d not in pre_dims]
-        if out_dim is None:
-            out_dim = dims[0]
-            for d in dims[1:]:
-                out_dim = out_dim * d
         source = source.copy_transpose(tuple(pre_dims) + tuple(dims) + tuple(post_dims), allow_int=False)
         out = Tensor(
             "merge_dims",
@@ -295,7 +289,7 @@ class TorchBackend(Backend[torch.Tensor]):
         out.raw_tensor = torch.reshape(source.raw_tensor, out_shape)
         if source.feature_dim and source.feature_dim in dims:
             out.feature_dim = out_dim
-        return out, out_dim
+        return out
 
     @staticmethod
     def split_dims(
@@ -1185,8 +1179,8 @@ class TorchBackend(Backend[torch.Tensor]):
         return out
 
     @staticmethod
-    def flip(source: Tensor, *, axis: Dim) -> Tensor:
-        """flip"""
+    def flip_no_mask(source: Tensor, *, axis: Dim) -> Tensor:
+        """flip, ignoring masking"""
         axis_int = source.get_axis_from_description(axis, allow_int=False)
         out = source.copy_template("flip")
         out.raw_tensor = torch.flip(source.raw_tensor, [axis_int])
@@ -1224,6 +1218,8 @@ class TorchBackend(Backend[torch.Tensor]):
     @staticmethod
     def sort(source: Tensor, *, axis: Dim, descending: bool, stable: bool) -> Tuple[Tensor, Tensor, Dim]:
         """sort. return values and indices"""
+        if axis.need_masking():
+            raise NotImplementedError(f"sort: dynamic axis {axis} not supported")
         axis_int = source.get_axis_from_description(axis, allow_int=False)
         # Move to last axis. Should be more efficient.
         source = source.copy_move_axis(axis_int, -1)

{returnn-1.20250225.145151 → returnn-1.20250226.104737/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250225.145151
+Version: 1.20250226.104737
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250225.145151/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250225.145151'
-long_version = '1.20250225.145151+git.8debd7e'