returnn 1.20241211.145755__tar.gz → 1.20241213.155203__tar.gz

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241211.145755
+Version: 1.20241213.155203
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241213.155203/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20241213.155203'
+long_version = '1.20241213.155203+git.64d234b'

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/frontend/_backend.py

@@ -496,21 +496,6 @@ class Backend(Generic[T]):
         """
         raise NotImplementedError
 
-    @staticmethod
-    def cum_concat_step(source: Tensor, *, prev_accum: Tensor, axis: Dim, out_spatial_dim: Dim) -> Tensor:
-        """
-        Concatenates all previous frames over a time-axis.
-        See RETURNN :class:`CumConcatLayer` for details.
-
-        :param source: same dims as prev_accum except for the accum axis
-        :param prev_accum: previous accumulated tensor, shape {..., axis}
-        :param axis: the axis to accumulate over
-        :param out_spatial_dim: the spatial dim of the output will be this dim. like axis+1.
-        :return: accumulated. accumulated shape {..., out_spatial_dim},
-            same shape as prev_accum with axis replaced by out_spatial_dim.
-        """
-        raise NotImplementedError
-
     @staticmethod
     def stack(sources: Sequence[Tensor], *, out_dim: Dim) -> Tensor:
         """
@@ -1095,6 +1080,15 @@ class Backend(Generic[T]):
         out.raw_tensor = source.raw_tensor
         return out
 
+    @staticmethod
+    def set_sparse_dim(source: Tensor, sparse_dim: Dim) -> Tensor:
+        """set sparse dim"""
+        # This default implementation works fine as long as the backend
+        # does not have special treatments of Tensor and dim tags itself (like TF net dict backend).
+        out = source.copy()
+        out.sparse_dim = sparse_dim
+        return out
+
     _AllowedReduceModes = {"sum", "max", "min", "mean", "logsumexp", "any", "all", "argmin", "argmax"}
 
     @staticmethod

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/frontend/array_.py

@@ -367,6 +367,7 @@ def concat(
     *sources: Tuple[Tensor, Dim],
     allow_broadcast: bool = False,
     out_dim: Optional[Dim] = None,
+    handle_dynamic_dims: Optional[bool] = None,
 ) -> Tuple[Tensor, Dim]:
     """
     Concatenates multiple sources in the specified dimension.
@@ -376,6 +377,7 @@ def concat(
     :param sources: list of (tensor, dim) pairs. dim is the axis to concatenate on.
     :param allow_broadcast: if True, the sources can have different dims, and the result will be broadcasted.
     :param out_dim: reuse existing dim for the resulting concatenated dim, if given
+    :param handle_dynamic_dims:
     :return: concatenated tensor, out_dim
     """
     assert sources
@@ -385,6 +387,9 @@ def concat(
             assert src.dims_set - {dim} == dims, f"concat {sources}, need allow_broadcast=True"
     if not out_dim:
         out_dim = sum(d for _, d in sources)
+    if handle_dynamic_dims is None or handle_dynamic_dims:
+        for src, dim in sources[:-1]:
+            assert dim.is_static(), f"concat {sources}, dim {dim} is not static, not yet implemented..."
     # noinspection PyProtectedMember
     return sources[0][0]._raw_backend.concat(*sources, allow_broadcast=allow_broadcast, out_dim=out_dim), out_dim
 
@@ -507,13 +512,18 @@ def cum_concat_step(
     :return: (accumulated, out_spatial_dim). accumulated shape {..., out_spatial_dim},
         same shape as prev_accum with axis replaced by out_spatial_dim.
     """
+    # Note: Before, we had a backend function just for this.
+    # In case of TF-layers, this was using CumConcatLayer.
+    # This would allow for automatic optimization when inside a RecLayer.
+    # However, we don't really need this for eager frameworks,
+    # and we want to simplify this for now,
+    # using pure RF code.
     if not out_spatial_dim:
         out_spatial_dim = axis + 1
-    # noinspection PyProtectedMember
-    return (
-        source._raw_backend.cum_concat_step(source, prev_accum=prev_accum, axis=axis, out_spatial_dim=out_spatial_dim),
-        out_spatial_dim,
+    out, (out_spatial_dim,) = rf.pad(
+        prev_accum, axes=[axis], padding=[(0, 1)], out_dims=[out_spatial_dim], value=source, handle_dynamic_dims=True
     )
+    return out, out_spatial_dim
 
 
 def stack(sources: Sequence[Tensor], *, out_dim: Optional[Dim] = None) -> Tuple[Tensor, Dim]:

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/frontend/attention.py

@@ -862,13 +862,24 @@ def _make_indices(
         query_spatial_dim_m1 = query_spatial_dim - 1
         q_pos_vec = rf.range_over_dim(query_spatial_dim_m1)  # [q_len-1]
 
+        # The masking in the output is quite custom (left+right masking), so our seq lens don't make sense,
+        # and might even cause to fail some tests (that e.g. max(q_seq_len+k_seq_len-1) == shape).
+        out_spatial_dim = Dim(
+            query_spatial_dim_m1.get_dim_value_tensor() + key_value_spatial_dim.get_dim_value_tensor(),
+            name=f"2*{query_spatial_dim.description}-1"
+            if (query_spatial_dim == key_value_spatial_dim)
+            else f"{query_spatial_dim.description}+{key_value_spatial_dim.description}-1",
+        )
+
         # We want to have all distances as in rf.combine_bc(kv_pos_vec, "-", q_pos_vec) with shape [q_len,kv_len].
         # We want to store only non-duplicates.
         # The min value is with kv_pos=0, q_pos=q_len-1: -(q_len-1)
         # The max value is with kv_pos=kv_len-1, q_pos=0: k_len-1
-        indices, out_spatial_dim = rf.concat(
+        indices, _ = rf.concat(
             (q_pos_vec - query_spatial_dim_m1.get_dim_value_tensor(), query_spatial_dim_m1),
             (kv_pos_vec, key_value_spatial_dim),
+            out_dim=out_spatial_dim,
+            handle_dynamic_dims=False,
         )
     if query_offset is not None:
         indices = indices - query_offset

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/frontend/dims.py

@@ -15,6 +15,7 @@ __all__ = [
     "range_over_dim_strided",
     "range_over_merged_dims",
     "replace_dim",
+    "set_sparse_dim",
     "dim_match_priority_when_needed",
     "num_elements_of_shape",
     "masked_fraction_of_shape",
@@ -94,6 +95,16 @@ def replace_dim(source: Tensor, *, in_dim: Dim, out_dim: Optional[Dim] = None) -
     return source._raw_backend.replace_dim(source, in_dim=in_dim, out_dim=out_dim), out_dim
 
 
+def set_sparse_dim(source: Tensor, sparse_dim: Dim) -> Tensor:
+    """
+    :param source:
+    :param sparse_dim:
+    :return: source with sparse_dim set
+    """
+    # noinspection PyProtectedMember
+    return source._raw_backend.set_sparse_dim(source, sparse_dim)
+
+
 def dim_match_priority_when_needed(dim: Dim, *other_dims: Dim) -> Dim:
     """
     :return: maybe copy of dim with higher match_priority if needed to distinguish from other_dims

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/tf/frontend_layers/_backend.py

@@ -342,7 +342,20 @@ class ReturnnLayersBackend(Backend[Layer]):
         opts = {}
         if allow_broadcast:
             opts["allow_broadcast"] = True
-        out_dim = sum(d for _, d in sources)
+        dim_deps = rfl.get_dim_deps(out_dim)
+        sources_dims = set()
+        for source, _ in sources:
+            sources_dims.update(source.dims)
+        need_explicit_dim_deps = False
+        for dim in dim_deps:
+            if dim not in sources_dims:
+                need_explicit_dim_deps = True
+                break
+        if need_explicit_dim_deps:
+            source0 = rfl.make_layer(
+                {"class": "copy", "from": sources[0][0], "extra_deps": dim_deps}, name="concat_extra_dim_deps"
+            )
+            sources = ((source0, sources[0][1]),) + sources[1:]
         return rfl.make_layer(
             {"class": "concat", "from": sources, "out_dim": out_dim, **opts},
             name="concat",
@@ -375,20 +388,6 @@ class ReturnnLayersBackend(Backend[Layer]):
             name="pad",
         )
 
-    @staticmethod
-    def cum_concat_step(source: Tensor, *, prev_accum: Tensor, axis: Dim, out_spatial_dim: Dim) -> Tensor:
-        """cum_concat_step"""
-        return rfl.make_layer(
-            {
-                "class": "cum_concat",
-                "from": source,
-                "state": {"state": prev_accum},
-                "out_spatial_dim": out_spatial_dim,
-                "axis": axis,
-            },
-            name="cum_concat",
-        )
-
     @staticmethod
     def activation(tensor: Tensor, func: str) -> Tensor:
         """activation"""
@@ -774,6 +773,14 @@ class ReturnnLayersBackend(Backend[Layer]):
             {"class": "reinterpret_data", "set_dim_tags": {in_dim: out_dim}, "from": source}, name="new_dim"
         )
 
+    @staticmethod
+    def set_sparse_dim(source: Tensor, sparse_dim: Dim) -> Tensor:
+        """set sparse dim"""
+        return rfl.make_layer(
+            {"class": "reinterpret_data", "set_sparse": True, "set_sparse_dim": sparse_dim, "from": source},
+            name="set_sparse_dim",
+        )
+
     @staticmethod
     def reduce(source: Tensor, *, mode: str, axis: Union[Dim, Sequence[Dim]], use_mask: bool = True) -> Tensor:
         """Reduce"""

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/tf/layers/basic.py

@@ -517,11 +517,12 @@ class ConcatLayer(LayerBase):
             dimension = 0
             for tag in concat_dim_tags:
                 dimension += tag.dimension
+        sum_concat_dim_tags: Dim = sum(concat_dim_tags)
         if not out_dim:
-            out_dim = sum(concat_dim_tags)
+            out_dim = sum_concat_dim_tags
             assert isinstance(out_dim, Dim)
-        else:
-            sum(concat_dim_tags).declare_same_as(out_dim)
+        elif not out_dim.is_dim_known():
+            sum_concat_dim_tags.declare_same_as(out_dim)
         assert out_dim.dimension == dimension
 
         def _as_common(x, axis):

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/torch/frontend/_backend.py

@@ -452,9 +452,6 @@ class TorchBackend(Backend[torch.Tensor]):
     ) -> Tensor:
         """pad"""
         assert len(out_dims) == len(axes) == len(padding)
-        out = source.copy_template_new_dim_tags(
-            [out_dims[axes.index(dim)] if dim in axes else dim for dim in source.dim_tags], keep_special_axes=True
-        )
         remaining_dims = set(axes)
         raw_pad = []
         for dim in reversed(source.dims):
@@ -469,10 +466,24 @@ class TorchBackend(Backend[torch.Tensor]):
             ]
             if not remaining_dims:
                 break
-        if isinstance(value, Tensor):
-            assert value.dims == (), f"value {value} must be a scalar"
-            value = value.raw_tensor
-        out.raw_tensor = torch.nn.functional.pad(source.raw_tensor, pad=raw_pad, mode=mode, value=value)
+        # Use torch.nn.functional.pad if possible.
+        if (isinstance(value, Tensor) and value.dims == ()) or (not isinstance(value, Tensor)):
+            if isinstance(value, Tensor):
+                assert value.dims == ()
+                value = value.raw_tensor
+            out = source.copy_template_new_dim_tags(
+                [out_dims[axes.index(dim)] if dim in axes else dim for dim in source.dim_tags], keep_special_axes=True
+            )
+            out.raw_tensor = torch.nn.functional.pad(source.raw_tensor, pad=raw_pad, mode=mode, value=value)
+        else:  # Fallback to concat.
+            assert isinstance(value, Tensor)
+            assert all(dim in source.dims and dim not in axes for dim in value.dims)
+            assert len(axes) == 1  # not implemented otherwise currently...
+            ext_dim = Dim(1, name="ext")
+            value_ext = rf.expand_dim(value, ext_dim)
+            out = TorchBackend.concat(
+                (source, axes[0]), (value_ext, ext_dim), allow_broadcast=True, out_dim=out_dims[0]
+            )
         if any(dim.need_masking() for dim in out_dims) and handle_dynamic_dims:
             if all(right == 0 for right in raw_pad[1::2]) and mode != "circular":
                 pass  # no masking needed
@@ -490,24 +501,12 @@ class TorchBackend(Backend[torch.Tensor]):
                                 rf.copy_to_device((left + middle).dyn_size_ext, out.device),
                             )
                             out.raw_tensor = torch.where(
-                                mask.copy_compatible_to(out, check_dtype=False, check_sparse=False).raw_tensor,
+                                mask.copy_compatible_to_dims_raw(out.dims),
                                 out.raw_tensor,
-                                value,
+                                value.copy_compatible_to_dims_raw(out.dims) if isinstance(value, Tensor) else value,
                             )
         return out
 
-    @staticmethod
-    def cum_concat_step(source: Tensor, *, prev_accum: Tensor, axis: Dim, out_spatial_dim: Dim) -> Tensor:
-        """cum concat step"""
-        out = prev_accum.copy_template_replace_dim_tag(
-            axis=prev_accum.get_axis_from_description(axis),
-            new_dim_tag=out_spatial_dim,
-            name=f"{source.name}/cum_concat_step",
-        )
-        source_raw = source.copy_compatible_to_dims_raw(prev_accum.dims)
-        out.raw_tensor = torch.cat((prev_accum.raw_tensor, source_raw), dim=prev_accum.get_axis_from_description(axis))
-        return out
-
     @staticmethod
     def stack(sources: Sequence[Tensor], *, out_dim: Dim) -> Tensor:
         """stack"""

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn/util/file_cache.py

@@ -472,7 +472,7 @@ class _TouchFilesThread(Thread):
         """thread main loop"""
         while True:
             all_files = {}  # dict to have order deterministic
-            for filename in self.files:
+            for filename in self.files.copy():  # copy dict under GIL to avoid modifications during iteration
                 all_files[filename] = True
                 all_files.update({k: True for k in _all_parent_dirs(filename, base_dir=self.cache_base_dir)})
             for filename in all_files:

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241211.145755
+Version: 1.20241213.155203
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/tests/test_rf_array.py

@@ -275,6 +275,46 @@ def test_pad_time_right():
         assert all(out_.raw_tensor[b, seq_len] == 1.0)
 
 
+def test_pad_time_right_non_scalar():
+    time_dim = Dim(Tensor("time", [batch_dim], dtype="int32"))
+    in_dim = Dim(7, name="in")
+    extern_data = TensorDict(
+        {
+            "data": Tensor("data", [batch_dim, time_dim, in_dim], dtype="float32"),
+            "value": Tensor("value", [batch_dim], dtype="float32"),
+        }
+    )
+
+    # noinspection PyShadowingNames
+    def _forward_step(*, extern_data: TensorDict, **_kwargs):
+        data, value = extern_data["data"], extern_data["value"]
+        data.mark_as_output("data", shape=(batch_dim, time_dim, in_dim))
+        value.mark_as_output("value", shape=(batch_dim,))
+        out, (new_time,) = rf.pad(data, axes=[time_dim], padding=[(0, 1)], value=value)
+        out.mark_as_default_output(shape=(batch_dim, new_time, in_dim))
+
+    # TF-layers currently does not support this.
+    res = run_model(extern_data, lambda **_kwargs: rf.Module(), _forward_step, test_tensorflow=False)
+    data_: Tensor = res["data"]
+    value_: Tensor = res["value"]
+    out_: Tensor = res["output"]
+    assert data_.dims == (batch_dim, time_dim, in_dim)
+    new_time_dim = out_.dims[1]
+    assert out_.dims == (batch_dim, new_time_dim, in_dim) and new_time_dim != time_dim
+    assert time_dim.dyn_size_ext.dims == new_time_dim.dyn_size_ext.dims == (batch_dim,)
+    batch_size = batch_dim.get_dim_value()
+    assert batch_size > 1
+    assert len(set(time_dim.dyn_size_ext.raw_tensor)) > 1  # not all the same
+    for b in range(batch_size):
+        seq_len = time_dim.dyn_size_ext.raw_tensor[b]
+        new_seq_len = new_time_dim.dyn_size_ext.raw_tensor[b]
+        print(f"batch {b}, seq_len {seq_len}, new_seq_len {new_seq_len}")
+        assert new_seq_len == seq_len + 1
+        np.testing.assert_allclose(data_.raw_tensor[b, :seq_len], out_.raw_tensor[b, :seq_len])
+        print(out_.raw_tensor[b])
+        assert all(out_.raw_tensor[b, seq_len] == value_.raw_tensor[b])
+
+
 def test_stack():
     batch_dim_ = Dim(3, name="batch")
     time_dim = Dim(5, name="time")

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/tests/test_rf_attention.py

@@ -542,6 +542,27 @@ def test_relative_positional_encoding():
     run_model(extern_data, lambda *, epoch, step: _Net(), _forward_step)
 
 
+def test_relative_positional_encoding_cross():
+    enc_spatial_dim = Dim(Tensor("enc_spatial", [batch_dim], dtype="int32"))
+    dec_spatial_dim = Dim(Tensor("dec_spatial", [batch_dim], dtype="int32"))
+    in_dim = Dim(8, name="in")
+    extern_data = TensorDict(
+        {
+            "enc": Tensor("enc", [batch_dim, enc_spatial_dim, in_dim], dtype="float32"),
+            "dec": Tensor("dec", [batch_dim, dec_spatial_dim, in_dim], dtype="float32"),
+        }
+    )
+
+    # noinspection PyShadowingNames
+    def _forward_step(**_kwargs):
+        out, dim = rf.relative_positional_encoding(
+            key_value_spatial_dim=enc_spatial_dim, query_spatial_dim=dec_spatial_dim, feat_dim=in_dim
+        )
+        out.mark_as_default_output(shape=(dim, in_dim))
+
+    run_model(extern_data, lambda **_kwargs: rf.Module(), _forward_step)
+
+
 def test_rel_pos_self_attention():
     time_dim = Dim(Tensor("time", [batch_dim], dtype="int32"))
     in_dim = Dim(8, name="in")
@@ -550,6 +571,7 @@ def test_rel_pos_self_attention():
             "data": Tensor("data", [batch_dim, time_dim, in_dim], dtype="float32"),
         }
     )
+    check_batching = False
 
     class _Net(rf.Module):
         def __init__(self):
@@ -565,6 +587,32 @@ def test_rel_pos_self_attention():
 
         def __call__(self, x: Tensor, *, axis: Dim) -> Tensor:
             """forward"""
+            nonlocal check_batching
+            if check_batching:
+                assert rf.is_executing_eagerly()
+                assert batch_dim in x.dims and axis != batch_dim
+                y = self.self_att(x, axis=axis)
+                for b in range(batch_dim.get_dim_value()):
+                    x_b = rf.gather(x, axis=batch_dim, indices=b)
+                    assert batch_dim in axis.dyn_size_ext.dims  # current assumption...
+                    seq_len = rf.gather(axis.dyn_size_ext, axis=batch_dim, indices=b)
+                    axis_b = Dim(seq_len)
+                    # Note: The current order (replace_dim and then slice) is somewhat dependent
+                    # on the current internal behavior of gather and replace_dim,
+                    # which might change at some point...
+                    x_b, _ = rf.replace_dim(x_b, in_dim=axis, out_dim=axis_b)
+                    x_b, _ = rf.slice(x_b, axis=axis_b, start=0, end=seq_len, out_dim=axis_b)
+                    y_b = self.self_att(x_b, axis=axis_b)
+                    y_b_ = rf.gather(y, axis=batch_dim, indices=b)
+                    y_b_, _ = rf.replace_dim(y_b_, in_dim=axis, out_dim=axis_b)
+                    y_b_, _ = rf.slice(y_b_, axis=axis_b, start=0, end=seq_len, out_dim=axis_b)
+                    y_b_ = y_b_.copy_transpose(y_b.dims)
+                    # Assuming PyTorch...
+                    np.testing.assert_almost_equal(
+                        y_b.raw_tensor.cpu().detach().numpy(), y_b_.raw_tensor.cpu().detach().numpy(), decimal=5
+                    )
+                return y
+
             return self.self_att(x, axis=axis)
 
     # noinspection PyShadowingNames
@@ -573,6 +621,8 @@ def test_rel_pos_self_attention():
         out.mark_as_default_output(shape=(batch_dim, time_dim, model.out_dim))
 
     run_model(extern_data, lambda *, epoch, step: _Net(), _forward_step)
+    check_batching = True
+    run_model(extern_data, lambda *, epoch, step: _Net(), _forward_step, test_tensorflow=False)
 
 
 def test_sinusoidal_positional_encoding():

{returnn-1.20241211.145755 → returnn-1.20241213.155203}/tests/test_rf_encoder_conformer.py

@@ -359,15 +359,16 @@ def test_e_branchformer():
                 (batch_dim, num_heads_dim, enc_spatial_dim, key_dim_per_head),
             ),
             # Check RelPositionalEncoding vs our relative_positional_encoding
-            (
-                (rf.RelPosSelfAttention.__call__, 0, "pos_emb", 0),
-                (RelPositionMultiHeadedAttention.forward, 0, "pos_emb", 0),
-                lambda x, **_: _tensor(
-                    _reorder_rel_pos_emb_espnet_to_rf(x.squeeze(dim=0)),
-                    "pos_emb",
-                    [enc_spatial_dim - 1 + enc_spatial_dim, model_dim],
-                ),
-            ),
+            # Currently disabled this check, as the dim tags are different now...
+            # (
+            #     (rf.RelPosSelfAttention.__call__, 0, "pos_emb", 0),
+            #     (RelPositionMultiHeadedAttention.forward, 0, "pos_emb", 0),
+            #     lambda x, **_: _tensor(
+            #         _reorder_rel_pos_emb_espnet_to_rf(x.squeeze(dim=0)),
+            #         "pos_emb",
+            #         [enc_spatial_dim - 1 + enc_spatial_dim, model_dim],
+            #     ),
+            # ),
             (
                 (EBranchformerLayer.__call__, 0, "x_mhsa", 0),
                 (EBranchformerEncoderLayer.forward, 0, "x_att", 0),

returnn-1.20241211.145755/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20241211.145755'
-long_version = '1.20241211.145755+git.4c6e245'