returnn 1.20250225.201207__py3-none-any.whl → 1.20250226.115259__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250225.201207
+Version: 1.20250226.115259
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn/_setup_info_generated.py

@@ -1,2 +1,2 @@
-version = '1.20250225.201207'
-long_version = '1.20250225.201207+git.c7cfe6c'
+version = '1.20250226.115259'
+long_version = '1.20250226.115259+git.0d32534'

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
 

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(

returnn/tensor/_dim_extra.py

@@ -390,15 +390,15 @@ class _DimMixin:
             if dim_extra:
                 dim_extra.cache_dyn_size_ext_dev.clear()
                 dim_extra.cache_seq_mask.clear()
+                if dim.dyn_size_ext is not None or dim.dimension is None:
+                    dim_extra.cache_dim_math.clear()
+                else:
+                    dim_extra.cache_dim_math.clear_dynamic()
             if only_self:
                 return
             if dim_extra:
                 # Any dims via dim math could also contain raw tensors,
                 # so iterate through them.
-                if dim.dyn_size_ext is not None or dim.dimension is None:
-                    dim_extra.cache_dim_math.clear()
-                else:
-                    dim_extra.cache_dim_math.clear_dynamic()
                 queue += dim_extra.cache_dim_math.values()
                 if dim_extra.same_as:
                     queue.append(dim_extra.same_as)

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(

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(

returnn/torch/util/debug_inf_nan.py

@@ -39,6 +39,7 @@ from io import TextIOBase
 import traceback
 from types import FrameType
 import torch
+import tree
 
 # noinspection PyProtectedMember
 from torch.utils._python_dispatch import TorchDispatchMode
@@ -96,6 +97,7 @@ def debug_inf_nan(
 
 # For efficiency, and to be less spammy
 _TraceFuncNameBlacklist = {
+    "aten::empty.memory_format",
     "aten::zeros_like",
     "aten::ones_like",
     "aten::full",
@@ -113,6 +115,7 @@ _TraceFuncNameBlacklist = {
     "aten::split_with_sizes",
     "aten::slice.Tensor",
     "aten::select.int",
+    "aten::max_pool2d_with_indices",
 }
 
 
@@ -140,19 +143,20 @@ class _TraceOps(TorchDispatchMode):
         if self.report_every_op_call:
             print(f"--- op {func.name()}", file=self.file)
         out = func(*args, **kwargs)
-        if isinstance(out, torch.Tensor):
-            with no_python_dispatcher():
-                got_nan_inf_t = torch.stack([torch.isnan(out).any(), torch.isinf(out).any()]).cpu()
-                got_nan = got_nan_inf_t[0].item()
-                got_inf = got_nan_inf_t[1].item()
-                if got_nan or got_inf:
-                    s = "/".join([s_ for s_, b in [("nan", got_nan), ("inf", got_inf)] if b])
-                    print(f"--> {s} in {func}: {out}", file=self.file)
-                    traceback.print_list(
-                        _extract_stack_up_to(skip_top_num_frames=1, root_frame=self.root_frame), file=self.file
-                    )
-                    if self.stop_reporting_after_first_inf_nan:
-                        self.enabled = False
+        for out_ in tree.flatten(out):
+            if isinstance(out_, torch.Tensor):
+                with no_python_dispatcher():
+                    got_nan_inf_t = torch.stack([torch.isnan(out_).any(), torch.isinf(out_).any()]).cpu()
+                    got_nan = got_nan_inf_t[0].item()
+                    got_inf = got_nan_inf_t[1].item()
+                    if got_nan or got_inf:
+                        s = "/".join([s_ for s_, b in [("nan", got_nan), ("inf", got_inf)] if b])
+                        print(f"--> {s} in {func}: {out_}", file=self.file)
+                        traceback.print_list(
+                            _extract_stack_up_to(skip_top_num_frames=1, root_frame=self.root_frame), file=self.file
+                        )
+                        if self.stop_reporting_after_first_inf_nan:
+                            self.enabled = False
         return out
 
 

{returnn-1.20250225.201207.dist-info → returnn-1.20250226.115259.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250225.201207
+Version: 1.20250226.115259
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250225.201207.dist-info → returnn-1.20250226.115259.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=MlICTMbISeiq6sz_1NI8XnxGZLlQzE0eH9VfGGDjlKs,5215
+returnn/PKG-INFO,sha256=2Ws--V5aicc3WJ-I6OrqPbbrvVNTH3Cnno6L7yeIyKY,5215
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=qBFbuB1yN3adgVM5pXt2-Yq9vorjRNchNPL8kDKx44M,31752
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=DwJubgEQQUSxnLSgr9-UFixkYOeM2bKYKKxyIW_3L3w,77
+returnn/_setup_info_generated.py,sha256=3ur2a8rg2h6MJe2vAo7Tq4axfkV1GYJMcaQdnsmshb8,77
 returnn/config.py,sha256=3tmKhB6FnQZaNdtcYsiB61JnEY--iZ2qmJ4yq0b6tE0,29140
 returnn/forward_iface.py,sha256=A_OJiaXsX4MlXQRzST86ylyxSUZbC402PQL1REcqHjM,911
 returnn/learning_rate_control.py,sha256=ZvWryAn_tv9DhV8sh1LV3eE34Yltl3On3mYZAG4hR9s,34684
@@ -75,12 +75,12 @@ returnn/extern/graph_editor/subgraph.py,sha256=R3uIFqWgiL7L5S4YATm9o9a3wfEa_mSb4
 returnn/extern/graph_editor/transform.py,sha256=d9fEgu0JC342q0g9niVxRWMKzkQQA9mrrajBGcU1o_s,29349
 returnn/extern/graph_editor/util.py,sha256=QMrQeQZ7lJwsrNQub9tof0h3quEaoHiGJaZmogQ7jXE,18707
 returnn/frontend/__init__.py,sha256=2aS7nbxXniIrBp2DODl0xN0f3IJ_dX4Bi9ZlR7W5_DE,1472
-returnn/frontend/_backend.py,sha256=VWTe2ps8UK9BQpbnZRqNfbesQ6PGH5WHqkEa4ai8btw,50353
+returnn/frontend/_backend.py,sha256=TNkEdj9GKxJfSM1ZMQ_SdAQzn2TU7SQbG6JGdaWhUeI,50374
 returnn/frontend/_cache.py,sha256=JAhi7L-raQ3A-NC3JUYDtdRTwT3BGJJGGZxrZ8MfEWQ,8403
 returnn/frontend/_numpy_backend.py,sha256=2oCtG0YCWL_89v4cD_jDj8em1O_Fp-_YWl5EblGi_yo,7858
 returnn/frontend/_random_journal.py,sha256=_ktP_mjgx8vtQQGX_DofdhewJj0aPiczefTWeemPkmo,5457
 returnn/frontend/_utils.py,sha256=4A3MSRM0i86J77550uR_AjcBEPu6nymLUZ9Xd1V3Fkc,12073
-returnn/frontend/array_.py,sha256=CYk8lQinS2EDINBttl4UqSYP2BhqikeSjnbNy9Mzpx4,48013
+returnn/frontend/array_.py,sha256=M5vCeH0nlwJ-zrdjbZpsMLN6StOLn0iM7PnXvGLLE3g,49154
 returnn/frontend/attention.py,sha256=GKt-Xqnz8sIyXVrE0i4VCS7J2Wu7dmoH_BA0Cu8CrXQ,45769
 returnn/frontend/backend.py,sha256=iQ9w4xl8Ea7bgpb0VUaCKq50rV5Bl2E5J8Rhd-oqD_c,883
 returnn/frontend/build_from_dict.py,sha256=rfWa2rjjhIR_kIQED_nMrygrQBunS6unegzWTLVbC98,3017
@@ -154,7 +154,7 @@ returnn/sprint/extern_interface.py,sha256=l-v1X-Yg0UpTFe7Y3c4FwWOqpSNuv9Oy5EzqlK
 returnn/sprint/interface.py,sha256=_IGNQlOFcJcwsSeVkKcM-y8g2NDJv07jFhii47KfWtg,36490
 returnn/tensor/README.md,sha256=X6BqcRLrPLPnwF9yR69uqIFrMnNluj9pBkOPHwNgzuo,501
 returnn/tensor/__init__.py,sha256=on6j5PEOQpck50UcsR4nJzJSDmoVy34z1Oq4efv6Ax0,154
-returnn/tensor/_dim_extra.py,sha256=ywfU-vMewufkDYeM1UE3Gfee3NhAUkoSZxvC6L-lkH8,122551
+returnn/tensor/_dim_extra.py,sha256=kL_nnGNaRpKIQLlvCo6TJ35WynS_jIssNZusFmtOAE0,122551
 returnn/tensor/_tensor_extra.py,sha256=v8oacDyrNMlDTRF0XR0LcU04snr5I1D9_yidw1ZWKk4,164859
 returnn/tensor/_tensor_mixin_base.py,sha256=H5z86I0NejxrSgMH1c5oXQzBqS6L9HpvP4y7oegBaSc,643
 returnn/tensor/_tensor_op_overloads.py,sha256=kVVcnYtcZdW7Vjj78V1Im_yVX2M2r6dUTgeiAQZ37X0,5449
@@ -177,7 +177,7 @@ returnn/tf/sprint.py,sha256=Yqjh0-6sCWHpdDPQCzHKx7TwQCOjJyjfd0KHtnYdd-8,5471
 returnn/tf/updater.py,sha256=St4Z5iBjlkWaB6CiS-K1VNc_iLaan2e6-mVMTTPldzk,72034
 returnn/tf/frontend_layers/README.md,sha256=P4vVl_EK-4jT55m40mq-K4Nr9yFY0tJR5fmDzTHSDFE,1096
 returnn/tf/frontend_layers/__init__.py,sha256=MGUn7rv6fOefbtkX-5pq6fC1T6Y5h0oh1uOPSEcv1_I,506
-returnn/tf/frontend_layers/_backend.py,sha256=6bT_4fjfV0IRcFqcZ0kcWLx0eYZGRqAJDTEfWSRIFnA,47451
+returnn/tf/frontend_layers/_backend.py,sha256=U7rbRY9XgMkxxyWY2D8KG-KesSOEGLCxn-Gl6dgwmPc,47277
 returnn/tf/frontend_layers/_utils.py,sha256=ijByaDOqPDod5mZC9EoTkt8PHBEODXHsWbkwDOF9XW4,4205
 returnn/tf/frontend_layers/cond.py,sha256=yQ2h5W0sgMZndJdrWv2EE9k9yIcspQ1U0HwBSh3hOKE,14830
 returnn/tf/frontend_layers/config_entry_points.py,sha256=t01RWOiaZohzuqPXX-MLV0P5yCOfE0dz-9dZ77_pK4c,5751
@@ -216,7 +216,7 @@ returnn/torch/data/queued_data_iter.py,sha256=PoOsGHdHVZjTmcyfq_ZOw--P6hyfTdmAWI
 returnn/torch/data/returnn_dataset_wrapper.py,sha256=1Bw82-Ge_8m_DSDXZNqQ3zGDic2HQlp6jysELL0NVK0,7369
 returnn/torch/data/tensor_utils.py,sha256=-Teqi--LLbt6q_5mDRdoHZHmPgSdC83W706ukif_YiU,1284
 returnn/torch/frontend/__init__.py,sha256=AA48HZnC17ASuKA0EWy8loZ-Bib_yUtqF4T1wYvjst4,62
-returnn/torch/frontend/_backend.py,sha256=8rCnNRoiUf_Sqmb1u2Y7Mf89Hmzd0LkrroLoXVKn6ww,101468
+returnn/torch/frontend/_backend.py,sha256=TqyDWNP4XCvJNNGn8jyxaT8BOEjVE24QCUR3qsTIS3A,101242
 returnn/torch/frontend/_rand.py,sha256=1JgIkV2XmpgJD86zXZ-NCAe-QuoP2swr6NaS1oz3Qa8,1830
 returnn/torch/frontend/bridge.py,sha256=Z2_UW8AagezC7zsXDc5PKcd8G9WwisV7j9SWGHU0m4U,7840
 returnn/torch/frontend/raw_ops.py,sha256=lF0h-KtYYsdaaqQADylVZp9qzPskOOXA4MfmYDyx5IU,296
@@ -226,7 +226,7 @@ returnn/torch/optim/lion.py,sha256=jV_qfwyyO5HAgqW94caap-ALkVjU688RpRgkZyLNZ5Y,5
 returnn/torch/util/README.md,sha256=AW-6ueWhgcwDcm57md6sm227QXNkvLnlRLwaH7NlS-w,193
 returnn/torch/util/__init__.py,sha256=AOXYUjzPm0XrzFJCPAXo9Jj_FvqD1XH3FfKtho80Vl8,26
 returnn/torch/util/array_.py,sha256=ell3VZvn01SLtF9Pw2fvPzFNO-XDQ7tSB9VCrVSKmSA,2556
-returnn/torch/util/debug_inf_nan.py,sha256=v0IzLy4kRKBWChSV70O4x829QtEuXMwB9mBqAyE4O2o,6223
+returnn/torch/util/debug_inf_nan.py,sha256=pXAHwgyn1aimLjD-XUblY2syBRCK0J20ioWgpvWfHvg,6400
 returnn/torch/util/diagnose_gpu.py,sha256=PYMmSk7iQ-jC3RXKKNXlYx1Q744C0LXqz0SB6ympwQg,5844
 returnn/torch/util/exception_helper.py,sha256=4e7YEf9D42aAUEkM3uSjnOxpNEYgtyPSpNV0-1L6PSU,4319
 returnn/torch/util/gradient_checkpoint.py,sha256=iLy-FB65DC8O6LxzmMvFjnSdpIVpko87ppIvRKAbtpQ,27995
@@ -253,8 +253,8 @@ returnn/util/sig_proc.py,sha256=Tjz0VOAVyqu2qDCF5HZ1JjALjcFsHcNkcd96WgZeKfE,7265
 returnn/util/task_system.py,sha256=y4sMVXQ25Qd2z0rx03uOlXlkE-jbCYC1Sjfn-XlraVU,26003
 returnn/util/train_proc_manager.py,sha256=Pjht28k6uz6BNQ47uW6Gf880iyq5q4wx7P_K2tmoAM8,3266
 returnn/util/watch_memory.py,sha256=BR5P2kvBN6UI81cE0_1WAA6Hd1SByLbBaiDxvLhPOew,4213
-returnn-1.20250225.201207.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250225.201207.dist-info/METADATA,sha256=MlICTMbISeiq6sz_1NI8XnxGZLlQzE0eH9VfGGDjlKs,5215
-returnn-1.20250225.201207.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-returnn-1.20250225.201207.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250225.201207.dist-info/RECORD,,
+returnn-1.20250226.115259.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250226.115259.dist-info/METADATA,sha256=2Ws--V5aicc3WJ-I6OrqPbbrvVNTH3Cnno6L7yeIyKY,5215
+returnn-1.20250226.115259.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+returnn-1.20250226.115259.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250226.115259.dist-info/RECORD,,