returnn 1.20231211.133727__tar.gz → 1.20231212.184511__tar.gz

{returnn-1.20231211.133727/returnn.egg-info → returnn-1.20231212.184511}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231211.133727
+Version: 1.20231212.184511
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231212.184511/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20231212.184511'
+long_version = '1.20231212.184511+git.01b9e12'

{returnn-1.20231211.133727 → returnn-1.20231212.184511}/returnn/frontend/_backend.py

@@ -931,6 +931,17 @@ class Backend(Generic[T]):
         """where"""
         raise NotImplementedError
 
+    @staticmethod
+    def clip_by_value(
+        x: Tensor,
+        clip_value_min: Union[Tensor, rf.RawTensorTypes],
+        clip_value_max: Union[Tensor, rf.RawTensorTypes],
+        *,
+        allow_broadcast_all_sources: bool = False,
+    ) -> Tensor:
+        """clip by value"""
+        raise NotImplementedError
+
     @staticmethod
     def matmul(a: Tensor[T], b: Tensor[T], *, reduce: Union[Dim, Sequence[Dim]], use_mask: bool = True) -> Tensor[T]:
         """

{returnn-1.20231211.133727 → returnn-1.20231212.184511}/returnn/frontend/math_.py

@@ -38,6 +38,7 @@ __all__ = [
     "opt_logical_or",
     "maximum",
     "minimum",
+    "clip_by_value",
     "identity",
     "exp",
     "expm1",
@@ -357,6 +358,20 @@ def minimum(a: Tensor, b: Union[Tensor, _RawTensorTypes], *other_tensors) -> Ten
     return res
 
 
+def clip_by_value(
+    x: Tensor,
+    clip_value_min: Union[Tensor, _RawTensorTypes],
+    clip_value_max: Union[Tensor, _RawTensorTypes],
+    *,
+    allow_broadcast_all_sources: bool = False,
+) -> Tensor:
+    """clip by value"""
+    # noinspection PyProtectedMember
+    return x._raw_backend.clip_by_value(
+        x, clip_value_min, clip_value_max, allow_broadcast_all_sources=allow_broadcast_all_sources
+    )
+
+
 def identity(x: Tensor) -> Tensor:
     """
     Identity function. Just to have one canonical. Does nothing, returns the input.

{returnn-1.20231211.133727 → returnn-1.20231212.184511}/returnn/tf/frontend_layers/_backend.py

@@ -704,6 +704,27 @@ class ReturnnLayersBackend(Backend[Layer]):
             {"class": "switch", "condition": cond, "true_from": true_, "false_from": false_}, name="where"
         )
 
+    @staticmethod
+    def clip_by_value(
+        x: Tensor,
+        clip_value_min: Union[Tensor, rf.RawTensorTypes],
+        clip_value_max: Union[Tensor, rf.RawTensorTypes],
+        *,
+        allow_broadcast_all_sources: bool = False,
+    ) -> Tensor:
+        """clip by value"""
+        clip_value_min = rf.convert_to_tensor(clip_value_min, _backend=ReturnnLayersBackend)
+        clip_value_max = rf.convert_to_tensor(clip_value_max, _backend=ReturnnLayersBackend)
+        return rfl.make_layer(
+            {
+                "class": "eval",
+                "eval": "tf.clip_by_value(source(0), source(1), source(2))",
+                "from": [x, clip_value_min, clip_value_max],
+                "allow_broadcast_all_sources": allow_broadcast_all_sources,
+            },
+            name="clip_by_value",
+        )
+
     @staticmethod
     def matmul(a: Tensor, b: Tensor, *, reduce: Union[Dim, Sequence[Dim]], use_mask: bool = True) -> Tensor:
         """matmul"""

{returnn-1.20231211.133727 → returnn-1.20231212.184511}/returnn/torch/frontend/_backend.py

@@ -1045,13 +1045,39 @@ class TorchBackend(Backend[torch.Tensor]):
             [true_, false_, cond], allow_broadcast_all_sources=allow_broadcast_all_sources, name="where"
         )
         out.dtype = true_.dtype
-        out.sparse_dim = true_.sparse_dim
+        out.sparse_dim = true_.sparse_dim or false_.sparse_dim
+        out.feature_dim = true_.feature_dim or false_.feature_dim
         cond_bc_raw = cond.copy_compatible_to_dims_raw(out.dims)
         true_bc_raw = true_.copy_compatible_to_dims_raw(out.dims)
         false_bc_raw = false_.copy_compatible_to_dims_raw(out.dims)
         out.raw_tensor = torch.where(cond_bc_raw, true_bc_raw, false_bc_raw)
         return out
 
+    @staticmethod
+    def clip_by_value(
+        x: Tensor,
+        clip_value_min: Union[Tensor, rf.RawTensorTypes],
+        clip_value_max: Union[Tensor, rf.RawTensorTypes],
+        *,
+        allow_broadcast_all_sources: bool = False,
+    ) -> Tensor:
+        """clip by value"""
+        clip_value_min = rf.convert_to_tensor(clip_value_min, _backend=TorchBackend, device=x.device)
+        clip_value_max = rf.convert_to_tensor(clip_value_max, _backend=TorchBackend, device=x.device)
+        out = Tensor.get_common_data(
+            [x, clip_value_min, clip_value_max],
+            allow_broadcast_all_sources=allow_broadcast_all_sources,
+            name="clip_by_value",
+        )
+        out.dtype = x.dtype
+        out.sparse_dim = x.sparse_dim
+        out.feature_dim = x.feature_dim
+        x_bc_raw = x.copy_compatible_to_dims_raw(out.dims)
+        min_bc_raw = clip_value_min.copy_compatible_to_dims_raw(out.dims)
+        max_bc_raw = clip_value_max.copy_compatible_to_dims_raw(out.dims)
+        out.raw_tensor = torch.clamp(x_bc_raw, min_bc_raw, max_bc_raw)
+        return out
+
     @staticmethod
     def matmul(a: _TT, b: _TT, *, reduce: Union[Dim, Sequence[Dim]], use_mask: bool = True) -> _TT:
         """

{returnn-1.20231211.133727 → returnn-1.20231212.184511/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231211.133727
+Version: 1.20231212.184511
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231211.133727/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20231211.133727'
-long_version = '1.20231211.133727+git.1a65545'