returnn 1.20250105.184816__tar.gz → 1.20250108.104903__tar.gz

{returnn-1.20250105.184816/returnn.egg-info → returnn-1.20250108.104903}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250105.184816
+Version: 1.20250108.104903
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250108.104903/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250108.104903'
+long_version = '1.20250108.104903+git.fee68a8'

{returnn-1.20250105.184816 → returnn-1.20250108.104903}/returnn/frontend/parameter.py

@@ -4,12 +4,13 @@ Parameter / variable
 
 from __future__ import annotations
 from typing import Optional, Union, TypeVar, Sequence
+from contextlib import contextmanager
 from returnn.tensor import Tensor, Dim
 import returnn.frontend as rf
 from ._backend import global_backend as _global_backend
 
 
-__all__ = ["Parameter"]
+__all__ = ["Parameter", "set_parameter_copy_behavior_ctx"]
 
 
 T = TypeVar("T")
@@ -97,8 +98,11 @@ class Parameter(Tensor[T]):
         self.initial = initial  # use setter
 
     def __copy__(self):
+        if _copy_behavior == "share":
+            return self
         # Should return new copy. https://github.com/rwth-i6/returnn_common/pull/215#issuecomment-1269651064
         # Note that the values are *not* copied, but rather it will use the same param init scheme.
+        assert _copy_behavior == "copy_init"
         res = type(self)(
             dims=self.dims,
             dtype=self.dtype,
@@ -111,10 +115,14 @@ class Parameter(Tensor[T]):
         return res
 
     def __deepcopy__(self, memo=None):
+        if _copy_behavior == "share":
+            return self
+
         # Should return new copy. https://github.com/rwth-i6/returnn_common/pull/215#issuecomment-1269651064
         # Note that the values are *not* copied, but rather it will use the same param init scheme.
         from copy import deepcopy
 
+        assert _copy_behavior == "copy_init"
         res = self.__copy__()
         if isinstance(self.initial, rf.init.ParamInit):
             res.initial = deepcopy(self.initial, memo=memo)  # noqa
@@ -241,3 +249,27 @@ class Parameter(Tensor[T]):
     @non_critical_for_restore.setter
     def non_critical_for_restore(self, value: bool):
         self._non_critical_for_restore = value
+
+
+_copy_behavior = "copy_init"
+
+
+@contextmanager
+def set_parameter_copy_behavior_ctx(mode: str):
+    """
+    Set the copy behavior for parameters.
+
+    :param mode: "copy_init" or "share".
+        "copy_init" (default when not set): When copying a parameter, a new parameter will be created.
+            It is not a shared parameter.
+            It does not copy the values, but it just copies the param init scheme.
+        "share": The parameter will not be copied, it will just return itself.
+    """
+    global _copy_behavior
+    assert mode in ("copy_init", "share")
+    old = _copy_behavior
+    try:
+        _copy_behavior = mode
+        yield
+    finally:
+        _copy_behavior = old

{returnn-1.20250105.184816 → returnn-1.20250108.104903}/returnn/pretrain.py

@@ -30,35 +30,6 @@ class WrapEpochValue:
         return self.func(epoch=epoch)
 
 
-def find_pretrain_wrap_values(net_json):
-    """
-    See also :func:`Pretrain._resolve_wrapped_values`.
-    Recursively goes through dicts, tuples and lists.
-    This is a simple check to see if this is needed,
-    i.e. if there are any :class:`WrapEpochValue` used.
-
-    :param dict[str] net_json: network dict
-    :return: whether there is some :class:`WrapEpochValue` in it
-    :rtype: bool
-    """
-    assert isinstance(net_json, dict)
-
-    def _check(d):
-        if isinstance(d, WrapEpochValue):
-            return True
-        if isinstance(d, dict):
-            for k, v in sorted(d.items()):
-                if _check(v):
-                    return True
-        if isinstance(d, (tuple, list)):
-            for v in d:
-                if _check(v):
-                    return True
-        return False
-
-    return _check(net_json)
-
-
 class Pretrain:
     """
     Start with 1 hidden layers up to N hidden layers -> N pretrain steps -> N epochs (with repetitions == 1).
@@ -198,7 +169,6 @@ class Pretrain:
         """
         Resolves any :class:`WrapEpochValue` in the net dicts.
         Recursively goes through dicts, tuples and lists.
-        See also :func:`find_pretrain_wrap_values`.
         """
 
         # noinspection PyShadowingNames

{returnn-1.20250105.184816 → returnn-1.20250108.104903/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250105.184816
+Version: 1.20250108.104903
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250105.184816/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250105.184816'
-long_version = '1.20250105.184816+git.2c5e602'