returnn 1.20240712.3448__tar.gz → 1.20240712.22135__tar.gz

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240712.3448
+Version: 1.20240712.22135
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240712.22135/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240712.022135'
+long_version = '1.20240712.022135+git.3da9c4b'

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/returnn/frontend/conv.py

@@ -4,6 +4,7 @@ Convolution, transposed convolution, pooling
 
 from __future__ import annotations
 from typing import Optional, Sequence, Tuple, Union
+from returnn.util.basic import next_type_attrib_in_mro_chain
 from returnn.tensor import Tensor, Dim
 import returnn.frontend as rf
 
@@ -98,7 +99,7 @@ class _ConvOrTransposedConv(rf.Module):
         out_spatial_dim: Optional[Dim] = None,
     ) -> Tuple[Tensor, Dim]:
         assert self.nd == 1
-        out, (out_spatial_dim,) = self.__class__.__base__.__call__(
+        out, (out_spatial_dim,) = next_type_attrib_in_mro_chain(self.__class__, "__call__", self.__class__._call_nd1)(
             self,
             source,
             in_spatial_dims=[in_spatial_dim],

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/returnn/frontend/parametrize.py

@@ -48,10 +48,7 @@ def register_parametrization(
         # (see https://docs.python.org/3/howto/descriptor.html)
         # and also we do not want to modify the original class.
         cls = module.__class__
-        param_cls = _new_classes.get(cls, None)
-        if not param_cls:
-            param_cls = _Metaclass(f"Parametrized{cls.__name__}", (cls,), {})
-            _new_classes[cls] = param_cls
+        param_cls = _Metaclass(f"Parametrized{cls.__name__}", (cls,), {})
         module.__class__ = param_cls
 
     if hasattr(module.__class__, param_name):
@@ -153,9 +150,6 @@ _ParametrizationType = Union[
 ]
 
 
-_new_classes: weakref.WeakKeyDictionary[type, type] = weakref.WeakKeyDictionary()
-
-
 class _Metaclass(type):
     """
     https://stackoverflow.com/a/75943813/133374

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/returnn/util/basic.py

@@ -2396,6 +2396,31 @@ def help_on_type_error_wrong_args(cls, kwargs):
         print("Args mismatch? Missing are %r, unknowns are %r. Kwargs %r." % (mandatory_args, unknown_args, kwargs))
 
 
+def type_attrib_mro_chain(cls: type, attr_name: str) -> list:
+    """
+    :return: list of all attributes with the given name in the MRO chain
+    """
+    attribs = []
+    for cls_ in cls.mro():
+        if attr_name in cls_.__dict__:
+            attribs.append(cls_.__dict__[attr_name])
+    return attribs
+
+
+def next_type_attrib_in_mro_chain(cls: type, attr_name: str, attr):
+    """
+    :param cls:
+    :param attr_name:
+    :param attr: must be in the attrib MRO chain
+    :return: next attribute in the MRO chain
+    """
+    attribs = type_attrib_mro_chain(cls, attr_name)
+    assert attr in attribs
+    idx = attribs.index(attr)
+    assert idx + 1 < len(attribs)
+    return attribs[idx + 1]
+
+
 def custom_exec(source: str, source_filename: str, user_ns: Dict[str, Any], user_global_ns: Dict[str, Any]):
     """
     :param source:

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240712.3448
+Version: 1.20240712.22135
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/tests/test_Util.py

@@ -376,6 +376,23 @@ def test_get_func_kwargs():
     assert_equal(list(getargspec(dummy_func).args), ["net", "var", "update_ops"])
 
 
+def test_next_type_attrib_in_mro_chain():
+    class Base:
+        def method(self):
+            return id(self)
+
+    class Foo(Base):
+        def method(self):
+            return 2
+
+    class Bar(Foo):
+        pass
+
+    assert_equal(type_attrib_mro_chain(Foo, "method"), [Foo.method, Base.method])
+    assert_equal(type_attrib_mro_chain(Bar, "method"), [Foo.method, Base.method])
+    assert_equal(next_type_attrib_in_mro_chain(Bar, "method", Foo.method), Base.method)
+
+
 def test_simple_obj_repr():
     class X:
         def __init__(self, a, b=13):

{returnn-1.20240712.3448 → returnn-1.20240712.22135}/tests/test_rf_base.py

@@ -504,3 +504,22 @@ def test_weight_noise():
     assert set(params.keys()) == {"weight", "bias"}
     assert params["weight"] is orig_weight
     assert params["bias"] is orig_bias
+
+    class _Model(rf.Module):
+        def __init__(self):
+            super().__init__()
+            self.linear = rf.Linear(in_dim, out_dim)
+
+    model = _Model()
+    for mod in model.modules():
+        for param_name, param in mod.named_parameters(recurse=False):
+            if param_name.endswith("bias"):  # no bias
+                continue
+            print("***", mod, param_name, param)
+            rf.weight_noise(mod, param_name, std=0.0025)
+
+    # https://github.com/rwth-i6/returnn/issues/1580
+    conv = rf.Conv1d(in_dim, out_dim, 3, padding="same")
+    rf.weight_noise(conv, "filter", std=0.1)
+    time_dim = Dim(11, name="time")
+    conv(rf.random_normal([time_dim, in_dim]), in_spatial_dim=time_dim)

returnn-1.20240712.3448/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240712.003448'
-long_version = '1.20240712.003448+git.f0a1a4f'