returnn 1.20250819.10249__py3-none-any.whl → 1.20250820.171158__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250819.10249
+Version: 1.20250820.171158
 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 = '1.20250819.010249'
-long_version = '1.20250819.010249+git.9c1f159'
+version = '1.20250820.171158'
+long_version = '1.20250820.171158+git.d60d270'

returnn/torch/updater.py

@@ -5,11 +5,10 @@ and model param update logic in general.
 
 from __future__ import annotations
 
-from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Iterator, Set, Dict, List, Tuple
+from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Set, Dict, List, Tuple
 import os
 import gc
 import torch
-import typing
 
 import returnn
 from returnn.log import log
@@ -130,8 +129,8 @@ class Updater:
             else:
                 raise NotImplementedError("not implemented for not callable dynamic_learning_rate")
 
-        self._optimizer_opts = None
-        self.optimizer = None  # type: typing.Optional[torch.optim.Optimizer]
+        self._optimizer_opts: Optional[Dict[str, Any]] = None
+        self.optimizer: Optional[torch.optim.Optimizer] = None
 
         self._grad_clip = self.config.float("gradient_clip", 0.0)
         self._grad_clip_global_norm = self.config.float("gradient_clip_global_norm", 0.0)
@@ -481,7 +480,7 @@ class Updater:
 
     def _get_optimizer_param_groups(
         self, optim_class: Type[torch.optim.Optimizer], optimizer_opts: Dict[str, Any]
-    ) -> Union[List[Dict[str, Any]], Iterator[torch.nn.Parameter]]:
+    ) -> Union[Iterable[Dict[str, Any]], Iterable[torch.nn.Parameter]]:
         """
         The weight_decay parameter from AdamW affects the weights of layers such as LayerNorm and Embedding.
         This function creates a blacklist of network modules and splits the optimizer groups in two:
@@ -514,10 +513,17 @@ class Updater:
         if custom_param_groups is not None:
             assert callable(custom_param_groups), f"invalid param_groups_custom {custom_param_groups!r}"
             rf_model = wrapped_pt_module_to_rf_module(self.network)
-            custom_param_groups = custom_param_groups(
-                model=self.network, rf_model=rf_model, optimizer_class=optim_class, optimizer_opts=optimizer_opts
+            custom_param_groups_ = custom_param_groups(
+                model=self.network,
+                rf_model=rf_model,
+                optimizer_class=optim_class,
+                optimizer_opts=optimizer_opts,
+                **get_fwd_compat_kwargs(),
             )
-            return custom_param_groups
+            assert isinstance(custom_param_groups_, Iterable) and all(
+                isinstance(group, dict) for group in custom_param_groups_
+            ), f"invalid param_groups_custom {custom_param_groups!r} result {custom_param_groups_!r} type"
+            return custom_param_groups_
 
         network_params = self.network.parameters()
 
@@ -545,7 +551,7 @@ class Updater:
         # Parameters without weight decay: biases + LayerNorm/Embedding layers.
         wd_params = set()
         no_wd_params = set()
-        blacklist_wd_modules = optimizer_opts.pop("weight_decay_modules_blacklist", None)
+        blacklist_wd_modules: Any = optimizer_opts.pop("weight_decay_modules_blacklist", None)
         if blacklist_wd_modules is None:
             blacklist_wd_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
         else:

returnn/util/basic.py

@@ -2459,8 +2459,12 @@ class DictRefKeys(Generic[K, V]):
     Like `dict`, but hash and equality of the keys
     """
 
-    def __init__(self):
+    def __init__(self, items: Union[None, Iterable[Tuple[K, V]], Dict[K, V]] = None, /, **kwargs):
         self._d = {}  # type: Dict[RefIdEq[K], V]
+        if items is not None:
+            self.update(items)
+        if kwargs:
+            self.update(kwargs)
 
     def __repr__(self):
         return "DictRefKeys(%s)" % ", ".join(["%r: %r" % (k, v) for (k, v) in self.items()])
@@ -2489,6 +2493,15 @@ class DictRefKeys(Generic[K, V]):
     def __contains__(self, item: K):
         return RefIdEq(item) in self._d
 
+    def update(self, other: Union[Dict[K, V], Iterable[Tuple[K, V]]], /):
+        """
+        :param other: dict or iterable of (key, value) tuples
+        """
+        if isinstance(other, dict):
+            other = other.items()
+        for k, v in other:
+            self[k] = v
+
 
 def make_dll_name(basename):
     """

{returnn-1.20250819.10249.dist-info → returnn-1.20250820.171158.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250819.10249
+Version: 1.20250820.171158
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250819.10249.dist-info → returnn-1.20250820.171158.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=40ciCZzddEgWfHHnfFmRo7cpK8dukyBH8HYxTaEd5XY,5214
+returnn/PKG-INFO,sha256=Cbw-LFRDg3cxVzUdgT7yaNpnAUDTjOw5nRYgq8jjH8A,5215
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=lHyZcu_0yc9f7Vf_Kfdy9PmeU0T76XVXnpalHi5WKro,31740
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=poqBmOb1nT6ZUEe9pGqrp91VojHENemIF3zMYt80T4g,77
+returnn/_setup_info_generated.py,sha256=OQkFtzn37F7h2FgPXE84vkON338V4uRqbMen7VsZWR8,77
 returnn/config.py,sha256=3tmKhB6FnQZaNdtcYsiB61JnEY--iZ2qmJ4yq0b6tE0,29140
 returnn/forward_iface.py,sha256=A_OJiaXsX4MlXQRzST86ylyxSUZbC402PQL1REcqHjM,911
 returnn/learning_rate_control.py,sha256=ZvWryAn_tv9DhV8sh1LV3eE34Yltl3On3mYZAG4hR9s,34684
@@ -208,7 +208,7 @@ returnn/torch/README.md,sha256=jzJ2FpOHW02vxN69yKaV97C9LI-hmvjBglKfdZXIDdc,85
 returnn/torch/__init__.py,sha256=MHEUyNHB20Vy89uKAqZoj6FxJKF1Gq3HW-i6ra1pNcI,24
 returnn/torch/distributed.py,sha256=_lyJR71HIoCHpMi5GztGM7YwrX54Am8zSkjnDkE1Lbk,7524
 returnn/torch/engine.py,sha256=JSsQZZiVs9TxRyFEJuR3iH-YZb9sRw7TzoIAIqmplZY,78275
-returnn/torch/updater.py,sha256=Vyh5w6ZFVc1hQvyyoWpeienQdlBVLZ2HYfjFZRQB3cQ,30035
+returnn/torch/updater.py,sha256=-v_uY-8jDhreXPmjJYR4cgrlW_7ZI4kt2X2xIZdX_DE,30377
 returnn/torch/data/__init__.py,sha256=6cLNEi8KoGI12PF6akN7mI_mtjlx-0hcQAfMYoExwik,132
 returnn/torch/data/extern_data.py,sha256=5al706ZaYtHWLp5VH2vS-rW69YXP3NHyOFRKY0WY714,7810
 returnn/torch/data/pipeline.py,sha256=HgIL0jQsPcgvh_SPC4wQ6BzclmrnpFja-UiboF_GPN4,29459
@@ -233,7 +233,7 @@ returnn/torch/util/gradient_checkpoint.py,sha256=iLy-FB65DC8O6LxzmMvFjnSdpIVpko8
 returnn/torch/util/module.py,sha256=MXHIrF9Isu575DDJIa81212ULKwdqu1oOLxDVZecVSk,1693
 returnn/torch/util/scaled_gradient.py,sha256=C5e79mpqtxdtw08OTSy413TSBSlOertRisc-ioiFIaU,3191
 returnn/util/__init__.py,sha256=UIG1qw4idqhW71BV60ha7h9PktxvEVcBIu0lYRossK8,336
-returnn/util/basic.py,sha256=9Ig-7XLtvXk3yfycmBEhdJG-WVNDtoND3DmDyXOl018,142627
+returnn/util/basic.py,sha256=UjHujX9pSu_dOgTxozWD0ujj5eSpyj_zD5vFU6bfyms,143096
 returnn/util/better_exchook.py,sha256=39yvRecluDgYhViwSkaQ8crJ_cBWI63KeEGuK4RKe5w,70843
 returnn/util/bpe.py,sha256=LWFhICZsEOnMwNws0lybPNzKRX6rSr8yKCvP65vjl9Y,19656
 returnn/util/debug.py,sha256=wuRzdg9zB84WWCGyTjmRR_zYypu8gXxlc0nZ6si9OC8,28224
@@ -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-1.20250819.10249.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250819.10249.dist-info/METADATA,sha256=40ciCZzddEgWfHHnfFmRo7cpK8dukyBH8HYxTaEd5XY,5214
-returnn-1.20250819.10249.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-returnn-1.20250819.10249.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250819.10249.dist-info/RECORD,,
+returnn-1.20250820.171158.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250820.171158.dist-info/METADATA,sha256=Cbw-LFRDg3cxVzUdgT7yaNpnAUDTjOw5nRYgq8jjH8A,5215
+returnn-1.20250820.171158.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+returnn-1.20250820.171158.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250820.171158.dist-info/RECORD,,