returnn 1.20241113.92146__tar.gz → 1.20241113.145057__tar.gz

{returnn-1.20241113.92146 → returnn-1.20241113.145057}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241113.92146
+Version: 1.20241113.145057
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241113.145057/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20241113.145057'
+long_version = '1.20241113.145057+git.e22a5d0'

{returnn-1.20241113.92146 → returnn-1.20241113.145057}/returnn/torch/updater.py

@@ -5,7 +5,7 @@ and model param update logic in general.
 
 from __future__ import annotations
 
-from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Set, Dict, List, Tuple
+from typing import Optional, Union, Any, Type, Callable, Sequence, Iterable, Iterator, Set, Dict, List, Tuple
 import os
 import gc
 import torch
@@ -428,8 +428,8 @@ class Updater:
         lr = lr * opt_kwargs.pop("learning_rate_multiplier", 1.0)
         opt_kwargs["lr"] = lr
 
-        param_groups = self._get_optimizer_param_groups(optim_class, opt_kwargs)
-        optimizer = optim_class(param_groups, **opt_kwargs)
+        params_or_param_groups = self._get_optimizer_param_groups(optim_class, opt_kwargs)
+        optimizer = optim_class(params_or_param_groups, **opt_kwargs)
         print("Optimizer: %s" % optimizer, file=log.v1)
         assert isinstance(optimizer, torch.optim.Optimizer)
 
@@ -447,7 +447,7 @@ class Updater:
 
     def _get_optimizer_param_groups(
         self, optim_class: Type[torch.optim.Optimizer], optimizer_opts: Dict[str, Any]
-    ) -> List[Dict[str, Any]]:
+    ) -> Union[List[Dict[str, Any]], Iterator[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:
@@ -502,11 +502,11 @@ class Updater:
             ), "weight_decay not accepted by the chosen optimizer. Accepted values: %s" % ", ".join(
                 "%s" % optim_name for optim_name in cls_init_kwargs
             )
-            return [{"params": network_params}]
+            return network_params
 
         weight_decay = optimizer_opts.get("weight_decay", 0.0)
         if not weight_decay:
-            return [{"params": network_params}]
+            return network_params
 
         # Distinguish between parameters with and without weight_decay/L2 regularization.
         # Parameters without weight decay: biases + LayerNorm/Embedding layers.

{returnn-1.20241113.92146 → returnn-1.20241113.145057}/returnn/util/multi_proc_non_daemonic_spawn.py

@@ -196,29 +196,34 @@ class _AtExitCleanupProcess:
         self.proc_pid = proc_pid
 
     def __call__(self):
-        if os.getpid() != self.cur_pid:  # e.g. in fork
-            return  # ignore
-        if self.proc_pid is None:  # already cleaned
-            return
-        # First try SIGINT. See NonDaemonicSpawnProcess docstring.
-        # Then, if that fails, try the others.
-        signals = ["SIGINT"] * 3 + ["SIGTERM", "SIGKILL"]
-        for sig in signals:
-            # The proc might have been killed by some other code. That's ok.
-            try:
-                print(f"Send signal {sig} to pid {self.proc_pid}")
-                os.kill(self.proc_pid, getattr(signal, sig))
-            except ProcessLookupError:
-                break
-            if self._waitpid():
-                break
-            time.sleep(1)
-            if self._waitpid():
-                break
-            time.sleep(4)
-            if self._waitpid():
-                break
-        self.proc_pid = None
+        try:
+            if os.getpid() != self.cur_pid:  # e.g. in fork
+                return  # ignore
+            if self.proc_pid is None:  # already cleaned
+                return
+            # First try SIGINT. See NonDaemonicSpawnProcess docstring.
+            # Then, if that fails, try the others.
+            signals = ["SIGINT"] * 3 + ["SIGTERM", "SIGKILL"]
+            for sig in signals:
+                # The proc might have been killed by some other code. That's ok.
+                try:
+                    print(f"Send signal {sig} to pid {self.proc_pid}")
+                    os.kill(self.proc_pid, getattr(signal, sig))
+                except ProcessLookupError:
+                    break
+                if self._waitpid():
+                    break
+                time.sleep(1)
+                if self._waitpid():
+                    break
+                time.sleep(4)
+                if self._waitpid():
+                    break
+            self.proc_pid = None
+        except KeyboardInterrupt:  # someone is killing us. let it happen
+            pass
+        except RuntimeError:  # e.g. PyTorch SIGCHLD handler does this. just quit
+            pass
 
     def _waitpid(self) -> bool:
         try:

{returnn-1.20241113.92146 → returnn-1.20241113.145057}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241113.92146
+Version: 1.20241113.145057
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241113.92146/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20241113.092146'
-long_version = '1.20241113.092146+git.55cdd2f'