returnn 1.20240713.231521__tar.gz → 1.20240714.161251__tar.gz

{returnn-1.20240713.231521 → returnn-1.20240714.161251}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240713.231521
+Version: 1.20240714.161251
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240714.161251/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240714.161251'
+long_version = '1.20240714.161251+git.2e54d6a'

{returnn-1.20240713.231521 → returnn-1.20240714.161251}/returnn/torch/frontend/_backend.py

@@ -516,6 +516,8 @@ class TorchBackend(Backend[torch.Tensor]):
         out.raw_tensor = torch.stack([s.copy_compatible_to_dims_raw(out_dims[1:]) for s in sources], dim=0)
         return out
 
+    _ActivationFuncMapping = {"log_sigmoid": torch.nn.functional.logsigmoid}
+
     @staticmethod
     def activation_raw(raw_tensor: torch.Tensor, func: str) -> torch.Tensor:
         """
@@ -524,11 +526,12 @@ class TorchBackend(Backend[torch.Tensor]):
         :return: raw tensor after activation
         """
         assert func in Backend._AllowedActivationFuncs
-        if hasattr(torch, func):
-            f = getattr(torch, func)
-        elif hasattr(torch.nn.functional, func):
-            f = getattr(torch.nn.functional, func)
-        else:
+        f = getattr(torch, func, None)
+        if not f:
+            f = getattr(torch.nn.functional, func, None)
+        if not f:
+            f = TorchBackend._ActivationFuncMapping.get(func)
+        if not f:
             raise ValueError(f"unknown activation function {func!r}")
         return f(raw_tensor)
 

{returnn-1.20240713.231521 → returnn-1.20240714.161251}/returnn/torch/util/gradient_checkpoint.py

@@ -26,6 +26,7 @@ from dataclasses import dataclass, field
 import contextlib
 from weakref import ref, WeakSet
 import threading
+import atexit
 
 import torch
 from torch.utils.weak import WeakTensorKeyDictionary  # needs Torch >=2.0.0
@@ -178,6 +179,8 @@ class gradient_checkpoint_scope:
                 self.exit_saved_tensors_hooks_scope()
 
     def __del__(self):
+        if _python_exit:
+            return
         # Note, be very careful what we do in __del__ because it might be called in a different thread!
         # Note that the __del__ will likely be called very late,
         # as the reference to the _Graph is kept alive until we used it for backprop,
@@ -220,6 +223,8 @@ class gradient_checkpoint_scope:
         return x
 
     def _tensor_del_hook(self):
+        if _python_exit:
+            return
         # Some of the relevant tensors got deleted.
         # If we are in the right thread, maybe we can do the cleanup now.
         self._maybe_exit_saved_tensors_hooks_scope()
@@ -601,3 +606,12 @@ def _custom_saved_tensors_hooks_call_callbacks():
             assert not _custom_saved_tensors_hooks_tls_ctx.callbacks and not _custom_saved_tensors_hooks_tls_ctx.stack
     finally:
         _custom_saved_tensors_hooks_tls_ctx.in_callback = False
+
+
+def _python_exit_handler():
+    global _python_exit
+    _python_exit = True
+
+
+_python_exit = False
+atexit.register(_python_exit_handler)

{returnn-1.20240713.231521 → returnn-1.20240714.161251}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240713.231521
+Version: 1.20240714.161251
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240713.231521 → returnn-1.20240714.161251}/tests/test_rf_math.py

@@ -92,3 +92,20 @@ def test_relu():
         out.mark_as_default_output(shape=(batch_dim, time_dim, in_dim))
 
     run_model(extern_data, lambda *, epoch, step: _Net(), _forward_step)
+
+
+def test_log_sigmoid():
+    time_dim = Dim(Tensor("time", [batch_dim], dtype="int32"))
+    in_dim = Dim(7, name="in")
+    extern_data = TensorDict(
+        {
+            "data": Tensor("data", [batch_dim, time_dim, in_dim], dtype="float32"),
+        }
+    )
+
+    # noinspection PyShadowingNames
+    def _forward_step(*, model: rf.Module, extern_data: TensorDict):
+        out = rf.log_sigmoid(extern_data["data"])
+        out.mark_as_default_output(shape=(batch_dim, time_dim, in_dim))
+
+    run_model(extern_data, lambda *, epoch, step: rf.Module(), _forward_step)

returnn-1.20240713.231521/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240713.231521'
-long_version = '1.20240713.231521+git.90f56c9'