returnn 1.20240310.10417__tar.gz → 1.20240322.111713__tar.gz

{returnn-1.20240310.10417/returnn.egg-info → returnn-1.20240322.111713}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240310.10417
+Version: 1.20240322.111713
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240322.111713/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240322.111713'
+long_version = '1.20240322.111713+git.85ce153'

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/frontend/_backend.py

@@ -25,6 +25,7 @@ class Backend(Generic[T]):
     """
 
     # class attribs set by derived classes
+    name: Optional[str] = None  # e.g. "tensorflow" or "torch"
     RawTensorType: Type[T]
     is_tensorflow: bool = False  # whether this framework uses TensorFlow
     is_backend_raw_tensor_dim_tag_independent: bool = True  # whether raw tensors of backend are independent of Dim
@@ -1301,6 +1302,40 @@ global_backend = object.__new__(Backend)
 _backend_tensor_type_dispatch_table = {}  # type: Dict[Type, Type[Backend]]
 
 
+def select_backend(name: str):
+    """
+    Select backend by name.
+
+    :param name: "torch", "tf", "returnn_layers_tf", "numpy"
+    """
+    if name == "tf":
+        select_backend_tf()
+    elif name == "returnn_layers_tf":
+        select_backend_returnn_layers_tf()
+    elif name == "torch":
+        select_backend_torch()
+    elif name == "numpy":
+        from ._numpy_backend import NumpyBackend
+
+        global_backend.__class__ = NumpyBackend
+    else:
+        raise ValueError(f"unknown backend {name!r}")
+
+
+def get_selected_backend() -> Optional[str]:
+    """
+    :return: the selected backend name, or None if not selected
+    """
+    return global_backend.__class__.name
+
+
+def is_executing_eagerly() -> bool:
+    """
+    :return: whether the current selected backend is executing eagerly
+    """
+    return global_backend.executing_eagerly()
+
+
 def select_backend_tf():
     """
     Selects the RETURNN layers backend (based on TF).

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/frontend/_numpy_backend.py

@@ -17,6 +17,7 @@ from ._backend import Backend
 class NumpyBackend(Backend[numpy.ndarray]):
     """Numpy backend"""
 
+    name = "numpy"
     RawTensorType = numpy.ndarray
 
     @staticmethod

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/frontend/backend.py

@@ -7,9 +7,22 @@ from __future__ import annotations
 from . import _backend
 
 # And some functions from the internal backend API.
-from ._backend import select_backend_torch, select_backend_returnn_layers_tf
+from ._backend import (
+    select_backend,
+    get_selected_backend,
+    select_backend_torch,
+    select_backend_returnn_layers_tf,
+    is_executing_eagerly,
+)
 
-__all__ = ["select_backend_torch", "select_backend_returnn_layers_tf", "is_backend_raw_tensor_dim_tag_independent"]
+__all__ = [
+    "select_backend",
+    "get_selected_backend",
+    "select_backend_torch",
+    "select_backend_returnn_layers_tf",
+    "is_executing_eagerly",
+    "is_backend_raw_tensor_dim_tag_independent",
+]
 
 
 def is_backend_raw_tensor_dim_tag_independent() -> bool:

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/tf/frontend_layers/_backend.py

@@ -35,6 +35,7 @@ class ReturnnLayersBackend(Backend[Layer]):
     RETURNN layers backend (using TF), where raw_tensor represents a RETURNN layer
     """
 
+    name = "returnn_layers_tf"
     RawTensorType = Layer
     is_tensorflow = True
     is_backend_raw_tensor_dim_tag_independent = False
@@ -1241,7 +1242,9 @@ def _random_replay_eval(*, self, source, idx: int, **_kwargs):
         out.set_shape(self.output.batch_shape)
         return out
 
-    with tf.control_dependencies(
-        [source(i, auto_convert=False) for i in range(len(self.sources))]
-    ) if self.sources else contextlib.nullcontext():
+    with (
+        tf.control_dependencies([source(i, auto_convert=False) for i in range(len(self.sources))])
+        if self.sources
+        else contextlib.nullcontext()
+    ):
         return _func()

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/tf/frontend_low_level/_backend.py

@@ -27,6 +27,7 @@ class TFBackend(Backend[tf.Tensor]):
     TensorFlow low-level backend, operating on tf.Tensor
     """
 
+    name = "tf"
     RawTensorType = tf.Tensor
     is_tensorflow = True
 

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/returnn/torch/frontend/_backend.py

@@ -34,6 +34,7 @@ class TorchBackend(Backend[torch.Tensor]):
     PyTorch backend
     """
 
+    name = "torch"
     RawTensorType = torch.Tensor
 
     @staticmethod

{returnn-1.20240310.10417 → returnn-1.20240322.111713/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240310.10417
+Version: 1.20240322.111713
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240310.10417 → returnn-1.20240322.111713}/tools/torch_avg_checkpoints.py

@@ -111,7 +111,7 @@ def merge_checkpoints(in_ckpts: Sequence[str], out_ckpt: str, extra_state: Optio
 
     # Average
     for k in out_model_state:
-        out_model_state[k] /= out_model_state_num[k]
+        out_model_state[k] = out_model_state[k] / out_model_state_num[k]
 
     if extra_state:
         out_state.update(extra_state)

returnn-1.20240310.10417/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240310.010417'
-long_version = '1.20240310.010417+git.3bd8b43'