returnn 1.20230519.110838__tar.gz → 1.20230519.123334__tar.gz

{returnn-1.20230519.110838/returnn.egg-info → returnn-1.20230519.123334}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20230519.110838
+Version: 1.20230519.123334
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20230519.123334/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20230519.123334'
+long_version = '1.20230519.123334+git.6ca9372'

{returnn-1.20230519.110838 → returnn-1.20230519.123334}/returnn/torch/engine.py

@@ -22,6 +22,7 @@ from returnn.tensor import TensorDict, Tensor
 from returnn.datasets.basic import init_dataset, Dataset
 from returnn.util import basic as util
 from returnn.util import NumbersDict
+from returnn.util.basic import NotSpecified
 from .updater import Updater
 from .data import pipeline as data_pipeline
 from .data import returnn_dataset_wrapper
@@ -64,10 +65,9 @@ class Engine(EngineBase):
         print("Using device:", self._device, file=log.v2)
 
         amp_options = self.config.typed_value("torch_amp")
-        grad_scaler_opts = None
+        grad_scaler_opts = self.config.typed_value("grad_scaler", NotSpecified)
         if amp_options is not None:
             self._use_autocast = True
-            grad_scaler_opts = {}
             if isinstance(amp_options, dict):
                 amp_options = util.CollectionReadCheckCovered(amp_options)
                 dtype = amp_options.get("dtype", None)
@@ -80,12 +80,14 @@ class Engine(EngineBase):
             if isinstance(dtype, str):
                 dtype = getattr(torch, dtype)
             assert isinstance(dtype, torch.dtype) or dtype is None
+            print(f"Using autocast (automatic mixed precision (AMP)) with dtype {dtype}", file=log.v2)
             self._autocast_dtype = dtype
 
-        if grad_scaler_opts is None:
-            grad_scaler_opts = self.config.typed_value("grad_scaler")
+        if grad_scaler_opts is NotSpecified:
+            grad_scaler_opts = {} if self._use_autocast else None
         if grad_scaler_opts is not None:
             assert isinstance(grad_scaler_opts, dict)
+            print("Using GradScaler with options:", grad_scaler_opts, file=log.v2)
             self._grad_scaler = amp.GradScaler(**grad_scaler_opts)
 
     def init_train_from_config(

{returnn-1.20230519.110838 → returnn-1.20230519.123334/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20230519.110838
+Version: 1.20230519.123334
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20230519.110838/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20230519.110838'
-long_version = '1.20230519.110838+git.907efda'