returnn 1.20230526.180133__tar.gz → 1.20230530.134944__tar.gz

{returnn-1.20230526.180133/returnn.egg-info → returnn-1.20230530.134944}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20230526.180133
+Version: 1.20230530.134944
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20230530.134944/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20230530.134944'
+long_version = '1.20230530.134944+git.cf36f68'

{returnn-1.20230526.180133 → returnn-1.20230530.134944}/returnn/torch/engine.py

@@ -332,30 +332,9 @@ class Engine(EngineBase):
         """
         :param dict[str, torch.Tensor] extern_data_raw: model inputs for the step
         """
-        assert isinstance(extern_data_raw, dict) and extern_data_raw
-        batch_dim = next(iter(self.extern_data.data.values())).dims[0]
-        batch_dim.dyn_size_ext = None  # if it is dynamic, reset now, and set it below
-        extern_data = TensorDict()
-        for k, data in self.extern_data.data.items():
-            data = data.copy_template()
-            raw_tensor = extern_data_raw[k].to(self._device)
-            data.dtype = str(raw_tensor.dtype).split(".")[-1]  # just overwrite for now...
-            data.raw_tensor = raw_tensor
-
-            if batch_dim.size is None and batch_dim.dyn_size_ext is None:
-                batch_dim.dyn_size_ext = Tensor(batch_dim.name or "batch", dims=[], dtype="int32")
-                batch_dim.dyn_size_ext.raw_tensor = torch.tensor(extern_data_raw[k].shape[0], dtype=torch.int32)
-
-            if k + ":seq_len" in extern_data_raw:
-                # Sequence lengths have to be on CPU for the later call to rnn.pack_padded_sequence
-                size = extern_data_raw[k + ":seq_len"].cpu()
-                size_dtype = str(size.dtype).split(".")[-1]
-                if data.dims[1].dyn_size_ext is None:
-                    data.dims[1].dyn_size_ext = Tensor(data.dims[1].name or "time", dims=[batch_dim], dtype=size_dtype)
-                data.dims[1].dyn_size_ext.dtype = size_dtype
-                data.dims[1].dyn_size_ext.raw_tensor = size
-
-            extern_data.data[k] = data
+        extern_data = _raw_dict_to_extern_data(
+            extern_data_raw, extern_data_template=self.extern_data, device=self._device
+        )
 
         rf.init_train_step_run_ctx(train_flag=train_flag)
 
@@ -509,6 +488,43 @@ def _to_raw(n: Union[int, float, Tensor]):
     raise TypeError(f"Unexpected {n} of type {type(n)}")
 
 
+def _raw_dict_to_extern_data(
+    extern_data_raw: Dict[str, torch.Tensor], *, extern_data_template: TensorDict, device: Union[str, torch.device]
+) -> TensorDict:
+    """
+    :param extern_data_raw: This comes out of the DataLoader.
+    :param extern_data_template: Specified via `extern_data` in the config.
+    :param device: E.g. the GPU.
+    :return: tensor dict, like extern_data_template, but with raw tensors set to Torch tensors, on the right device.
+    """
+    assert isinstance(extern_data_raw, dict) and extern_data_raw
+    batch_dim = next(iter(extern_data_template.data.values())).dims[0]
+    batch_dim.dyn_size_ext = None  # if it is dynamic, reset now, and set it below
+    extern_data = TensorDict()
+    for k, data in extern_data_template.data.items():
+        data = data.copy_template()
+        raw_tensor = extern_data_raw[k].to(device)
+        data.dtype = str(raw_tensor.dtype).split(".")[-1]  # just overwrite for now...
+        data.raw_tensor = raw_tensor
+
+        if batch_dim.size is None and batch_dim.dyn_size_ext is None:
+            batch_dim.dyn_size_ext = Tensor(batch_dim.name or "batch", dims=[], dtype="int32")
+            batch_dim.dyn_size_ext.raw_tensor = torch.tensor(extern_data_raw[k].shape[0], dtype=torch.int32)
+
+        if k + ":seq_len" in extern_data_raw:
+            # Sequence lengths have to be on CPU for the later call to rnn.pack_padded_sequence
+            size = extern_data_raw[k + ":seq_len"].cpu()
+            size_dtype = str(size.dtype).split(".")[-1]
+            if data.dims[1].dyn_size_ext is None:
+                data.dims[1].dyn_size_ext = Tensor(data.dims[1].name or "time", dims=[batch_dim], dtype=size_dtype)
+            data.dims[1].dyn_size_ext.dtype = size_dtype
+            data.dims[1].dyn_size_ext.raw_tensor = size
+
+        extern_data.data[k] = data
+
+    return extern_data
+
+
 def _print_process(report_prefix, step, eval_info):
     """
     Similar but simplified from TF engine _print_process.

{returnn-1.20230526.180133 → returnn-1.20230530.134944/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20230526.180133
+Version: 1.20230530.134944
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20230526.180133 → returnn-1.20230530.134944}/tests/test_demos.py

@@ -153,7 +153,6 @@ def test_demo_torch_task12ax():
 def test_demo_rf_torch_task12ax():
     cleanup_tmp_models("demos/demo-rf.config")
     out = run(py, "rnn.py", "demos/demo-rf.config", print_stdout=True)
-    # Currently this just uses linear layers, so it's not very good.
     # Also see test_demo_tf_task12ax above.
     fer = parse_last_fer(out)
     assert_less(fer, 0.02)
@@ -163,7 +162,6 @@ def test_demo_rf_torch_task12ax():
 def test_demo_rf_tf_task12ax():
     cleanup_tmp_models("demos/demo-rf.config")
     out = run(py, "rnn.py", "demos/demo-rf.config", "++backend", "tensorflow-net-dict", print_stdout=True)
-    # Currently this just uses linear layers, so it's not very good.
     # Also see test_demo_tf_task12ax above.
     fer = parse_last_fer(out)
     assert_less(fer, 0.02)

returnn-1.20230526.180133/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20230526.180133'
-long_version = '1.20230526.180133+git.327f43c'