returnn 1.20231220.104729__tar.gz → 1.20231220.144533__tar.gz

{returnn-1.20231220.104729/returnn.egg-info → returnn-1.20231220.144533}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231220.104729
+Version: 1.20231220.144533
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231220.144533/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20231220.144533'
+long_version = '1.20231220.144533+git.da66b62'

{returnn-1.20231220.104729 → returnn-1.20231220.144533}/returnn/frontend/decoder/transformer.py

@@ -27,6 +27,7 @@ class TransformerDecoder(rf.Module):
 
     def __init__(
         self,
+        encoder_dim: Dim,
         vocab_dim: Dim,
         model_dim: Dim = Dim(512, name="transformer-dec-default-model-dim"),
         *,
@@ -41,6 +42,7 @@ class TransformerDecoder(rf.Module):
         share_embedding: bool = False,
     ):
         """
+        :param encoder_dim:
         :param vocab_dim:
         :param model_dim: the output feature dimension
         :param num_layers: the number of encoder layers
@@ -55,6 +57,7 @@ class TransformerDecoder(rf.Module):
         """
         super().__init__()
 
+        self.encoder_dim = encoder_dim
         self.vocab_dim = vocab_dim
         self.model_dim = model_dim
 
@@ -69,6 +72,7 @@ class TransformerDecoder(rf.Module):
 
         if not decoder_layer or isinstance(decoder_layer, type):
             decoder_layer_opts_ = dict(
+                encoder_dim=encoder_dim,
                 out_dim=model_dim,
                 ff_dim=ff_dim,
                 ff_activation=ff_activation,
@@ -117,13 +121,14 @@ class TransformerDecoder(rf.Module):
         collected_outputs: Optional[Dict[str, Tensor]] = None,
     ) -> Tuple[Tensor, rf.State]:
         """
-        forward, single step or whole sequence
+        forward, single step or whole sequence.
 
-        :param source:
+        :param source: labels
         :param spatial_dim: single_step_dim or spatial dim of source
         :param state: e.g. via :func:`default_initial_state`
         :param encoder: via :func:`transform_encoder`
         :param collected_outputs:
+        :return: logits, new state
         """
         new_state = rf.State()
 
@@ -255,7 +260,14 @@ class FeedForward(rf.Module):
         FF -> Activation -> Dropout -> FF
     """
 
-    def __init__(self, out_dim: Dim, *, ff_dim: Dim, dropout: float, activation: Callable[[Tensor], Tensor]):
+    def __init__(
+        self,
+        out_dim: Dim,
+        *,
+        ff_dim: Optional[Dim] = NotSpecified,
+        dropout: float,
+        activation: Callable[[Tensor], Tensor],
+    ):
         """
         :param out_dim: output feature dimension
         :param ff_dim: dimension of the feed-forward layers
@@ -264,6 +276,9 @@ class FeedForward(rf.Module):
         """
         super().__init__()
 
+        if ff_dim is NotSpecified:
+            ff_dim = out_dim * 4
+
         self.out_dim = out_dim
         self.dropout = dropout
         self.dropout_broadcast = rf.dropout_broadcast_default()

{returnn-1.20231220.104729 → returnn-1.20231220.144533}/returnn/util/watch_memory.py

@@ -36,7 +36,8 @@ def _watch_memory_main(pid: int):
             f.write(f"watch memory")
 
     def _print(*args):
-        print(f"[{datetime.now().strftime('%Y-%m-%d %H:%M:%S')} pid:{cur_proc.pid}] MEMORY:", *args)
+        print("MEMORY:", *args)
+        sys.stdout.flush()
 
     cur_proc = psutil.Process(pid)
     procs = []
@@ -81,7 +82,11 @@ def _watch_memory_main(pid: int):
             for mem_info in mem_per_pid.values():
                 for k in res.keys():
                     res[k] += mem_info[k]
-            _print(f"total ({len(mem_per_pid)} procs):", _format_mem_info(res))
+            _print(
+                "total",
+                f"(main {cur_proc.pid}, {datetime.now().strftime('%Y-%m-%d, %H:%M:%S')}, {len(mem_per_pid)} procs):",
+                _format_mem_info(res),
+            )
 
         time.sleep(5)
 

{returnn-1.20231220.104729 → returnn-1.20231220.144533/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231220.104729
+Version: 1.20231220.144533
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231220.104729/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20231220.104729'
-long_version = '1.20231220.104729+git.8209aee'