returnn 1.20250905.121049__tar.gz → 1.20250909.90304__tar.gz

{returnn-1.20250905.121049/returnn.egg-info → returnn-1.20250909.90304}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250905.121049
+Version: 1.20250909.90304
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250909.90304/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250909.090304'
+long_version = '1.20250909.090304+git.9e941ab'

{returnn-1.20250905.121049 → returnn-1.20250909.90304}/returnn/frontend/nested.py

@@ -354,7 +354,10 @@ def _masked_scatter_merge_dims(
         if s == backup:
             return s
         if s in merged_dim_map:
+            # If this assert fails, see e.g. https://github.com/rwth-i6/returnn/pull/1759 for an example.
+            assert backup in merged_dim_map, f"nested masked_scatter: mismatch of s {s} vs backup {backup}"
             return merged_dim_map[s]
+        assert backup not in merged_dim_map, f"nested masked_scatter: mismatch of s {s} vs backup {backup}"
         # Note: s/backup might even be static dims.
         new_size = _masked_scatter(
             s.get_size_tensor(),

{returnn-1.20250905.121049 → returnn-1.20250909.90304}/returnn/torch/data/returnn_dataset_wrapper.py

@@ -20,12 +20,17 @@ ResetCallbackT = Callable[[], None]
 class ReturnnDatasetResetDefaultEpochCounterCallback:
     """
     Default for reset_callback.
-    Has an internal counter for the epoch, starting at epoch 1 (RETURNN convention).
+    Has an internal counter for the epoch, starting by default at epoch 1 (RETURNN convention).
     """
 
-    def __init__(self, dataset: ReturnnDataset):
+    def __init__(self, dataset: ReturnnDataset, epoch: int = 0):
+        """
+        :param dataset: RETURNN dataset.
+        :param epoch: Epoch from which the dataset sequence ordering should start.
+            By default 0 since next :func:`__call__` will increment, thus we start at epoch 1.
+        """
         self.dataset = dataset
-        self.epoch = 0  # next __call__ will increment, thus we start at epoch 1
+        self.epoch = epoch
 
     def __call__(self):
         # dataset is likely a copy of the original dataset, either in the main process or in a worker process

{returnn-1.20250905.121049 → returnn-1.20250909.90304}/returnn/torch/frontend/_backend.py

@@ -1791,6 +1791,7 @@ class TorchBackend(Backend[torch.Tensor]):
             dims=(out_dim,) + tuple(remaining_dims),
             dtype=tensor.dtype,
             sparse_dim=tensor.sparse_dim,
+            feature_dim=tensor.feature_dim,
             raw_tensor=out_raw,
         )
         return out, out_dim

{returnn-1.20250905.121049 → returnn-1.20250909.90304/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250905.121049
+Version: 1.20250909.90304
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250905.121049 → returnn-1.20250909.90304}/tests/test_rf_decoder_transformer.py

@@ -483,7 +483,6 @@ def test_transformer_decoder_time_sync_search():
                 raise ValueError(f"invalid recog_recomb {recomb!r}")
 
         if decoder is not None and got_new_label_cpu.raw_tensor.sum().item() > 0:
-            assert decoder_state["0"].self_att.accum_axis != decoder_state["1"].self_att.accum_axis
             (target_, decoder_state_, enc_), packed_new_label_dim, packed_new_label_dim_map = (
                 rf.nested.masked_select_nested(
                     (target, decoder_state, enc),
@@ -492,8 +491,6 @@ def test_transformer_decoder_time_sync_search():
                     dims=batch_dims + [beam_dim],
                 )
             )
-            assert decoder_state_["0"].self_att.accum_axis != decoder_state_["1"].self_att.accum_axis
-            assert decoder_state_["0"].self_att.accum_axis + 1 != decoder_state_["1"].self_att.accum_axis + 1
             # packed_new_label_dim_map: old dim -> new dim. see _masked_select_prepare_dims
             assert packed_new_label_dim.get_dim_value() > 0
 
@@ -503,7 +500,6 @@ def test_transformer_decoder_time_sync_search():
                 spatial_dim=single_step_dim,
                 state=decoder_state_,
             )  # Flat_Batch_Beam, Vocab / ...
-            assert decoder_state_["0"].self_att.accum_axis != decoder_state_["1"].self_att.accum_axis
             decoder_log_probs_ = rf.log_softmax(decoder_logits_, axis=target_dim)  # Flat_Batch_Beam, Vocab
 
             decoder_log_probs, decoder_state = rf.nested.masked_scatter_nested(

returnn-1.20250905.121049/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250905.121049'
-long_version = '1.20250905.121049+git.c688197'