returnn 1.20250909.170507__tar.gz → 1.20250922.113348__tar.gz

{returnn-1.20250909.170507/returnn.egg-info → returnn-1.20250922.113348}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250909.170507
+Version: 1.20250922.113348
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250922.113348/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250922.113348'
+long_version = '1.20250922.113348+git.58aca86'

{returnn-1.20250909.170507 → returnn-1.20250922.113348}/returnn/datasets/distrib_files.py

@@ -153,7 +153,7 @@ class DistributeFilesDataset(CachedDataset2):
             or a JSON file containing a list of arbitarily nested (JSON) objects.
         :param get_sub_epoch_dataset: callable which returns a dataset dict for a given subset of files
         :param preload_next_n_sub_epochs: how many sub epoch datasets to preload
-        :param buffer_size: buffer size for each worker, amount of seqs to prefetch
+        :param buffer_size: buffer size for each worker, number of seqs to prefetch
         :param distrib_shard_files: set to true to shard the data across worker processes in
             distributed training scenaria
         :param _meta_info_cache: for internal use

{returnn-1.20250909.170507 → returnn-1.20250922.113348}/returnn/datasets/multi_proc.py

@@ -42,7 +42,7 @@ class MultiProcDataset(CachedDataset2):
         """
         :param dataset: the dataset to use
         :param num_workers: number of workers to use
-        :param buffer_size: buffer size for each worker, amount of seqs to prefetch
+        :param buffer_size: buffer size for each worker, number of seqs to prefetch
         :param sharding_method: which method to use for sharding the data across the worker procs.
             The default is ``seq_order``, which fetches the full list of seq indices,
             and then distributes shards of that to the other workers.

{returnn-1.20250909.170507 → returnn-1.20250922.113348/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250909.170507
+Version: 1.20250922.113348
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250909.170507 → returnn-1.20250922.113348}/tests/test_rf_encoder_conformer.py

@@ -496,3 +496,25 @@ def test_e_branchformer_meta_dev_num_params():
         num_params_ = sum(parameter.numel() for parameter in e_branchformer_pt.parameters())
         print(f"E-Branchformer {num_layers}x{dim_} net params #: {num_params_}")
         assert num_params_ == 413034496
+
+        e_branchformer = ConformerEncoder(
+            in_dim=in_dim,
+            input_layer=None,
+            num_layers=num_layers,
+            out_dim=dim,
+            encoder_layer=rf.build_dict(
+                EBranchformerLayer,
+                ff=rf.build_dict(
+                    rf.encoder.conformer.ConformerPositionwiseFeedForward,
+                    ff_dim=dim * 2,
+                    activation=rf.build_dict(rf.relu_square),
+                    with_bias=False,
+                ),
+                num_heads=8,
+            ),
+        )
+        e_branchformer_pt = rf_module_to_pt_module(e_branchformer)
+        assert all(parameter.device.type == "meta" for parameter in e_branchformer_pt.parameters())
+        num_params_ = sum(parameter.numel() for parameter in e_branchformer_pt.parameters())
+        print(f"E-Branchformer {num_layers}x{dim} (FF: {dim}*2) net params #: {num_params_}")
+        assert num_params_ == 404930560

returnn-1.20250909.170507/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250909.170507'
-long_version = '1.20250909.170507+git.4dc7088'