returnn 1.20240606.160130__tar.gz → 1.20240607.35919__tar.gz

{returnn-1.20240606.160130/returnn.egg-info → returnn-1.20240607.35919}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240606.160130
+Version: 1.20240607.35919
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240607.35919/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240607.035919'
+long_version = '1.20240607.035919+git.ac594c5'

{returnn-1.20240606.160130 → returnn-1.20240607.35919}/returnn/datasets/concat_files.py

@@ -10,6 +10,7 @@ from typing import Optional, Any, Dict, Tuple, List, Sequence, Collection, Calla
 import os
 import sys
 import numpy
+from returnn.log import log
 from returnn.util.basic import try_run
 from returnn.util.multi_proc_non_daemonic_spawn import NonDaemonicSpawnContext
 from returnn.config import SubProcCopyGlobalConfigPreInitFunc
@@ -274,7 +275,9 @@ class ConcatFilesDataset(CachedDataset2):
                 continue
             full_epoch_0idx_ = (ep_ - 1) // self.partition_epoch
             files_order: List[List[FileTree]] = self._files_order_cache[full_epoch_0idx_]
-            dataset_dict, exit_hook = self._get_sub_dataset_dict(files=files_order[(ep_ - 1) % self.partition_epoch])
+            files_for_subep = files_order[(ep_ - 1) % self.partition_epoch]
+            print(f"{self}: using files for epoch {ep_}: {files_for_subep}", file=log.v5)
+            dataset_dict, exit_hook = self._get_sub_dataset_dict(files=files_for_subep)
             worker = _WorkerProcParent(
                 name=f"{self.__class__.__name__} {self.name} ep {epoch}",
                 epoch=ep_,
@@ -294,8 +297,11 @@ class ConcatFilesDataset(CachedDataset2):
             dataset_dict["random_seed_offset"] = self.random_seed_offset
         if dataset_dict.get("partition_epoch", 1) != 1:
             raise ValueError(f"{self}: sub dataset should not have partition_epoch, got: {dataset_dict}")
-        if "seq_ordering" not in dataset_dict:
-            raise ValueError(f"{self}: sub dataset should have explicit seq_ordering, got: {dataset_dict}")
+        if "seq_ordering" not in dataset_dict and "seq_order_control_dataset" not in dataset_dict:
+            raise ValueError(
+                f"{self}: sub dataset should have explicit seq_ordering "
+                "(or seq_order_control_dataset for MetaDataset), got: {dataset_dict}"
+            )
         self._lazy_init_file_cache_proc()
         dataset_dict, exit_hook = self._file_cache.handle_cached_files_in_config(dataset_dict)
         return dataset_dict, exit_hook
@@ -367,6 +373,10 @@ class ConcatFilesDataset(CachedDataset2):
         seq.seq_idx = seq_idx
         return seq
 
+    def have_seqs(self) -> bool:
+        """have seqs"""
+        return bool(self.files)
+
     @property
     def num_seqs(self) -> int:
         """num seqs"""

{returnn-1.20240606.160130 → returnn-1.20240607.35919}/returnn/torch/frontend/_backend.py

@@ -2018,6 +2018,7 @@ class TorchBackend(Backend[torch.Tensor]):
 
         out = source.copy_template_replace_dim_tag(axis=-1, new_dim_tag=out_dim, name="lstm")
         out.feature_dim = out_dim
+        out.dtype = TorchBackend.get_dtype_name_raw(out_raw)  # maybe due to AMP
         out.raw_tensor = out_raw
 
         new_state_h = state_h.copy_template()

{returnn-1.20240606.160130 → returnn-1.20240607.35919/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240606.160130
+Version: 1.20240607.35919
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240606.160130/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240606.160130'
-long_version = '1.20240606.160130+git.bc0d5bf'