returnn 1.20240610.94006__tar.gz → 1.20240610.111814__tar.gz

{returnn-1.20240610.94006/returnn.egg-info → returnn-1.20240610.111814}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240610.94006
+Version: 1.20240610.111814
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240610.111814/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240610.111814'
+long_version = '1.20240610.111814+git.b8dc968'

{returnn-1.20240610.94006 → returnn-1.20240610.111814}/returnn/datasets/basic.py

@@ -28,6 +28,9 @@ from returnn.util.basic import try_run, NumbersDict, OptionalNotImplementedError
 from returnn.tensor import TensorDict
 
 
+RANDOM_SEED_OFFSET_ENV_VAR = "RETURNN_RANDOM_SEED_OFFSET"
+
+
 class Dataset(object):
     """
     Base class for any dataset. This defines the dataset API.
@@ -245,7 +248,11 @@ class Dataset(object):
         config = get_global_config(raise_exception=False)
         if not config:
             return 0
-        if config.typed_value("torch_distributed") is not None:
+
+        env_val = os.environ.get(RANDOM_SEED_OFFSET_ENV_VAR)
+        if env_val is not None:
+            return int(env_val)
+        elif config.typed_value("torch_distributed") is not None:
             import returnn.torch.distributed
 
             return returnn.torch.distributed.get_ctx(config=config).rank() * 16127

{returnn-1.20240610.94006 → returnn-1.20240610.111814}/returnn/datasets/concat_files.py

@@ -11,10 +11,10 @@ import os
 import sys
 import numpy
 from returnn.log import log
-from returnn.util.basic import try_run
+from returnn.util.basic import override_env_var, try_run
 from returnn.util.multi_proc_non_daemonic_spawn import NonDaemonicSpawnContext
 from returnn.config import SubProcCopyGlobalConfigPreInitFunc
-from .basic import init_dataset, DatasetSeq
+from .basic import init_dataset, DatasetSeq, RANDOM_SEED_OFFSET_ENV_VAR
 from .cached2 import CachedDataset2
 
 # noinspection PyProtectedMember
@@ -423,6 +423,9 @@ class _WorkerProcParent:
         buffer_size: int,
         exit_hook: Optional[Callable[[], None]] = None,
     ):
+        # the dataset makes sure this is set
+        assert "random_seed_offset" in dataset_dict
+
         self.epoch = epoch
         self.full_epoch_0idx = full_epoch_0idx
         self.dataset_dict = dataset_dict
@@ -431,13 +434,16 @@ class _WorkerProcParent:
         parent_conn, child_conn = _mp.Pipe()
         self.parent_conn: mpConnection = parent_conn
 
-        self.worker_proc = _mp.Process(
-            name=f"{name} worker ep {epoch}",
-            target=_worker_proc_loop,
-            args=(epoch, buffer_size, dataset_dict, child_conn),
-            daemon=True,
-        )
-        self.worker_proc.start()
+        # the env will be forwarded to the child process
+        with override_env_var(RANDOM_SEED_OFFSET_ENV_VAR, str(dataset_dict["random_seed_offset"])):
+            self.worker_proc = _mp.Process(
+                name=f"{name} worker ep {epoch}",
+                target=_worker_proc_loop,
+                args=(epoch, buffer_size, dataset_dict, child_conn),
+                daemon=True,
+            )
+            self.worker_proc.start()
+
         # Make sure the child connection is closed here.
         # It stays open in the child, until the child dies.
         # When that happens, now any consecutive read on the pipe

{returnn-1.20240610.94006 → returnn-1.20240610.111814}/returnn/datasets/meta.py

@@ -949,7 +949,15 @@ class CombinedDataset(CachedDataset2):
         self.sampling_sizes = sampling_sizes
 
         # This will only initialize datasets needed for features occurring in data_map
-        self.datasets = {key: init_dataset(datasets[key]) for key in self.dataset_keys}
+        self.datasets = {
+            key: init_dataset(
+                {
+                    **datasets[key],
+                    "random_seed_offset": datasets[key].get("random_seed_offset", self.random_seed_offset),
+                }
+            )
+            for key in self.dataset_keys
+        }
 
         self._estimated_num_seqs = sum([self.datasets[k].estimated_num_seqs for k in sorted(self.datasets.keys())])
         self.estimated_num_seq_per_subset = [self.datasets[k].estimated_num_seqs for k in sorted(self.datasets.keys())]

{returnn-1.20240610.94006 → returnn-1.20240610.111814}/returnn/util/basic.py

@@ -4519,3 +4519,23 @@ def find_libcudart_from_runtime():
             return fn
     _find_libcudart_from_runtime_cached = [None]
     return None
+
+
+@contextlib.contextmanager
+def override_env_var(var_name: str, value: str):
+    """
+    context manager for temporarily overriding the value of an env var
+
+    :param var_name: the name of the environment variable to override
+    :param value: the value to set while the context mgr is active
+    """
+
+    cur_val = os.environ.get(var_name)
+    os.environ[var_name] = value
+    try:
+        yield
+    finally:
+        if cur_val is not None:
+            os.environ[var_name] = cur_val
+        else:
+            os.environ.pop(var_name)

{returnn-1.20240610.94006 → returnn-1.20240610.111814/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240610.94006
+Version: 1.20240610.111814
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240610.94006/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240610.094006'
-long_version = '1.20240610.094006+git.260ab31'