returnn 1.20241016.173036__tar.gz → 1.20241017.4429__tar.gz

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241016.173036
+Version: 1.20241017.4429
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241017.4429/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20241017.004429'
+long_version = '1.20241017.004429+git.c3878ec'

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/basic.py

@@ -35,7 +35,7 @@ if TYPE_CHECKING:
 RANDOM_SEED_OFFSET_ENV_VAR = "RETURNN_RANDOM_SEED_OFFSET"
 
 
-class Dataset(object):
+class Dataset:
     """
     Base class for any dataset. This defines the dataset API.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/bundle_file.py

@@ -1,4 +1,4 @@
-class BundleFile(object):
+class BundleFile:
     """Holds paths to HDF dataset files."""
 
     def __init__(self, filePath):

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/hdf.py

@@ -476,7 +476,7 @@ class HDFDataset(CachedDataset):
 # ------------------------------------------------------------------------------
 
 
-class StreamParser(object):
+class StreamParser:
     """
     Stream parser.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/map.py

@@ -9,7 +9,7 @@ from returnn.datasets.cached2 import CachedDataset2
 from returnn.util.basic import OptionalNotImplementedError
 
 
-class MapDatasetBase(object):
+class MapDatasetBase:
     """
     This dataset can be used as template to implement user-side Datasets,
     where the data can be access in arbitrary order.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/normalization_data.py

@@ -6,7 +6,7 @@ from .bundle_file import BundleFile
 from returnn.util.basic import long
 
 
-class NormalizationData(object):
+class NormalizationData:
     """This class holds normalization data for inputs and outputs.
     It also contains methods to create the normalization HDF file.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/sprint.py

@@ -1064,7 +1064,7 @@ class SprintCacheDataset(CachedDataset2):
     For alignments, you need to provide all options for the AllophoneLabeling class, such as allophone file, etc.
     """
 
-    class SprintCacheReader(object):
+    class SprintCacheReader:
         """
         Helper class to read a Sprint cache directly.
         """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/datasets/util/vocabulary.py

@@ -22,7 +22,7 @@ from returnn.log import log
 from returnn.util.basic import NotSpecified
 
 
-class Vocabulary(object):
+class Vocabulary:
     """
     Represents a vocabulary (set of words, and their ids).
     Used by :class:`BytePairEncoding`.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/extern/graph_editor/reroute.py

@@ -61,7 +61,7 @@ def _check_ts_compatibility(ts0, ts1):
             raise ValueError("Shapes {} and {} are not compatible.".format(shape0, shape1))
 
 
-class _RerouteMode(object):
+class _RerouteMode:
     """Enums for reroute's mode.
 
     swap: the end of tensors a and b are swapped.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/extern/graph_editor/subgraph.py

@@ -68,7 +68,7 @@ def _check_within_range(mapping, n, repetition):
         raise ValueError("Found repetition in mapping: {}".format(mapping))
 
 
-class SubGraphView(object):
+class SubGraphView:
     """A subgraph view on an existing `tf.Graph`.
 
     An instance of this class is a subgraph view on an existing `tf.Graph`.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/extern/graph_editor/transform.py

@@ -161,7 +161,7 @@ def copy_op_handler(info, op, new_inputs):
     return op_, op_.outputs
 
 
-class TransformerInfo(object):
+class TransformerInfo:
     """ "Contains information about the result of a transform operation."""
 
     def __init__(self, info):
@@ -282,7 +282,7 @@ class TransformerInfo(object):
         return res.getvalue()
 
 
-class _TmpInfo(object):
+class _TmpInfo:
     """Transformer temporary data.
 
     An instance of this class holds all the information relevant to a call
@@ -333,7 +333,7 @@ class _TmpInfo(object):
         return name_
 
 
-class Transformer(object):
+class Transformer:
     """Transform a subgraph into another one.
 
     By default, the constructor create a transform which copy a subgraph and
@@ -614,6 +614,7 @@ def copy_with_input_replacements(
         raise TypeError("Expected a tf.Graph, got: {}".format(type(dst_graph)))
 
     copier = Transformer()
+
     # Replace tensor if possible.
     def replace_t_with_replacement_handler(info, t):
         if t in replacement_ts:

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/extern/graph_editor/util.py

@@ -62,7 +62,7 @@ def concatenate_unique(la, lb):
 
 
 # TODO(fkp): very generic code, it should be moved in a more generic place.
-class ListView(object):
+class ListView:
     """Immutable list wrapper.
 
     This class is strongly inspired by the one in tf.Operation.
@@ -333,7 +333,7 @@ def get_consuming_ops(ts):
     return ops
 
 
-class ControlOutputs(object):
+class ControlOutputs:
     """The control outputs topology."""
 
     def __init__(self, graph):

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/learning_rate_control.py

@@ -9,12 +9,12 @@ from typing import Optional, Any, Dict
 import typing
 import os
 import returnn.util.basic as util
-from returnn.util.basic import better_repr, simple_obj_repr, ObjAsDict, unicode
+from returnn.util.basic import better_repr, simple_obj_repr, unicode
 from returnn.log import log
 import numpy
 
 
-class LearningRateControl(object):
+class LearningRateControl:
     """
     Base class for learning rate control / scheduling.
     """
@@ -531,7 +531,7 @@ class LearningRateControl(object):
         Loads the saved epoch data from file (self.filename).
         """
         s = open(self.filename).read()
-        self.epoch_data = eval(s, {"nan": float("nan"), "inf": float("inf")}, ObjAsDict(self))
+        self.epoch_data = eval(s, {"EpochData": self.EpochData, "nan": float("nan"), "inf": float("inf"), "np": numpy})
 
 
 class ConstantLearningRate(LearningRateControl):

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/native_op.py

@@ -17,7 +17,7 @@ import typing
 from returnn.util.basic import make_hashable, unicode
 
 
-class NativeOpBaseMixin(object):
+class NativeOpBaseMixin:
     """
     The purpose of having this as a separate base class is to make this independent of any Theano specific
     functionality so that we can also use this base for example for TensorFlow.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/sprint/cache.py

@@ -698,7 +698,7 @@ def is_sprint_cache_file(filename):
         return header == FileArchive.SprintCacheHeader
 
 
-class AllophoneLabeling(object):
+class AllophoneLabeling:
     """
     Allophone labeling.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/sprint/interface.py

@@ -128,7 +128,7 @@ def init_python_feature_scorer(config, **kwargs):
     return cls(sprint_opts=sprint_opts, **kwargs)
 
 
-class PythonFeatureScorer(object):
+class PythonFeatureScorer:
     """
     Sprint API.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/data_pipeline.py

@@ -134,7 +134,7 @@ import returnn.tf.horovod as tf_horovod
 from returnn.log import log
 
 
-class DataProviderBase(object):
+class DataProviderBase:
     """
     Base class which wraps up the logic in this class. See derived classes.
     """
@@ -507,7 +507,7 @@ class FeedDictDataProvider(DataProviderBase):
         return self.batches.completed_frac()
 
 
-class InputContext(object):
+class InputContext:
     """
     This object will be passed to the dataset pipeline function
     (``dataset_pipeline`` in the config)

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/engine.py

@@ -46,7 +46,7 @@ class CancelTrainingException(Exception):
     """
 
 
-class Runner(object):
+class Runner:
     """
     This encapsulates the logic around TF ``session.run``, i.e. iterating over the dataset.
     """
@@ -2838,7 +2838,7 @@ class Engine(EngineBase):
         assert not os.path.exists(output_file), "Already existing output file %r." % output_file
         print("Compute priors, using output layer %r, writing to %r." % (output_layer, output_file), file=log.v2)
 
-        class Accumulator(object):
+        class Accumulator:
             """
             Also see PriorEstimationTaskThread for reference.
             """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/layers/base.py

@@ -18,7 +18,7 @@ from returnn.tf.util.basic import OutputWithActivation, CustomUpdate, reuse_name
 from returnn.log import log
 
 
-class LayerBase(object):
+class LayerBase:
     """
     This is the base class for all layers.
     Every layer by default has a list of source layers `sources`
@@ -2585,7 +2585,7 @@ class ReuseParams:
         return getter(name=name, shape=shape, dtype=dtype, **kwargs)
 
 
-class SearchChoices(object):
+class SearchChoices:
     """
     In beam search, after expanding the beam and then selecting the N best (beam) (see :class:`ChoiceLayer`),
     when doing this multiple times, we need to keep reference where each beam came from,
@@ -2824,7 +2824,7 @@ class SearchChoices(object):
         return common_choices.translate_to_this_search_beam(layer_desc)
 
 
-class Loss(object):
+class Loss:
     """
     Base class for all losses.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/layers/rec.py

@@ -1340,7 +1340,7 @@ class RecLayer(_ConcatInputLayer):
         return []
 
 
-class _SubnetworkRecCell(object):
+class _SubnetworkRecCell:
     """
     This class is used by :class:`RecLayer` to implement
     the generic subnetwork logic inside the recurrency.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/layers/signal_processing.py

@@ -1033,7 +1033,7 @@ class ParametricWienerFilterLayer(LayerBase):
 
         super(ParametricWienerFilterLayer, self).__init__(**kwargs)
 
-        class _NoiseEstimator(object):
+        class _NoiseEstimator:
             """
             Noise estimator. For details, see tfSi6Proc.audioProcessing.enhancement.singleChannel
             """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/native_op.py

@@ -69,7 +69,7 @@ class OpDescription(native_op.NativeOpBaseMixin):
         return OpDescription(**kwargs)
 
 
-class OpMaker(object):
+class OpMaker:
     """
     https://www.tensorflow.org/guide/extend/op
     """
@@ -730,7 +730,7 @@ def make_lstm_op(**kwargs):
     return make_op(native_op.LstmGenericBase, **kwargs)
 
 
-class RecSeqCellOp(object):
+class RecSeqCellOp:
     """
     In TF terminology, this is a "fused" cell, i.e. the op loops over the time.
     Similar is e.g. :class:`tf.contrib.rnnLSTMBlockFusedCell`.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/network.py

@@ -539,7 +539,7 @@ class _NetworkConstructionStack:
         assert False, "we should not get here"
 
 
-class TFNetwork(object):
+class TFNetwork:
     """
     The main neural network, i.e. collection of interconnected layers, i.e. computation graph with trainable params.
     """
@@ -3679,7 +3679,7 @@ class Subnetwork:
             raise new_exc
 
 
-class TFNetworkParamsSerialized(object):
+class TFNetworkParamsSerialized:
     """
     Holds all the params as numpy arrays, including auxiliary params.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/updater.py

@@ -112,7 +112,7 @@ def get_optimizer_class(class_name):
     return _OptimizerClassesDict[class_name.lower()]
 
 
-class Updater(object):
+class Updater:
     """
     This will create the :class:`tf.compat.v1.train.Optimizer` instance given the config
     and the update-op for all trainable vars.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/util/basic.py

@@ -72,7 +72,7 @@ def have_min_tf_version(version):
     return tf_version >= version
 
 
-class CustomUpdate(object):
+class CustomUpdate:
     """
     Custom updates will be handled by :class:`TFUpdater`.
     """
@@ -295,7 +295,7 @@ def copy_compatible_reduce(source, target, reduce_type):
     return source.copy_compatible_to(target, check_sparse=False, check_dtype=False)
 
 
-class OutputWithActivation(object):
+class OutputWithActivation:
     """
     Stores some tensor before and after some activation function,
     and also the activation function itself.
@@ -645,7 +645,7 @@ def get_root_graph(graph=None):
     return graph
 
 
-class _ScaledGradientBuilder(object):
+class _ScaledGradientBuilder:
     """
     Use the ``scaled_gradient`` instance.
     tf.identity in forward pass, but scales the gradient in backprop.
@@ -2633,7 +2633,7 @@ def matrix_triangular(shape, dtype=tf.float32, lower=False, upper=False):
     return tf_compat.v1.matrix_band_part(x, num_lower=-1 if lower else 0, num_upper=-1 if upper else 0)
 
 
-class VariableAssigner(object):
+class VariableAssigner:
     """
     Object helper to assign some var.
     (This is mostly obsolete now.)
@@ -2766,7 +2766,7 @@ def get_tf_gpp_path():
     return _tf_gpp_path
 
 
-class CudaEnv(object):
+class CudaEnv:
     """
     Information about the Nvidia CUDA environment, and library.
     Also path to ``nvcc``, the CUDA compiler.
@@ -3224,7 +3224,7 @@ def add_scaled_noise_to_gradients(grads_and_vars, gradient_noise_scale, sparse_g
     return list(zip(noisy_gradients, variables))
 
 
-class CustomGradient(object):
+class CustomGradient:
     """
     Utility functions to specify a custom gradient for a given function,
     which will be wrapped around via TF :func:`Defun`.
@@ -3333,7 +3333,7 @@ class CustomGradient(object):
 custom_gradient = CustomGradient()
 
 
-class MetaLosses(object):
+class MetaLosses:
     """
     This provides a way to use an alternative gradient,
     or to use the original gradient (error signal) and do something with it.
@@ -3361,7 +3361,7 @@ class MetaLosses(object):
             self.name = name
             self.source = source
 
-    class Scope(object):
+    class Scope:
         """
         Defines the scope for a synthetic gradient.
         Create this object via :func:`MetaLosses.enter_gradient_scope`.

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/tf/util/data.py

@@ -99,7 +99,7 @@ class BatchInfo:
     However, we maybe can make use of some of the functions in :mod:`tf.ragged`.
     """
 
-    class VirtualDimBase(object):
+    class VirtualDimBase:
         """
         Represents one virtual dim, flattened into the batch dim.
         """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/torch/updater.py

@@ -92,7 +92,7 @@ class Updater:
         :param float initial_learning_rate:
         """
         self.config = config
-        self.learning_rate = initial_learning_rate
+        self.learning_rate = float(initial_learning_rate)
         self._effective_learning_rate = self.learning_rate
         self.network = network
         self._device = device
@@ -150,7 +150,7 @@ class Updater:
 
         :param float value: New learning rate.
         """
-        self.learning_rate = value
+        self.learning_rate = float(value)
         self._update_effective_learning_rate()
 
     def get_effective_learning_rate(self) -> float:
@@ -162,9 +162,10 @@ class Updater:
     def _update_effective_learning_rate(self):
         self._effective_learning_rate = self.learning_rate
         if self.learning_rate_function is not None:
-            self._effective_learning_rate = self.learning_rate_function(
+            lr = self.learning_rate_function(
                 global_train_step=self._current_train_step, epoch=self._current_epoch, learning_rate=self.learning_rate
             )
+            self._effective_learning_rate = float(lr)
         if self.optimizer:
             for param_group in self.optimizer.param_groups:
                 param_group["lr"] = self._effective_learning_rate

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/util/basic.py

@@ -52,7 +52,7 @@ V = TypeVar("V")
 returnn_root_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 
 
-class NotSpecified(object):
+class NotSpecified:
     """
     This is just a placeholder, to be used as default argument to mark that it is not specified.
     """
@@ -3194,7 +3194,7 @@ def get_cache_dir():
     return get_temp_dir()
 
 
-class LockFile(object):
+class LockFile:
     """
     Simple lock file.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/util/better_exchook.py

@@ -339,7 +339,7 @@ def debug_shell(user_ns, user_global_ns, traceback=None, execWrapper=None):
             # noinspection PyPackageRequirements,PyUnresolvedReferences
             import IPython.terminal.embed
 
-            class DummyMod(object):
+            class DummyMod:
                 """Dummy module"""
 
             module = DummyMod()

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/util/native_code_compiler.py

@@ -11,7 +11,7 @@ import sys
 from . import basic as util
 
 
-class NativeCodeCompiler(object):
+class NativeCodeCompiler:
     """
     Helper class to compile native C/C++ code on-the-fly.
     """

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn/util/sig_proc.py

@@ -48,7 +48,7 @@ def inv_greenwood_function(freq, scaling_constant=165.4, constant_of_integration
     return numpy.log10(freq / scaling_constant + constant_of_integration) / slope
 
 
-class GammatoneFilterbank(object):
+class GammatoneFilterbank:
     """
     Class representing a gammatone filterbank.
     Based on

{returnn-1.20241016.173036 → returnn-1.20241017.4429}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241016.173036
+Version: 1.20241017.4429
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241016.173036/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20241016.173036'
-long_version = '1.20241016.173036+git.0035a76'