returnn 1.20240628.104058__tar.gz → 1.20240701.135153__tar.gz

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240628.104058
+Version: 1.20240701.135153
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240701.135153/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240701.135153'
+long_version = '1.20240701.135153+git.5239436'

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/datasets/generating.py

@@ -2543,7 +2543,7 @@ class Enwik8Corpus(CachedDataset2):
         num_test_chars = 5000000
 
         raw_data = zipfile.ZipFile(self._zip_filename).read("enwik8").decode("utf8")
-        raw_data = numpy.fromstring(raw_data, dtype=numpy.uint8)
+        raw_data = numpy.fromstring(raw_data, dtype=numpy.uint8)  # noqa
         unique, data = numpy.unique(raw_data, return_inverse=True)
 
         train_data = data[: -2 * num_test_chars]

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/datasets/util/feature_extraction.py

@@ -523,7 +523,7 @@ def _get_random_permuted_audio(audio, sample_rate, opts, random_state):
             warnings.simplefilter("ignore")
             # Alternative: scipy.interpolate.interp2d
             factor = random_state.uniform(opts.get("rnd_zoom_lower", 0.9), opts.get("rnd_zoom_upper", 1.1))
-            audio = scipy.ndimage.zoom(audio, factor, order=opts.get("rnd_zoom_order", 3))
+            audio = scipy.ndimage.zoom(audio, factor, order=opts.get("rnd_zoom_order", 3))  # noqa
     if opts.get("rnd_stretch_switch", 1.0) > 0.0:
         opts.get("rnd_stretch_lower"), opts.get("rnd_stretch_upper")  # Mark as read.
     if random_state.uniform(0.0, 1.0) < opts.get("rnd_stretch_switch", 0.2):

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/sprint/control.py

@@ -123,6 +123,7 @@ def init(name, reference, config, sprint_unit=None, version_number=None, callbac
 
 # Start Sprint PythonSegmentOrder interface. {
 
+
 # Keep names for compatibility.
 # noinspection PyPep8Naming,PyUnusedLocal
 def getSegmentList(corpusName, segmentList, config, **kwargs):
@@ -268,7 +269,7 @@ class SprintNnPythonLayer:
         seg_len = errorSignalIn.shape[1]
         PythonControl.instance.set_current_seg_error_signal(seg_len=seg_len, error_signal=errorSignalIn.T)
         # must return a 1-tuple
-        return (numpy.zeros((self.input_size, seg_len), dtype="float32"),)
+        return (numpy.zeros((self.input_size, seg_len), dtype="float32"),)  # noqa
 
 
 # End SprintNnPythonLayer. }
@@ -400,7 +401,7 @@ class PythonControl:
         with self.cond:
             self.loss = loss
             self.error_signal = error_signal
-            self.cond.notifyAll()
+            self.cond.notify_all()
 
     def _get_loss_and_error_signal_via_sprint_callback(self, seg_name, orthography, posteriors):
         """
@@ -484,13 +485,13 @@ class PythonControl:
             self.loss = None
             self.asked_for_posteriors = False
             self.notified_for_segment = False
-            self.cond.notifyAll()
+            self.cond.notify_all()
         loss, error_signal = self.callback("get_loss_and_error_signal", seg_name, seg_len, posteriors)
         assert error_signal.shape == posteriors.shape
         with self.cond:
             self.control_thread__have_new_error_signal = True
             self.posteriors = None
-            self.cond.notifyAll()
+            self.cond.notify_all()
         numpy_set_unused(posteriors)
         error_signal = error_signal.astype("float32", copy=False)
         return loss, error_signal
@@ -544,14 +545,14 @@ class PythonControl:
         with self.cond:
             assert not self.control_loop_started
             self.control_loop_started = True
-            self.cond.notifyAll()
+            self.cond.notify_all()
         try:
             while True:
                 self.handle_next()
         finally:
             with self.cond:
                 self.control_loop_exited = True
-                self.cond.notifyAll()
+                self.cond.notify_all()
 
     # noinspection PyMethodMayBeStatic
     def exit(self, **kwargs):
@@ -635,7 +636,7 @@ class PythonControl:
         with self.cond:
             assert self.seg_name == segment_name
             self.notified_for_segment = True
-            self.cond.notifyAll()
+            self.cond.notify_all()
 
     def notify_segment_loss(self, segment_name, loss):
         """
@@ -654,7 +655,7 @@ class PythonControl:
             assert self.seg_len == seg_len
             assert self.posteriors.shape[0] == seg_len
             self.asked_for_posteriors = True
-            self.cond.notifyAll()
+            self.cond.notify_all()
             return self.posteriors
 
     def set_current_seg_error_signal(self, seg_len, error_signal):
@@ -667,7 +668,7 @@ class PythonControl:
             assert error_signal.ndim == 2
             assert error_signal.shape[0] == seg_len
             self.error_signal = error_signal
-            self.cond.notifyAll()
+            self.cond.notify_all()
 
     def set_current_seg_loss(self, seg_name, loss):
         """
@@ -678,7 +679,7 @@ class PythonControl:
             if seg_name:
                 assert self.seg_name == seg_name
             self.loss = loss
-            self.cond.notifyAll()
+            self.cond.notify_all()
 
     # noinspection PyMethodMayBeStatic
     def _default_skipped_loss(self):
@@ -702,7 +703,7 @@ class PythonControl:
                 self.loss = self._default_skipped_loss()
             if self.error_signal is None:
                 self.error_signal = self._default_skipped_error_signal(self.posteriors)
-            self.cond.notifyAll()
+            self.cond.notify_all()
 
     def _wait_for_control_loop_error_signal(self):
         while True:

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/sprint/interface.py

@@ -102,6 +102,7 @@ def init(name=None, sprint_unit=None, **kwargs):
 # <editor-fold desc="PythonFeatureScorer">
 # Start Sprint PythonFeatureScorer interface. {
 
+
 # noinspection PyShadowingNames
 def init_python_feature_scorer(config, **kwargs):
     """
@@ -324,6 +325,7 @@ class PythonFeatureScorer(object):
 # <editor-fold desc="PythonSegmentOrder">
 # Start Sprint PythonSegmentOrder interface. {
 
+
 # Names need to stay that way for compatibility.
 # noinspection PyPep8Naming
 def getSegmentList(corpusName, segmentList, **kwargs):
@@ -399,6 +401,7 @@ def getSegmentList(corpusName, segmentList, **kwargs):
 # <editor-fold desc="PythonTrainer">
 # Start Sprint PythonTrainer interface. {
 
+
 # noinspection PyPep8Naming,PyShadowingNames
 def init_python_trainer(inputDim, outputDim, config, targetMode, **kwargs):
     """
@@ -996,7 +999,7 @@ def make_criterion_class():
         errorSignal = None
 
         def __eq__(self, other):
-            return type(self) == type(other)
+            return type(self) == type(other)  # nopep8
 
         def __hash__(self):
             return hash(type(self))

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/data_pipeline.py

@@ -395,7 +395,7 @@ class FeedDictDataProvider(DataProviderBase):
                 if enqueue_args is not None:
                     self.queue.put(enqueue_args)
                 with self.state_change_cond:
-                    self.state_change_cond.notifyAll()
+                    self.state_change_cond.notify_all()
                 self.batches.advance(1)
 
             self.reached_end = not self.batches.has_more()
@@ -407,7 +407,7 @@ class FeedDictDataProvider(DataProviderBase):
         finally:
             with self.state_change_cond:
                 self.thread_finished = True
-                self.state_change_cond.notifyAll()
+                self.state_change_cond.notify_all()
 
     def have_more_data(self, session):
         """

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/hyper_param_tuning.py

@@ -230,7 +230,7 @@ class HyperParam:
             x = 1.0 - eps
         import scipy.special
 
-        return self.dtype(scipy.special.ndtri(x))
+        return self.dtype(scipy.special.ndtri(x))  # noqa
 
     def get_initial_value(self):
         """

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/layers/rec.py

@@ -6291,9 +6291,11 @@ class ChoiceLayer(BaseChoiceLayer):
                         out=self._debug_out,
                         x=labels[0],
                         args={
-                            "step": self.network.get_rec_step_index()
-                            if self.network.have_rec_step_info()
-                            else tf.constant(-1),
+                            "step": (
+                                self.network.get_rec_step_index()
+                                if self.network.have_rec_step_info()
+                                else tf.constant(-1)
+                            ),
                             "base_beam_in": base_beam_in,
                             "scores_in_orig": self.sources[0].output.placeholder,
                             "scores_in": scores_in,
@@ -8213,6 +8215,7 @@ class SelfAttentionLayer(_ConcatInputLayer):
         :param tf.Tensor v: [B,H,T,V]
         :param tf.Tensor mask: shape [B,T]
         """
+
         # noinspection PyUnusedLocal,PyShadowingNames
         def _masked_transform_mask(value, prev_value, mask):
             """
@@ -8506,33 +8509,35 @@ class KenLmStateLayer(_ConcatInputLayer):
                 tf.less_equal(prev_step, 2),
                 lambda: py_print(
                     new_rel_scores,
-                    [
-                        str(self),
-                        "; step: ",
-                        prev_step,
-                        "; input shape: ",
-                        tf.shape(self.input_data.placeholder),
-                        str(self.input_data),
-                        "; input: ",
-                        self.input_data.placeholder,
-                        "; strings shape: ",
-                        tf.shape(next_strings),
-                        "; strings: ",
-                        "'" + next_strings + "'",
-                        "; new_abs_scores: ",
-                        new_abs_scores,
-                        "; sparse rel scores: ",
-                        new_abs_scores - prev_scores,
-                        "; min/max/mean rel scores: ",
-                        tf.reduce_min(new_rel_scores),
-                        "/",
-                        tf.reduce_max(new_rel_scores),
-                        "/",
-                        tf.reduce_mean(new_rel_scores),
-                    ]
-                    + ["; vocab: ", self.tf_vocab]
-                    if self.tf_vocab is not None
-                    else [],
+                    (
+                        [
+                            str(self),
+                            "; step: ",
+                            prev_step,
+                            "; input shape: ",
+                            tf.shape(self.input_data.placeholder),
+                            str(self.input_data),
+                            "; input: ",
+                            self.input_data.placeholder,
+                            "; strings shape: ",
+                            tf.shape(next_strings),
+                            "; strings: ",
+                            "'" + next_strings + "'",
+                            "; new_abs_scores: ",
+                            new_abs_scores,
+                            "; sparse rel scores: ",
+                            new_abs_scores - prev_scores,
+                            "; min/max/mean rel scores: ",
+                            tf.reduce_min(new_rel_scores),
+                            "/",
+                            tf.reduce_max(new_rel_scores),
+                            "/",
+                            tf.reduce_mean(new_rel_scores),
+                        ]
+                        + ["; vocab: ", self.tf_vocab]
+                        if self.tf_vocab is not None
+                        else []
+                    ),
                 ),
                 lambda: new_rel_scores,
             )
@@ -9897,6 +9902,7 @@ class BaseRNNCell(rnn_cell.RNNCell):
         return x
 
 
+# noinspection PyAbstractClass
 class VanillaLSTMCell(BaseRNNCell):
     """
     Just a vanilla LSTM cell, which is compatible to our NativeLSTM (v1 and v2).
@@ -9964,6 +9970,7 @@ class VanillaLSTMCell(BaseRNNCell):
             return new_h, rnn_cell.LSTMStateTuple(new_c, new_h)
 
 
+# noinspection PyAbstractClass
 class RHNCell(BaseRNNCell):
     """
     Recurrent Highway Layer.
@@ -10119,6 +10126,7 @@ class RHNCell(BaseRNNCell):
         return current_state, current_state
 
 
+# noinspection PyAbstractClass
 class _WrapBaseCell(BaseRNNCell):
     """
     Simpler helper wrapper class, for :class:`BaseRNNCell`.
@@ -10172,6 +10180,7 @@ class _WrapBaseCell(BaseRNNCell):
         return self.cell.call(inputs, state)
 
 
+# noinspection PyAbstractClass
 class BlocksparseLSTMCell(_WrapBaseCell):
     """
     Standard LSTM but uses OpenAI blocksparse kernels to support bigger matrices.
@@ -10269,6 +10278,7 @@ class BlocksparseLSTMCell(_WrapBaseCell):
         m1["bias"].load(b_new, session=session)
 
 
+# noinspection PyAbstractClass
 class BlocksparseMultiplicativeMultistepLSTMCell(_WrapBaseCell):
     """
     Multiplicative LSTM with multiple steps, as in the OpenAI blocksparse paper.
@@ -10324,6 +10334,7 @@ class BlocksparseMultiplicativeMultistepLSTMCell(_WrapBaseCell):
         return y
 
 
+# noinspection PyAbstractClass
 class LayerNormVariantsLSTMCell(BaseRNNCell):
     """LSTM unit with layer normalization and recurrent dropout
 
@@ -10988,6 +10999,7 @@ class TwoDLSTMLayer(LayerBase):
         return y
 
 
+# noinspection PyAbstractClass
 class ZoneoutLSTMCell(BaseRNNCell):
     """
     Wrapper for tf LSTM to create Zoneout LSTM Cell.

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/layers/signal_processing.py

@@ -625,7 +625,7 @@ class MultiChannelMultiResolutionStftLayer(_ConcatInputLayer):
                     # noinspection PyPackageRequirements
                     import scipy.signal
 
-                    window = tf.constant(scipy.signal.windows.blackman(frame_size), dtype=tf.float32)
+                    window = tf.constant(scipy.signal.windows.blackman(frame_size), dtype=tf.float32)  # noqa
                 elif self._window == "None" or self._window == "ones":
                     window = tf.ones((window_length,), dtype=dtype)
                 else:

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/native_op.py

@@ -359,9 +359,9 @@ class OpMaker(object):
                 code_set_io += "outputs[%i] = &output_%i;\n" % (out_idx, out_idx)
                 cshape = "TensorShape({%s})" % ", ".join(
                     [
-                        str(dim)
-                        if isinstance(dim, int)
-                        else ("inputs[%i]->dim_size(%i)" % dim)  # also see make_dim_str
+                        (
+                            str(dim) if isinstance(dim, int) else ("inputs[%i]->dim_size(%i)" % dim)
+                        )  # also see make_dim_str
                         for dim in v["shape"]
                     ]
                 )
@@ -704,7 +704,7 @@ def load_dump_file(filename):
         data = _read_bytes()
         assert len(data) == numpy.prod(dims) * dtype.itemsize
         # noinspection PyTypeChecker
-        v_flat = numpy.fromstring(data, dtype=dtype)
+        v_flat = numpy.fromstring(data, dtype=dtype)  # noqa
         v = v_flat.reshape(dims)
         return v
 

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/updater.py

@@ -931,9 +931,11 @@ class Updater(object):
 
                 grads_and_vars = [
                     (
-                        hvd.allreduce(grad, average=self.config.is_true("horovod_avg_grad"))
-                        if grad is not None
-                        else None,
+                        (
+                            hvd.allreduce(grad, average=self.config.is_true("horovod_avg_grad"))
+                            if grad is not None
+                            else None
+                        ),
                         var,
                     )
                     for (grad, var) in grads_and_vars
@@ -1028,6 +1030,7 @@ def accum_grad_multiple_step(grad, var, train_step, num_accum_steps):
         )
 
 
+# noinspection PyAbstractClass
 class _KerasOptimizerWrapper(Optimizer):
     """
     Wraps a TF optimizer into a standard TF optimizer.
@@ -1090,6 +1093,7 @@ class _KerasOptimizerWrapper(Optimizer):
         return self.keras_optimizer._resource_apply_sparse(grad, handle, indices, None)
 
 
+# noinspection PyAbstractClass
 class BaseCustomOptimizer(Optimizer):
     """
     Base class for our own optimizer implementations.
@@ -1200,6 +1204,7 @@ class BaseCustomOptimizer(Optimizer):
         return dense
 
 
+# noinspection PyAbstractClass
 class CustomGradientDescentOptimizer(BaseCustomOptimizer):
     """
     Just an example implementation for simple gradient descent.
@@ -1217,6 +1222,7 @@ class CustomGradientDescentOptimizer(BaseCustomOptimizer):
         return self._assign_sub(ref=var, updates=lr * grad, indices=indices).op
 
 
+# noinspection PyAbstractClass
 class NormalizedSGD(CustomGradientDescentOptimizer):
     """
     All grads are L2 normalized (via :func:`tf.nn.l2_normalize`), otherwise it's standard SGD.
@@ -1234,6 +1240,7 @@ class NormalizedSGD(CustomGradientDescentOptimizer):
         return super(NormalizedSGD, self)._apply(grad=tf.nn.l2_normalize(grad, None), var=var, indices=indices)
 
 
+# noinspection PyAbstractClass
 class NeuralOptimizer1(BaseCustomOptimizer):
     """
     Via Neural Optimizer Search with Reinforcement Learning (https://proceedings.mlr.press/v70/bello17a/bello17a.pdf).
@@ -1282,6 +1289,7 @@ class NeuralOptimizer1(BaseCustomOptimizer):
         return tf.group(*[var_update, m_t])
 
 
+# noinspection PyAbstractClass
 class GradVarianceScaledOptimizer(BaseCustomOptimizer):
     """
     Let m be the running average of g.
@@ -1344,6 +1352,7 @@ class GradVarianceScaledOptimizer(BaseCustomOptimizer):
         return tf.group(*[var_update, m_t])
 
 
+# noinspection PyAbstractClass
 class NadamOptimizer(tf_compat.v1.train.AdamOptimizer):
     """
     Optimizer that implements the Nadam algorithm.
@@ -1439,6 +1448,7 @@ class NadamOptimizer(tf_compat.v1.train.AdamOptimizer):
         return control_flow_ops.group(*[var_update, m_bar, v_t])
 
 
+# noinspection PyAbstractClass
 class CustomAdamOptimizer(BaseCustomOptimizer):
     """
     Reimplementation of Adam.
@@ -1515,6 +1525,7 @@ class CustomAdamOptimizer(BaseCustomOptimizer):
         return tf.group(*update_ops + [update_beta1, update_beta2], name=name_scope)
 
 
+# noinspection PyAbstractClass
 class AMSGradOptimizer(Optimizer):
     """
     https://colab.research.google.com/notebook#fileId=1xXFAuHM2Ae-OmF5M8Cn9ypGCa_HHBgfG&scrollTo=N1-2wPHN1Otn
@@ -1555,7 +1566,7 @@ class AMSGradOptimizer(Optimizer):
                 learning_rate /= tf.sqrt(self.t)
             update_ops = []
 
-            for (g, var) in gradient_variables:
+            for g, var in gradient_variables:
                 m = self.m[var].assign(self.beta1 * self.m[var] + (1 - self.beta1) * g)
                 v = self.v[var].assign(self.beta2 * self.v[var] + (1 - self.beta2) * g * g)
                 v_hat = self.v_hat[var].assign(tf.maximum(self.v_hat[var], v))

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/util/basic.py

@@ -2271,7 +2271,11 @@ def constant_with_shape(x, shape, dtype=None, name="constant_with_shape"):
     :rtype: tf.Tensor
     """
     with tf.name_scope(name):
-        if type(x) in [int, float, bool] and type(shape) in [list, tuple] and all([type(d) == int for d in shape]):
+        if (
+            isinstance(x, (int, float, bool))
+            and isinstance(shape, (list, tuple))
+            and all(isinstance(d, int) for d in shape)
+        ):
             if dtype is None:
                 dtype = {int: tf.int32, float: tf.float32, bool: tf.bool}[type(x)]
             if x in (0, 0.0, False):
@@ -3033,9 +3037,9 @@ class OpCodeCompiler(NativeCodeCompiler):
                 "tf_version": describe_tensorflow_version(),
                 "with_cuda": self._with_cuda(),
                 "cuda_path": self._cuda_env.cuda_path if self._with_cuda() else None,
-                "nvcc_opts": (tuple(self._cuda_env.get_compiler_opts()) + tuple(self._nvcc_opts))
-                if self._with_cuda()
-                else None,
+                "nvcc_opts": (
+                    (tuple(self._cuda_env.get_compiler_opts()) + tuple(self._nvcc_opts)) if self._with_cuda() else None
+                ),
             }
         )
         return d
@@ -3161,7 +3165,7 @@ def make_var_tuple(v):
     :rtype: tuple[tf.Tensor]
     """
     if isinstance(v, (int, float, tf.Tensor, tf.Operation)):
-        return (v,)
+        return (v,)  # noqa
     if isinstance(v, list):
         return tuple(v)
     assert isinstance(v, tuple)
@@ -5073,7 +5077,7 @@ def simplify_add(a, b):
     if isinstance(b, int):
         b = numpy.int32(b)  # use right type
     if isinstance(b, float):
-        b = numpy.float(b)
+        b = numpy.float(b)  # noqa
     if a.op.type in {"Add", "AddV2"}:
         a_dyn_parts = []
         a_const_parts = [b] if numpy.count_nonzero(b) > 0 else []
@@ -5469,10 +5473,7 @@ def gaussian_kernel_2d(size, std):
     if isinstance(size, (tuple, list)):
         size_x, size_y = size
     else:
-        size_x, size_y, = (
-            size,
-            size,
-        )
+        size_x, size_y = size, size
     if isinstance(std, (tuple, list)):
         std_x, std_y = std
     else:

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/tf/util/data.py

@@ -39,6 +39,7 @@ _MarkedDim = MarkedDim
 # Provide some simple wrappers. https://github.com/rwth-i6/returnn/issues/782
 # Use CamelCase function names (invalidates PEP8) to make it look like a class instance.
 
+
 # noinspection PyPep8Naming
 def FeatureDim(description, dimension, **kwargs):
     """
@@ -324,7 +325,7 @@ class BatchInfo:
                     # We want to get a reasonable order.
                     same_type_last_idx = None
                     for i, dim_ in enumerate(all_virtual_dims):
-                        if type(dim_) == type(dim):
+                        if type(dim_) == type(dim):  # noqa
                             same_type_last_idx = i
                     if same_type_last_idx is not None:
                         all_virtual_dims.insert(same_type_last_idx + 1, dim)

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/torch/data/pipeline.py

@@ -27,6 +27,7 @@ import numpy
 import torch
 import torch.utils.data
 
+from returnn.log import log
 from returnn.util.basic import NumbersDict
 
 
@@ -328,6 +329,14 @@ def create_data_loader_from_batches(
                 # See _DataLoaderWorkerPreInitFunc below, https://github.com/rwth-i6/returnn/issues/1495.
                 process_pre_init_func=SubProcCopyGlobalConfigPreInitFunc()
             )
+    else:
+        log.print_warning(
+            "Not using dedicated worker processes for torch data loading.\n"
+            'It is strongly recommended to set torch_dataloader_opts = {"num_workers": 1} '
+            "to improve GPU utilization and to work around some issues.\n"
+            "See e.g. https://github.com/rwth-i6/returnn/issues/1560 for an issue that is fixed by "
+            "using dedicated worker processes."
+        )
 
     return torch.utils.data.DataLoader(
         batches_dataset,

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/util/basic.py

@@ -1204,7 +1204,7 @@ def obj_diff_list(self, other, **kwargs):
         return ["%sself is None and other is %r" % (prefix, other)]
     if self is not None and other is None:
         return ["%sother is None and self is %r" % (prefix, self)]
-    if type(self) != type(other):
+    if type(self) != type(other):  # noqa
         return ["%stype diff: self is %s but other is %s" % (prefix, type(self).__name__, type(other).__name__)]
 
     if allowed_mapping:
@@ -1349,7 +1349,7 @@ def init_thread_join_hack():
         # https://github.com/albertz/playground/blob/master/thread-join-block.py
         # https://github.com/albertz/playground/blob/master/cond-wait-block.py
         return
-    main_thread = threading.currentThread()
+    main_thread = threading.current_thread()
     # noinspection PyUnresolvedReferences,PyProtectedMember
     assert isinstance(main_thread, threading._MainThread)
     main_thread_id = thread.get_ident()
@@ -1487,7 +1487,7 @@ def interrupt_main():
     :return: nothing
     """
     # noinspection PyProtectedMember,PyUnresolvedReferences
-    is_main_thread = isinstance(threading.currentThread(), threading._MainThread)
+    is_main_thread = isinstance(threading.current_thread(), threading._MainThread)
     if is_quitting():  # ignore if we are already quitting
         if is_main_thread:  # strange to get again in main thread
             raise Exception("interrupt_main() from main thread while already quitting")

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/util/better_exchook.py

@@ -1442,7 +1442,7 @@ def dump_all_thread_tracebacks(exclude_thread_ids=None, file=None):
             thread = threads.get(tid)
             if thread:
                 assert isinstance(thread, threading.Thread)
-                if thread is threading.currentThread():
+                if thread is threading.current_thread():
                     tags += ["current"]
                 # noinspection PyProtectedMember,PyUnresolvedReferences
                 if isinstance(thread, threading._MainThread):

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn/util/debug.py

@@ -75,7 +75,7 @@ def dump_all_thread_tracebacks(exclude_thread_ids=None, exclude_self=False):
             thread_ = threads.get(tid)
             if thread_:
                 assert isinstance(thread_, threading.Thread)
-                if thread_ is threading.currentThread():
+                if thread_ is threading.current_thread():
                     tags += ["current"]
                 # noinspection PyUnresolvedReferences,PyProtectedMember
                 if isinstance(thread_, threading._MainThread):
@@ -136,7 +136,7 @@ def init_better_exchook():
         # noinspection PyBroadException
         try:
             # noinspection PyUnresolvedReferences,PyProtectedMember
-            is_main_thread = isinstance(threading.currentThread(), threading._MainThread)
+            is_main_thread = isinstance(threading.current_thread(), threading._MainThread)
         except Exception:  # Can happen at a very late state while quitting.
             if exc_type is KeyboardInterrupt:
                 return
@@ -147,12 +147,12 @@ def init_better_exchook():
                     return
                 # An unhandled exception in the main thread. This means that we are going to quit now.
                 sys.exited = True
-        print("Unhandled exception %s in thread %s, proc %i." % (exc_type, threading.currentThread(), os.getpid()))
+        print("Unhandled exception %s in thread %s, proc %i." % (exc_type, threading.current_thread(), os.getpid()))
         if exc_type is KeyboardInterrupt:
             return
 
         # noinspection PyUnresolvedReferences,PyProtectedMember
-        if isinstance(threading.currentThread(), threading._MainThread):
+        if isinstance(threading.current_thread(), threading._MainThread):
             main_thread_id = thread.get_ident()
             if not isinstance(exc_type, Exception):
                 # We are the main thread and we got an exit-exception. This is likely fatal.
@@ -403,7 +403,7 @@ def init_ipython_kernel():
     # Do in mainthread to avoid history sqlite DB errors at exit.
     # https://github.com/ipython/ipython/issues/680
     # noinspection PyUnresolvedReferences,PyProtectedMember
-    assert isinstance(threading.currentThread(), threading._MainThread)
+    assert isinstance(threading.current_thread(), threading._MainThread)
     try:
         ip = socket.gethostbyname(socket.gethostname())
         connection_file = "ipython-kernel-%s-%s.json" % (ip, os.getpid())

{returnn-1.20240628.104058 → returnn-1.20240701.135153}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240628.104058
+Version: 1.20240701.135153
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer