returnn 1.20240723.155747__tar.gz → 1.20240724.170410__tar.gz

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240723.155747
+Version: 1.20240724.170410
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240724.170410/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240724.170410'
+long_version = '1.20240724.170410+git.6775c30'

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/__main__.py

@@ -270,13 +270,13 @@ def print_task_properties():
         print("Train data:", file=log.v2)
         print("  input:", train_data.num_inputs, "x", train_data.window, file=log.v2)
         print("  output:", train_data.num_outputs, file=log.v2)
-        print(" ", train_data.len_info() or "no info", file=log.v2)
+        print(" ", train_data.len_info(fast=True) or "no info", file=log.v2)
     if dev_data:
         print("Dev data:", file=log.v2)
-        print(" ", dev_data.len_info() or "no info", file=log.v2)
+        print(" ", dev_data.len_info(fast=True) or "no info", file=log.v2)
     if eval_data:
         print("Eval data:", file=log.v2)
-        print(" ", eval_data.len_info() or "no info", file=log.v2)
+        print(" ", eval_data.len_info(fast=True) or "no info", file=log.v2)
 
 
 def init_engine():

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/audio.py

@@ -393,10 +393,12 @@ class OggZipDataset(CachedDataset2):
         self._lazy_init()
         return [self._get_tag_from_info_dict(seq) for seq in self._data]
 
-    def get_total_num_seqs(self):
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :rtype: int
         """
+        if fast and self._data is None:
+            raise Exception(f"{self} not initialized")
         self._lazy_init()
         return len(self._data)
 

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/basic.py

@@ -812,11 +812,12 @@ class Dataset(object):
         """
         raise OptionalNotImplementedError(f"{self} get_all_tags not implemented")
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
+        :param fast: if True, might raise an exception if not possible to get fast.
         :return: total number of seqs, without partition epoch.
-          Should be the same as len(self.get_all_tags()).
-          Note that this is not possible with all datasets.
+            Should be the same as len(self.get_all_tags()).
+            Note that this is not possible with all datasets.
         """
         raise OptionalNotImplementedError(f"{self} get_total_num_seqs not implemented")
 
@@ -994,16 +995,14 @@ class Dataset(object):
             return self.is_less_than_num_seqs(0)
         raise NotImplementedError(f"{self} have_seqs() is not implemented (and neither get_total_num_seqs())")
 
-    def len_info(self):
+    def len_info(self, *, fast: bool = False) -> str:
         """
-        :rtype: str
-        :returns a string to present the user as information about our len.
-        Depending on our implementation, we can give some more or some less information.
+        :return: string to present the user as information about our len.
         """
         return ", ".join(
             [
                 self.__class__.__name__,
-                "sequences: %s" % try_run(self.get_total_num_seqs, default="unknown"),
+                "sequences: %s" % try_run(self.get_total_num_seqs, kwargs=dict(fast=fast), default="unknown"),
                 "frames: %s" % try_run(self.get_num_timesteps, default="unknown"),
             ]
         )

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/generating.py

@@ -199,7 +199,7 @@ class GeneratingDataset(Dataset):
         """
         return self._num_seqs
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :return: total num seqs
         """
@@ -1189,7 +1189,7 @@ class StaticDataset(CachedDataset2):
         """
         return self.data[0][key].dtype
 
-    def get_total_num_seqs(self):
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :rtype: int
         """
@@ -2335,7 +2335,7 @@ class LibriSpeechCorpus(CachedDataset2):
         """
         return [self._get_tag(i) for i in range(len(self._reference_seq_order))]
 
-    def get_total_num_seqs(self):
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :rtype: int
         """

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/hdf.py

@@ -428,7 +428,7 @@ class HDFDataset(CachedDataset):
             tags += h5_file["seqTags"][...].tolist()
         return list(map(self._decode, tags))
 
-    def get_total_num_seqs(self):
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :rtype: int
         """
@@ -451,12 +451,6 @@ class HDFDataset(CachedDataset):
         """
         return self.data_dtype[key]
 
-    def len_info(self):
-        """
-        :rtype: str
-        """
-        return ", ".join(["HDF dataset", "sequences: %i" % self.num_seqs, "frames: %i" % self.get_num_timesteps()])
-
     def _get_file_index(self, real_seq_idx):
         # bisect() returns the position for which all elements to the left of the returned index are <= real_seq_idx,
         # so it actually returns the next file index in which the sequence can be found.

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/lm.py

@@ -7,6 +7,7 @@ and some related helpers.
 
 from __future__ import annotations
 
+from typing import Optional, Union, Callable, List, Tuple
 import typing
 import os
 import sys
@@ -17,7 +18,15 @@ import xml.etree.ElementTree as ElementTree
 import numpy
 from random import Random
 
-from returnn.util.basic import parse_orthography, parse_orthography_into_symbols, load_json, unicode, cf
+from returnn.util.basic import (
+    parse_orthography,
+    parse_orthography_into_symbols,
+    load_json,
+    unicode,
+    cf,
+    human_bytes_size,
+    hms,
+)
 from returnn.util.literal_py_to_pickle import literal_eval
 from returnn.log import log
 
@@ -115,8 +124,6 @@ class LmDataset(CachedDataset2):
         self._orth_replace_map_file = orth_replace_map_file
         self._phone_info = phone_info
 
-        if callable(corpus_file):
-            corpus_file = corpus_file()
         if callable(orth_symbols_file):
             orth_symbols_file = orth_symbols_file()
         if callable(orth_symbols_map_file):
@@ -124,8 +131,6 @@ class LmDataset(CachedDataset2):
         if callable(orth_replace_map_file):
             orth_replace_map_file = orth_replace_map_file()
 
-        print("LmDataset, loading file", corpus_file, file=log.v4)
-
         self.word_based = word_based
         self.word_end_symbol = word_end_symbol
         self.seq_end_symbol = seq_end_symbol
@@ -266,24 +271,74 @@ class LmDataset(CachedDataset2):
             self.num_outputs["delayed"] = self.num_outputs["data"]
             self.labels["delayed"] = self.labels["data"]
 
+        self._orth_tmp_file = None
+        self._orth_mmap = None
+        self._orths_offsets_and_lens: Optional[List[Tuple[int, int]]] = None  # will be loaded in _lazy_init
+
+        self.next_orth_idx = 0
+        self.next_seq_idx = 0
+        self.num_skipped = 0
+        self.num_unknown = 0
+
+    def _lazy_init(self):
+        if self._orths_offsets_and_lens is not None:
+            return
+
+        corpus_file = self._corpus_file
+        if callable(corpus_file):
+            corpus_file = corpus_file()
+
+        print("LmDataset, loading file", corpus_file, file=log.v4)
+
+        import tempfile
+        import mmap
+
+        tmp_file = tempfile.NamedTemporaryFile(prefix="returnn_lm_dataset_", suffix="_tmp.txt")
+        orths = []
+        offset = 0
+        self._orth_tmp_file = tmp_file
+        self._orths_offsets_and_lens = orths
+        start_time = time.time()
+        last_print_time = start_time
+
+        def _add_line(line: bytes):
+            nonlocal offset, last_print_time
+
+            orths.append((offset, len(line)))
+            tmp_file.write(line)
+            tmp_file.write(b"\n")
+            offset += len(line) + 1
+
+            if time.time() - last_print_time > 10:
+                print(
+                    f"  ..., loaded {len(self._orths_offsets_and_lens)} sequences,"
+                    f" {human_bytes_size(offset)},"
+                    f" after {hms(time.time() - start_time)}",
+                    file=log.v4,
+                )
+                last_print_time = time.time()
+
         if isinstance(corpus_file, list):  # If a list of files is provided, concatenate all.
-            self.orths = []
             for file_name in corpus_file:
-                if use_cache_manager:
+                if self._use_cache_manager:
                     file_name = cf(file_name)
-                self.orths += read_corpus(file_name, skip_empty_lines=skip_empty_lines)
+                iter_corpus(file_name, skip_empty_lines=self._skip_empty_lines, decode=False, callback=_add_line)
         else:
-            if use_cache_manager:
+            if self._use_cache_manager:
                 corpus_file = cf(corpus_file)
-            self.orths = read_corpus(corpus_file, skip_empty_lines=skip_empty_lines)
-        # It's only estimated because we might filter some out or so.
-        self._estimated_num_seqs = len(self.orths) // self.partition_epoch
-        print("  done, loaded %i sequences" % len(self.orths), file=log.v4)
+            iter_corpus(corpus_file, skip_empty_lines=self._skip_empty_lines, decode=False, callback=_add_line)
+
+        tmp_file.flush()
+        self._orth_mmap = mmap.mmap(tmp_file.fileno(), 0, flags=mmap.MAP_PRIVATE)
+        print(
+            f"  done, loaded {len(self._orths_offsets_and_lens)} sequences,"
+            f" {human_bytes_size(offset)},"
+            f" in {hms(time.time() - start_time)}",
+            file=log.v4,
+        )
 
-        self.next_orth_idx = 0
-        self.next_seq_idx = 0
-        self.num_skipped = 0
-        self.num_unknown = 0
+        # It's only estimated because we might filter some out or so.
+        self._estimated_num_seqs = len(self._orths_offsets_and_lens) // self.partition_epoch
 
     def get_data_keys(self):
         """
@@ -332,9 +387,14 @@ class LmDataset(CachedDataset2):
             self.seq_order = seq_order
         elif seq_list is not None:
             self.seq_order = [int(s[len(self._tag_prefix) :]) for s in seq_list]
+        elif epoch is None:
+            self.seq_order = []
         else:
+            self._lazy_init()
             self.seq_order = self.get_seq_order_for_epoch(
-                epoch=epoch, num_seqs=len(self.orths), get_seq_len=lambda i: len(self.orths[i])
+                epoch=epoch,
+                num_seqs=len(self._orths_offsets_and_lens),
+                get_seq_len=lambda i: self._orths_offsets_and_lens[i][1],
             )
         self.next_orth_idx = 0
         self.next_seq_idx = 0
@@ -348,9 +408,12 @@ class LmDataset(CachedDataset2):
         """supports sorting"""
         return True
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """total num seqs"""
-        return len(self.orths)
+        if fast and self._orths_offsets_and_lens is None:
+            raise Exception(f"{self} not initialized")
+        self._lazy_init()
+        return len(self._orths_offsets_and_lens)
 
     def _reduce_log_skipped_seqs(self):
         if isinstance(self.log_skipped_seqs, bool):
@@ -380,11 +443,14 @@ class LmDataset(CachedDataset2):
             if self.next_orth_idx >= len(self.seq_order):
                 assert self.next_seq_idx <= seq_idx, "We expect that we iterate through all seqs."
                 if self.num_skipped > 0:
-                    print("LmDataset: reached end, skipped %i sequences" % self.num_skipped)
+                    print("LmDataset: reached end, skipped %i sequences" % self.num_skipped, file=log.v2)
                 return None
             assert self.next_seq_idx == seq_idx, "We expect that we iterate through all seqs."
             true_idx = self.seq_order[self.next_orth_idx]
-            orth = self.orths[true_idx]  # get sequence for the next index given by seq_order
+            self._lazy_init()
+            # get sequence for the next index given by seq_order
+            offset, len_ = self._orths_offsets_and_lens[true_idx]
+            orth = self._orth_mmap[offset : offset + len_].decode("utf8")
             seq_tag = self._tag_prefix + str(true_idx)
             self.next_orth_idx += 1
             if orth == "</s>":
@@ -500,10 +566,10 @@ def _is_bliss(filename):
     return False
 
 
-def _iter_bliss(filename, callback):
+def _iter_bliss(filename: str, callback: Callable[[Union[str, bytes]], None], *, decode: bool = True):
     """
-    :param str filename:
-    :param (str)->None callback:
+    :param filename:
+    :param callback:
     """
     corpus_file = open(filename, "rb")
     if filename.endswith(".gz"):
@@ -534,51 +600,68 @@ def _iter_bliss(filename, callback):
         orth_split = orth_raw.split()
         orth = " ".join(orth_split)
 
+        if not decode:
+            orth = orth.encode("utf8")
         callback(orth)
 
 
-def _iter_txt(filename, callback, skip_empty_lines=True):
+def _iter_txt(
+    filename: str, callback: Callable[[Union[str, bytes]], None], *, skip_empty_lines: bool = True, decode: bool = True
+) -> None:
     """
-    :param str filename:
-    :param (str)->None callback:
-    :param bool skip_empty_lines:
+    :param filename:
+    :param callback:
+    :param skip_empty_lines:
+    :param decode:
     """
     f = open(filename, "rb")
     if filename.endswith(".gz"):
         f = gzip.GzipFile(fileobj=f)
 
     for line in f:
-        try:
-            line = line.decode("utf8")
-        except UnicodeDecodeError:
-            line = line.decode("latin_1")  # or iso8859_15?
+        if decode:
+            try:
+                line = line.decode("utf8")
+            except UnicodeDecodeError:
+                line = line.decode("latin_1")  # or iso8859_15?
         line = line.strip()
         if skip_empty_lines and not line:
             continue
         callback(line)
 
 
-def iter_corpus(filename, callback, skip_empty_lines=True):
+def iter_corpus(
+    filename: str, callback: Callable[[Union[str, bytes]], None], *, skip_empty_lines: bool = True, decode: bool = True
+) -> None:
     """
-    :param str filename:
-    :param ((str)->None) callback:
-    :param bool skip_empty_lines:
+    :param filename:
+    :param callback:
+    :param skip_empty_lines:
+    :param decode:
     """
     if _is_bliss(filename):
-        _iter_bliss(filename=filename, callback=callback)
+        _iter_bliss(filename=filename, callback=callback, decode=decode)
     else:
-        _iter_txt(filename=filename, callback=callback, skip_empty_lines=skip_empty_lines)
-
-
-def read_corpus(filename, skip_empty_lines=True):
-    """
-    :param str filename: either Bliss XML or line-based text
-    :param bool skip_empty_lines: in case of line-based text, skip empty lines
-    :return: list of orthographies
-    :rtype: list[str]
-    """
-    out_list = []
-    iter_corpus(filename=filename, callback=out_list.append, skip_empty_lines=skip_empty_lines)
+        _iter_txt(filename=filename, callback=callback, skip_empty_lines=skip_empty_lines, decode=decode)
+
+
+def read_corpus(
+    filename: str,
+    *,
+    skip_empty_lines: bool = True,
+    decode: bool = True,
+    out_list: Optional[Union[List[str], List[bytes]]] = None,
+) -> Union[List[str], List[bytes]]:
+    """
+    :param filename: either Bliss XML or line-based text
+    :param skip_empty_lines: in case of line-based text, skip empty lines
+    :param decode: if True, return str, otherwise bytes
+    :param out_list: if given, append to this list
+    :return: out_list, list of orthographies
+    """
+    if out_list is None:
+        out_list = []
+    iter_corpus(filename=filename, callback=out_list.append, skip_empty_lines=skip_empty_lines, decode=decode)
     return out_list
 
 

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/map.py

@@ -109,7 +109,7 @@ class MapDatasetWrapper(CachedDataset2):
             raise NotImplementedError("'num_seqs' is only known after calling init_seq_order().")
         return len(self._seq_order)
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :return: total number of seqs
         """

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/meta.py

@@ -459,7 +459,7 @@ class MetaDataset(CachedDataset2):
         """
         return self.seq_list_original[self.default_dataset_key]
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """
         :return: total number of seqs, without partition epoch
         """
@@ -1639,7 +1639,7 @@ class ConcatSeqsDataset(CachedDataset2):
         """
         return self.dataset.get_data_shape(key)
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """total num seqs"""
         return len(self.full_seq_list)
 
@@ -1879,9 +1879,9 @@ class VariableDataset(Dataset):
         """all tags"""
         return self._dataset.get_all_tags()
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """total num seqs"""
-        return self._dataset.get_total_num_seqs()
+        return self._dataset.get_total_num_seqs(fast=fast)
 
     def get_seq_length(self, sorted_seq_idx: int) -> NumbersDict:
         """seq len"""

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/multi_proc.py

@@ -332,7 +332,7 @@ class MultiProcDataset(CachedDataset2):
         """num seqs"""
         return self._num_seqs
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """total num seqs"""
         if self._total_num_seqs is not None:
             return self._total_num_seqs

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/numpy_dump.py

@@ -123,12 +123,6 @@ class NumpyDumpDataset(Dataset):
         """
         return self._num_seqs
 
-    def len_info(self):
-        """
-        :rtype: str
-        """
-        return "%s, %i seqs" % (self.__class__.__name__, self.num_seqs)
-
     # ------------ Seq cache management -----------
 
     def _cleanup_old_seq_cache(self, seq_end):

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/datasets/sprint.py

@@ -1223,7 +1223,7 @@ class SprintCacheDataset(CachedDataset2):
         self._num_seqs = len(self.seq_list_ordered)
         return True
 
-    def get_total_num_seqs(self) -> int:
+    def get_total_num_seqs(self, *, fast: bool = False) -> int:
         """total num seqs"""
         return len(self.seq_list_original)
 

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/frontend/decoder/transformer.py

@@ -1,5 +1,5 @@
 """
-(Label-sync) Transformer decoder, including cross attention to encoder
+(Label-sync) Transformer decoder, optionally including cross attention to encoder
 
 References:
 
@@ -81,7 +81,7 @@ class TransformerDecoder(rf.Module):
         if embed_dim:
             self.input_embedding_proj = rf.Linear(embed_dim, model_dim, with_bias=False)
 
-        # This could also be configurable...
+        # TODO This should be configurable...
         self.pos_enc = functools.partial(
             rf.sinusoidal_positional_encoding, feat_dim=embed_dim or model_dim, dtype=self.input_embedding.weight.dtype
         )

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn/torch/engine.py

@@ -18,6 +18,7 @@ from torch.utils.data import DataLoader
 from torch import autocast
 from torch.cuda import amp
 from random import random
+import math
 
 import returnn
 from returnn.config import Config
@@ -118,6 +119,7 @@ class Engine(EngineBase):
 
         self._log_memory_usage = config.bool("torch_log_memory_usage", False)
         self._log_batch_size = config.bool("log_batch_size", False) and log.verbose[5]
+        self._calculate_exp_loss = config.bool("calculate_exp_loss", False)
         self._reset_dev_memory_caches = config.bool("reset_dev_memory_caches", False)
         self._forward_auto_split_batch_on_oom = config.bool("forward_auto_split_batch_on_oom", False)
 
@@ -433,10 +435,11 @@ class Engine(EngineBase):
 
             accumulated_losses_dict += losses_dict
             accumulated_inv_norm_factors_dict += inv_norm_factors_dict
+            eval_info = self._maybe_extend_losses_info(losses_dict / inv_norm_factors_dict)
             _print_process(
                 f"ep {self.epoch} train",
                 step=step_idx,
-                eval_info=dict(losses_dict / inv_norm_factors_dict),
+                eval_info=dict(eval_info),
                 step_duration=step_duration,
                 batch_size_info=_get_batch_size_info(extern_data) if self._log_batch_size else None,
                 log_memory_usage_device=self._device if self._log_memory_usage else None,
@@ -463,6 +466,7 @@ class Engine(EngineBase):
         )
 
         accumulated_losses_dict = accumulated_losses_dict / accumulated_inv_norm_factors_dict
+        accumulated_losses_dict = self._maybe_extend_losses_info(accumulated_losses_dict)
         self.learning_rate_control.set_epoch_error(
             self.epoch, {f"train_loss_{k}": v for k, v in accumulated_losses_dict.items()}
         )
@@ -530,8 +534,8 @@ class Engine(EngineBase):
                     train_ctx = rf.get_run_ctx()
 
                     if score_keys is None:
-                        score_keys = [name for name, loss in train_ctx.losses.items() if not loss.as_error]
-                        error_keys = [name for name, loss in train_ctx.losses.items() if loss.as_error]
+                        score_keys = set(name for name, loss in train_ctx.losses.items() if not loss.as_error)
+                        error_keys = set(name for name, loss in train_ctx.losses.items() if loss.as_error)
 
                     losses_dict = NumbersDict(
                         {
@@ -549,16 +553,18 @@ class Engine(EngineBase):
 
                     accumulated_losses_dict += losses_dict
                     accumulated_inv_norm_factors_dict += inv_norm_factors_dict
+                    eval_info = self._maybe_extend_losses_info(losses_dict / inv_norm_factors_dict)
                     _print_process(
                         f"ep {self.epoch} {dataset_name} eval",
                         step=step_idx,
-                        eval_info=dict(losses_dict / inv_norm_factors_dict),
+                        eval_info=dict(eval_info),
                         log_memory_usage_device=self._device if self._log_memory_usage else None,
                     )
                     step_idx += 1
 
             assert step_idx > 0, f"No data in dataset {dataset_name!r}."
             accumulated_losses_dict = accumulated_losses_dict / accumulated_inv_norm_factors_dict
+            accumulated_losses_dict = self._maybe_extend_losses_info(accumulated_losses_dict)
 
             self.learning_rate_control.set_epoch_error(
                 self.epoch, {f"{dataset_name}_loss_{k}": v for k, v in accumulated_losses_dict.items()}
@@ -588,6 +594,23 @@ class Engine(EngineBase):
             assert isinstance(ls[0], self.learning_rate_control.EpochData)
             self.learning_rate_control.epoch_data[self.epoch] = ls[0]
 
+    def _maybe_extend_losses_info(self, losses: NumbersDict) -> NumbersDict:
+        """
+        :param losses:
+        :return: maybe extended losses
+        """
+        if self._calculate_exp_loss and losses.has_values():
+            # Assume the current run ctx still has info about the losses from the last step.
+            assert rf.get_run_ctx().losses
+            score_keys = set(k for k, v in rf.get_run_ctx().losses.items() if not v.as_error)
+            losses_ = {}
+            for key, value in losses.items():
+                losses_[key] = value
+                if key in score_keys:
+                    losses_[f"{key}:exp"] = math.exp(value)
+            losses = NumbersDict(losses_)
+        return losses
+
     def _create_data_loader(self, dataset: Dataset, *, train: bool = False) -> DataLoader:
         """
         :param dataset: RETURNN dataset

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240723.155747
+Version: 1.20240724.170410
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/tests/test_Dataset.py

@@ -722,12 +722,15 @@ def test_LmDataset_pickle():
             }
         )
         assert isinstance(dataset, LmDataset)
+        assert dataset._orths_offsets_and_lens is None  # not yet loaded, will be lazily loaded
 
         s = pickle.dumps(dataset)
         dataset = pickle.loads(s)
         assert isinstance(dataset, LmDataset)
+        assert dataset._orths_offsets_and_lens is None  # not yet loaded, will be lazily loaded
 
         dataset.init_seq_order(epoch=1)
+        assert dataset._orths_offsets_and_lens is not None  # loaded now
         dataset.load_seqs(0, 2)
         orth = dataset.get_data(0, "data")
         assert orth.tolist() == [1, 2]

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/tools/dump-dataset.py

@@ -281,7 +281,7 @@ def init(options):
     print("Dataset:", file=log.v2)
     print("  input:", dataset.num_inputs, "x", dataset.window, file=log.v2)
     print("  output:", dataset.num_outputs, file=log.v2)
-    print(" ", dataset.len_info() or "no info", file=log.v2)
+    print(" ", dataset.len_info(fast=True) or "no info", file=log.v2)
     if options.max_seq_length == "config":
         options.max_seq_length = config.typed_value("max_seq_length", sys.maxsize)
     elif options.max_seq_length:

{returnn-1.20240723.155747 → returnn-1.20240724.170410}/tools/hdf_dump.py

@@ -71,7 +71,7 @@ def init(config_filename, cmd_line_opts, dataset_config_str):
     else:
         assert dataset_config_str
         dataset = init_dataset(dataset_config_str)
-    print("Source dataset:", dataset.len_info(), file=log.v3)
+    print("Source dataset:", dataset.len_info(fast=True), file=log.v3)
     return dataset
 
 

returnn-1.20240723.155747/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240723.155747'
-long_version = '1.20240723.155747+git.4d772af'