returnn 1.20250304.101951__py3-none-any.whl → 1.20250304.113330__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250304.101951
+Version: 1.20250304.113330
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn/_setup_info_generated.py

@@ -1,2 +1,2 @@
-version = '1.20250304.101951'
-long_version = '1.20250304.101951+git.0fa434e'
+version = '1.20250304.113330'
+long_version = '1.20250304.113330+git.acf09da'

returnn/torch/engine.py

@@ -3,7 +3,7 @@ Main engine for PyTorch
 """
 
 from __future__ import annotations
-from typing import Optional, Any, Union, Callable, Dict, Set, Tuple
+from typing import Optional, Any, Union, Callable, Dict, Set
 from contextlib import nullcontext, ExitStack, contextmanager
 
 import gc
@@ -365,8 +365,6 @@ class Engine(EngineBase):
         zero_grad_next_step = True
         cur_count_grad_accum = 0
         extern_data = None
-        num_seqs = None
-        last_seq_idx = 0
 
         total_data_size_packed = NumbersDict()
         total_data_size_padded = NumbersDict()
@@ -400,20 +398,8 @@ class Engine(EngineBase):
                 )
 
                 complete_frac = float(extern_data_raw["complete_frac"])
-                num_seqs, last_seq_idx = _get_num_seqs_last_seq_idx(
-                    report_prefix=report_prefix,
-                    extern_data_raw=extern_data_raw,
-                    step_idx=step_idx,
-                    prev_num_seqs=num_seqs,
-                    prev_last_seq_idx=last_seq_idx,
-                )
-                epoch_continuous = (
-                    self.epoch - 1 + complete_frac
-                    if complete_frac >= 0.0
-                    else (self.epoch - 1 + (last_seq_idx + 1) / num_seqs)
-                    if num_seqs is not None
-                    else None
-                )
+                epoch_continuous = self.epoch - 1 + complete_frac if complete_frac >= 0.0 else None
+                num_seqs = int(extern_data_raw["num_seqs"])
 
                 # clear the gradients when every gradient accumulation loop starts
                 if zero_grad_next_step:
@@ -490,7 +476,7 @@ class Engine(EngineBase):
                     eval_info=dict(eval_info),
                     step_duration=step_duration,
                     start_elapsed=step_end_time - epoch_start_time,
-                    seq_idx=last_seq_idx,
+                    complete_frac=complete_frac,
                     num_seqs=num_seqs,
                     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,
@@ -629,13 +615,18 @@ class Engine(EngineBase):
             accumulated_losses_dict = NumbersDict()
             accumulated_inv_norm_factors_dict = NumbersDict()
             step_idx = 0
+            eval_start_time = time.monotonic()
 
+            report_prefix = f"ep {self.epoch} {dataset_name} eval"
             with torch.no_grad():
                 for extern_data_raw in data_loader:
                     if self._torch_distributed_ctx and step_idx % 100 == 0:
                         _has_data = torch.tensor([True], device="cpu", dtype=torch.int8)
                         torch.distributed.broadcast(_has_data, src=0)
 
+                    complete_frac = float(extern_data_raw["complete_frac"])
+                    num_seqs = int(extern_data_raw["num_seqs"])
+
                     extern_data = extern_data_util.raw_dict_to_extern_data(
                         extern_data_raw,
                         extern_data_template=self.extern_data,
@@ -644,6 +635,8 @@ class Engine(EngineBase):
                     )
 
                     self._run_step(extern_data, train_func=True)
+                    step_end_time = time.monotonic()
+
                     train_ctx = rf.get_run_ctx()
 
                     losses_dict = NumbersDict(
@@ -664,9 +657,12 @@ class Engine(EngineBase):
                     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",
+                        report_prefix,
                         step=step_idx,
                         eval_info=dict(eval_info),
+                        complete_frac=complete_frac,
+                        num_seqs=num_seqs,
+                        start_elapsed=step_end_time - eval_start_time,
                         log_memory_usage_device=self._device if self._log_memory_usage else None,
                     )
                     step_idx += 1
@@ -1290,8 +1286,6 @@ class Engine(EngineBase):
             new_dim.dyn_size_ext = _get_tensor_wo_batch_numpy(dim.dyn_size_ext)
             return new_dim
 
-        num_seqs = None
-        last_seq_idx = 0
         report_prefix = f"ep {self.epoch} {dataset.name} forward"
         with torch.no_grad():
             callback.init(model=self._orig_model)
@@ -1300,13 +1294,8 @@ class Engine(EngineBase):
             for extern_data_raw in data_loader:
                 step_begin_time = time.monotonic()
 
-                num_seqs, last_seq_idx = _get_num_seqs_last_seq_idx(
-                    report_prefix=report_prefix,
-                    extern_data_raw=extern_data_raw,
-                    step_idx=step_idx,
-                    prev_num_seqs=num_seqs,
-                    prev_last_seq_idx=last_seq_idx,
-                )
+                complete_frac = float(extern_data_raw["complete_frac"])
+                num_seqs = int(extern_data_raw["num_seqs"])
 
                 if self._forward_step_expected_outputs:
                     # Also resets any dyn dims, which might have been set in the prev step.
@@ -1354,7 +1343,7 @@ class Engine(EngineBase):
                     eval_info=None,
                     step_duration=step_duration,
                     start_elapsed=step_end_time - epoch_start_time,
-                    seq_idx=last_seq_idx,
+                    complete_frac=complete_frac,
                     num_seqs=num_seqs,
                     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,
@@ -1442,7 +1431,7 @@ def _print_process(
     batch_size_info: Optional[Dict[str, Any]] = None,
     step_duration: Optional[float] = None,
     start_elapsed: Optional[float] = None,
-    seq_idx: Optional[int] = None,
+    complete_frac: Optional[float] = None,
     num_seqs: Optional[int] = None,
     log_memory_usage_device: Optional[str] = None,
 ):
@@ -1455,11 +1444,14 @@ def _print_process(
     :param batch_size_info:
     :param step_duration: time elapsed for this step (secs)
     :param start_elapsed: time elapsed since epoch start (secs)
-    :param num_seqs: total number of sequences for this epoch
+    :param complete_frac: how much of the current epoch is already consumed
+    :param num_seqs: total number of seqs this epoch
     :param log_memory_usage_device: if given, will log memory usage (peak allocated memory)
     :return: nothing, will be printed to log
     """
     if log.verbose[5]:  # report every minibatch
+        if step == 0 and num_seqs is not None and num_seqs >= 0:
+            print(f"{report_prefix} num_seqs: {num_seqs}", file=log.v5)
         info = [report_prefix, "step %i" % step]
         if eval_info:  # Such as score.
             info += ["%s %s" % (k, _format_score_value(v)) for k, v in eval_info.items()]
@@ -1475,17 +1467,16 @@ def _print_process(
             info += ["%.3f sec/step" % step_duration]
         if start_elapsed is not None:
             info += ["elapsed %s" % hms(start_elapsed)]
-        if num_seqs is not None:
-            assert seq_idx is not None and start_elapsed is not None  # unexpected combination...
-            complete = (seq_idx + 1) / num_seqs
-            assert 1 >= complete > 0, f"{step} step, {num_seqs} num_seqs"
-            total_time_estimated = start_elapsed / complete
+        if complete_frac is not None:
+            assert 1 >= complete_frac > 0, f"{step} step, {complete_frac} complete_frac"
+            assert start_elapsed is not None
+            total_time_estimated = start_elapsed / complete_frac
             remaining_estimated = total_time_estimated - start_elapsed
             info += [
                 "exp. remaining %s" % hms(remaining_estimated),
-                "complete %.02f%%" % (complete * 100),
+                "complete %.02f%%" % (complete_frac * 100),
             ]
-        if start_elapsed is not None and num_seqs is None:
+        if start_elapsed is not None and complete_frac is None:
             info += ["(unk epoch len)"]
         print(", ".join(filter(None, info)), file=log.v5)
 
@@ -1634,27 +1625,3 @@ def _get_total_grad_norm(model: torch.nn.Module, p: float) -> float:
             p=p,
         ).item()
     )
-
-
-def _get_num_seqs_last_seq_idx(
-    *,
-    report_prefix: str,
-    extern_data_raw: Dict[str, Any],
-    step_idx: int,
-    prev_num_seqs: Optional[int],
-    prev_last_seq_idx: int,
-) -> Tuple[Optional[int], int]:
-    num_seqs = prev_num_seqs
-    num_seqs_ = int(extern_data_raw["num_seqs"]) if extern_data_raw.get("num_seqs", None) is not None else -1
-    # Note: The batches might have been shuffled,
-    # thus we cannot really assert that the seq_idx is always increasing.
-    last_seq_idx = max(int(extern_data_raw["seq_idx"].max()), prev_last_seq_idx)
-    if step_idx == 0:
-        if num_seqs_ >= 0:
-            print(f"{report_prefix} num_seqs: {num_seqs_}", file=log.v5)
-            num_seqs = num_seqs_
-    elif num_seqs_ >= 0:
-        assert num_seqs_ == num_seqs
-    if num_seqs is not None:
-        assert last_seq_idx < num_seqs
-    return num_seqs, last_seq_idx

{returnn-1.20250304.101951.dist-info → returnn-1.20250304.113330.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250304.101951
+Version: 1.20250304.113330
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250304.101951.dist-info → returnn-1.20250304.113330.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=33hja9F4qBUxg8Y8J2At3XpXgjYhstvueyHkyMN3GdI,5215
+returnn/PKG-INFO,sha256=BmSxZKkRxyL20E4Zsud1muiQ-rth9Ob9PMR-43IrAMw,5215
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=qBFbuB1yN3adgVM5pXt2-Yq9vorjRNchNPL8kDKx44M,31752
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=Sz2yGoXcP-7QpDNHDiaxopoVFqP8i7nfUQMH1Wss9YA,77
+returnn/_setup_info_generated.py,sha256=94BElbYUGmjpsoY8BzvfW39RUTXw9Fy3UwlPoEjrkU8,77
 returnn/config.py,sha256=3tmKhB6FnQZaNdtcYsiB61JnEY--iZ2qmJ4yq0b6tE0,29140
 returnn/forward_iface.py,sha256=A_OJiaXsX4MlXQRzST86ylyxSUZbC402PQL1REcqHjM,911
 returnn/learning_rate_control.py,sha256=ZvWryAn_tv9DhV8sh1LV3eE34Yltl3On3mYZAG4hR9s,34684
@@ -207,7 +207,7 @@ returnn/tf/util/open_fst.py,sha256=sZRDw4TbxvhGqpGdUJWy1ebvlZm4_RPhygpRw9uLAOQ,1
 returnn/torch/README.md,sha256=jzJ2FpOHW02vxN69yKaV97C9LI-hmvjBglKfdZXIDdc,85
 returnn/torch/__init__.py,sha256=MHEUyNHB20Vy89uKAqZoj6FxJKF1Gq3HW-i6ra1pNcI,24
 returnn/torch/distributed.py,sha256=skFyutdVztxgTEk3HHJ8S83qRWbNpkNT8Tj16Ic0_hE,6981
-returnn/torch/engine.py,sha256=sU9A96icaj65uaEkX4i4aUK3IrB2S19_Fb9_sueB_JE,77426
+returnn/torch/engine.py,sha256=2FLLb2m4sWFwYOQGREDSxQCheCKd_osnFJCdLa_4TzE,76400
 returnn/torch/updater.py,sha256=GqtBvZpElPVMm0lq84JPl4NVLFFETZAzAbR0rTomSao,28249
 returnn/torch/data/__init__.py,sha256=6cLNEi8KoGI12PF6akN7mI_mtjlx-0hcQAfMYoExwik,132
 returnn/torch/data/extern_data.py,sha256=_uT_9_gd5HIh1IoRsrebVG-nufSnb7fgC5jyU05GxJg,7580
@@ -253,8 +253,8 @@ returnn/util/sig_proc.py,sha256=Tjz0VOAVyqu2qDCF5HZ1JjALjcFsHcNkcd96WgZeKfE,7265
 returnn/util/task_system.py,sha256=y4sMVXQ25Qd2z0rx03uOlXlkE-jbCYC1Sjfn-XlraVU,26003
 returnn/util/train_proc_manager.py,sha256=Pjht28k6uz6BNQ47uW6Gf880iyq5q4wx7P_K2tmoAM8,3266
 returnn/util/watch_memory.py,sha256=BR5P2kvBN6UI81cE0_1WAA6Hd1SByLbBaiDxvLhPOew,4213
-returnn-1.20250304.101951.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250304.101951.dist-info/METADATA,sha256=33hja9F4qBUxg8Y8J2At3XpXgjYhstvueyHkyMN3GdI,5215
-returnn-1.20250304.101951.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-returnn-1.20250304.101951.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250304.101951.dist-info/RECORD,,
+returnn-1.20250304.113330.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250304.113330.dist-info/METADATA,sha256=BmSxZKkRxyL20E4Zsud1muiQ-rth9Ob9PMR-43IrAMw,5215
+returnn-1.20250304.113330.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+returnn-1.20250304.113330.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250304.113330.dist-info/RECORD,,