returnn 1.20241017.4429__tar.gz → 1.20241018.213651__tar.gz

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241017.4429
+Version: 1.20241018.213651
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20241018.213651/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20241018.213651'
+long_version = '1.20241018.213651+git.2ac447f'

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/returnn/torch/data/pipeline.py

@@ -59,6 +59,9 @@ def collate_batch(batch: List[Dict[str, numpy.ndarray]]) -> Dict[str, Union[torc
 
     res = {}
     for key in data_keys:
+        if key == "num_seqs":
+            res[key] = batch[0][key]  # it should always be the same
+            continue
         ls = [create_tensor(sample[key]) for sample in batch]
         if not ls:
             raise ValueError("batch is empty?")
@@ -116,7 +119,7 @@ class ChunkingIterDataPipe(torch.utils.data.IterDataPipe):
 
             if not chunking_data_keys:
                 chunking_data_keys = list(data_dict.keys())  # use all if not configured separately
-                chunking_data_key_black_list = ["seq_tag"]
+                chunking_data_key_black_list = ["seq_tag", "seq_idx", "num_seqs"]
                 for key in chunking_data_key_black_list:
                     if key in chunking_data_keys:
                         chunking_data_keys.remove(key)

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/returnn/torch/data/returnn_dataset_wrapper.py

@@ -75,6 +75,13 @@ class ReturnnDatasetIterDataPipe(torch.utils.data.IterDataPipe):
         """
         :return: generator providing data samples in the form of a dict data_key -> data
         """
+        # noinspection PyBroadException
+        try:
+            num_seqs = self._dataset.num_seqs
+        except Exception:  # might not work for all datasets
+            num_seqs = -1
+        num_seqs = numpy.array(num_seqs)
+
         try:
             data_keys = self._dataset.get_data_keys()
 
@@ -83,6 +90,10 @@ class ReturnnDatasetIterDataPipe(torch.utils.data.IterDataPipe):
                 self._dataset.load_seqs(seq_index, seq_index + 1)
                 data = {data_key: self._dataset.get_data(seq_index, data_key) for data_key in data_keys}
                 data["seq_tag"] = str_to_numpy_array(self._dataset.get_tag(seq_index))
+                data["seq_idx"] = numpy.array(seq_index)
+                # It's slightly redundant to have num_seqs in each entry,
+                # but it's difficult to pass this back to the main proc otherwise.
+                data["num_seqs"] = num_seqs
                 yield data
                 seq_index += 1
 

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/returnn/torch/engine.py

@@ -252,7 +252,9 @@ class Engine(EngineBase):
 
         # Update learning rate
         self._updater.set_learning_rate(self.learning_rate)
-        self._updater.set_current_train_step(global_train_step=self.global_train_step, epoch=self.epoch)
+        self._updater.set_current_train_step(
+            global_train_step=self.global_train_step, epoch=self.epoch, epoch_continuous=self.epoch - 1
+        )
 
         self.learning_rate_control.epoch_data[self.epoch].meta.update(
             {
@@ -311,7 +313,7 @@ class Engine(EngineBase):
         accumulated_losses_dict = NumbersDict()
         accumulated_inv_norm_factors_dict = NumbersDict()
         step_idx = 0
-        epoch_start_time = time.time()
+        epoch_start_time = time.monotonic()
 
         data_iter = iter(self._train_dataloader)
         elapsed_computation_time = 0
@@ -339,12 +341,14 @@ class Engine(EngineBase):
         zero_grad_next_step = True
         cur_count_grad_accum = 0
         extern_data = None
+        num_seqs = None
+        last_seq_idx = 0
         try:
             while True:
                 with torch.no_grad():
                     extern_data_raw = next(data_iter, None)
 
-                step_begin_time = time.time()
+                step_begin_time = time.monotonic()
 
                 _has_data = torch.tensor([extern_data_raw is not None], dtype=torch.int8)
                 if self._torch_distributed_ctx:
@@ -353,6 +357,22 @@ class Engine(EngineBase):
                     torch.distributed.all_reduce(_has_data, op=torch.distributed.ReduceOp.MIN)
                 if not _has_data[0]:
                     break
+                num_seqs_ = (
+                    int(extern_data_raw["num_seqs"]) if extern_data_raw.get("num_seqs", None) is not None else -1
+                )
+                last_seq_idx_ = extern_data_raw["seq_idx"].max()
+                assert last_seq_idx_ >= last_seq_idx
+                last_seq_idx = int(last_seq_idx_)
+                del last_seq_idx_
+                if step_idx == 0:
+                    if num_seqs_ >= 0:
+                        print(f"Epoch {self.epoch} num_seqs: {num_seqs_}", file=log.v5)
+                        num_seqs = num_seqs_
+                elif num_seqs_ >= 0:
+                    assert num_seqs_ == num_seqs
+                del num_seqs_
+                if num_seqs is not None:
+                    assert last_seq_idx < num_seqs
 
                 # clear the gradients when every gradient accumulation loop starts
                 if zero_grad_next_step:
@@ -404,7 +424,8 @@ class Engine(EngineBase):
                 if self._torch_distributed_ctx:
                     self._torch_distributed_ctx.step_after_param_update(module=self._pt_model, epoch_step_idx=step_idx)
 
-                step_duration = time.time() - step_begin_time
+                step_end_time = time.monotonic()
+                step_duration = step_end_time - step_begin_time
                 elapsed_computation_time += step_duration
 
                 accumulated_losses_dict += losses_dict
@@ -415,6 +436,9 @@ class Engine(EngineBase):
                     step=step_idx,
                     eval_info=dict(eval_info),
                     step_duration=step_duration,
+                    start_elapsed=step_end_time - epoch_start_time,
+                    seq_idx=last_seq_idx,
+                    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,
                 )
@@ -431,12 +455,16 @@ class Engine(EngineBase):
 
                 step_idx += 1
                 self.global_train_step += 1
-                self._updater.set_current_train_step(global_train_step=self.global_train_step, epoch=self.epoch)
+                self._updater.set_current_train_step(
+                    global_train_step=self.global_train_step,
+                    epoch=self.epoch,
+                    epoch_continuous=(self.epoch - 1 + (last_seq_idx + 1) / num_seqs) if num_seqs is not None else None,
+                )
         except Exception as exc:
             help_on_torch_exception(exc, step_idx=step_idx, model=self._orig_model, extern_data=extern_data)
             raise
 
-        elapsed = time.time() - epoch_start_time
+        elapsed = time.monotonic() - epoch_start_time
         elapsed_computation_percentage = elapsed_computation_time / elapsed
         print(
             "Trained %i steps, %s elapsed (%.1f%% computing time)"
@@ -1008,7 +1036,7 @@ class Engine(EngineBase):
         assert isinstance(dataset, Dataset)
         assert isinstance(callback, ForwardCallbackIface)
 
-        epoch_start_time = time.time()
+        epoch_start_time = time.monotonic()
         elapsed_computation_time = 0.0
 
         self._pt_model.eval()
@@ -1087,7 +1115,7 @@ class Engine(EngineBase):
 
             step_idx = 0
             for extern_data_raw in data_loader:
-                step_begin_time = time.time()
+                step_begin_time = time.monotonic()
                 if self._forward_step_expected_outputs:
                     # Also resets any dyn dims, which might have been set in the prev step.
                     self._forward_step_expected_outputs.reset_content()
@@ -1121,7 +1149,7 @@ class Engine(EngineBase):
                         model_outputs_per_batch.data[k] = _get_tensor_wo_batch_numpy(v)
                     callback.process_seq(seq_tag=seq_tag, outputs=model_outputs_per_batch)
 
-                elapsed_computation_time += time.time() - step_begin_time
+                elapsed_computation_time += time.monotonic() - step_begin_time
                 _print_process(
                     report_prefix,
                     step=step_idx,
@@ -1132,7 +1160,7 @@ class Engine(EngineBase):
 
             callback.finish()
 
-        elapsed = time.time() - epoch_start_time
+        elapsed = time.monotonic() - epoch_start_time
         elapsed_computation_percentage = elapsed_computation_time / elapsed
         print(
             "Forward %i steps, %s elapsed (%.1f%% computing time)"
@@ -1202,20 +1230,26 @@ def _to_raw(n: Union[int, float, Tensor]):
 
 def _print_process(
     report_prefix: str,
+    *,
     step: int,
     eval_info: Optional[Dict[str, Any]] = None,
     batch_size_info: Optional[Dict[str, Any]] = None,
     step_duration: Optional[float] = None,
+    start_elapsed: Optional[float] = None,
+    seq_idx: Optional[int] = None,
+    num_seqs: Optional[int] = None,
     log_memory_usage_device: Optional[str] = None,
 ):
     """
     Similar but simplified from TF engine _print_process.
 
     :param report_prefix:
-    :param step:
+    :param step: for this epoch
     :param eval_info:
     :param batch_size_info:
-    :param step_duration:
+    :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 log_memory_usage_device: if given, will log memory usage (peak allocated memory)
     :return: nothing, will be printed to log
     """
@@ -1233,6 +1267,20 @@ def _print_process(
                 ]
         if step_duration is not None:
             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
+            remaining_estimated = total_time_estimated - start_elapsed
+            info += [
+                "exp. remaining %s" % hms(remaining_estimated),
+                "complete %.02f%%" % (complete * 100),
+            ]
+        if start_elapsed is not None and num_seqs is None:
+            info += ["(unk epoch len)"]
         print(", ".join(filter(None, info)), file=log.v5)
 
 

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/returnn/torch/updater.py

@@ -13,7 +13,7 @@ import typing
 
 import returnn
 from returnn.log import log
-from returnn.util.basic import RefIdEq
+from returnn.util.basic import RefIdEq, get_fwd_compat_kwargs
 import returnn.frontend as rf
 from returnn.torch.frontend.bridge import wrapped_pt_module_to_rf_module
 
@@ -96,8 +96,11 @@ class Updater:
         self._effective_learning_rate = self.learning_rate
         self.network = network
         self._device = device
+        # Just set the very first step as initial values here.
+        # They will be overwritten via set_current_train_step() below.
         self._current_train_step = 0
-        self._current_epoch = 0
+        self._current_epoch = 1
+        self._current_epoch_continuous = 0.0
 
         self.learning_rate_function = self.config.typed_value("dynamic_learning_rate", None)
         if self.learning_rate_function is not None:
@@ -163,19 +166,38 @@ class Updater:
         self._effective_learning_rate = self.learning_rate
         if self.learning_rate_function is not None:
             lr = self.learning_rate_function(
-                global_train_step=self._current_train_step, epoch=self._current_epoch, learning_rate=self.learning_rate
+                global_train_step=self._current_train_step,
+                epoch=self._current_epoch,
+                epoch_continuous=self._current_epoch_continuous,
+                learning_rate=self.learning_rate,
+                **get_fwd_compat_kwargs(),
             )
             self._effective_learning_rate = float(lr)
         if self.optimizer:
             for param_group in self.optimizer.param_groups:
                 param_group["lr"] = self._effective_learning_rate
 
-    def set_current_train_step(self, *, global_train_step: int, epoch: int):
+    def set_current_train_step(self, *, global_train_step: int, epoch: int, epoch_continuous: Optional[float] = None):
         """
         Obtains an updated learning rate for the current training step inside a (sub)epoch.
+
+        :param global_train_step: Current global training step over the whole training process.
+            In the first epoch, this starts at 0.
+        :param epoch: Current epoch. (First epoch is 1 by RETURNN convention.)
+        :param epoch_continuous: How much of the epoch is finished.
+            In the first step of the first epoch, this starts at 0.0,
+            and when the fist epoch is finished, this reaches 1.0,
+            and the values in between are the fraction of the epoch that is finished.
+            The second epoch (epoch=2) starts at 1.0,
+            and when the second epoch is finished, this reaches 2.0, and so on.
+            We usually calculate this based on ``epoch-1+(last_seq_idx+1)/num_seqs``,
+            if the dataset can provide ``num_seqs``.
+            Other schemes based on the step_idx might be used as well to calculate this,
+            if the number of steps per epoch is known in advance.
         """
         self._current_train_step = global_train_step
         self._current_epoch = epoch
+        self._current_epoch_continuous = epoch_continuous
         self._update_effective_learning_rate()
 
     def step(self, *, grad_scaler: Optional[torch.cuda.amp.GradScaler] = None):

{returnn-1.20241017.4429 → returnn-1.20241018.213651}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20241017.4429
+Version: 1.20241018.213651
 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 → returnn-1.20241018.213651}/tests/test_torch_engine.py

@@ -4,6 +4,7 @@ Tests for PyTorch engine.
 
 from __future__ import annotations
 import _setup_test_env  # noqa
+from typing import Optional
 import sys
 import unittest
 import tempfile
@@ -294,7 +295,6 @@ def test_forward_beam_seq_lens():
 
 
 def test_min_seq_len():
-
     from returnn.datasets.generating import DummyDataset
 
     config = Config({"min_seq_length": 2, "batch_size": 3})
@@ -318,7 +318,6 @@ def test_min_seq_len():
 
 
 def test_max_seq_len():
-
     from returnn.datasets.generating import DummyDataset
 
     config = Config({"max_seq_length": 4, "batch_size": 3})
@@ -520,6 +519,71 @@ def test_torch_engine_train_exception():
             raise Exception("did not get expected exception")
 
 
+def test_dynamic_learning_rate():
+    num_epochs = 3
+    last_global_train_step: Optional[float] = None
+    last_epoch_continuous: Optional[float] = None
+    epoch_continuous_diffs = []
+
+    def _dynamic_learning_rate(
+        *, global_train_step: int, epoch: int, epoch_continuous: float, learning_rate: float, **_kwargs
+    ) -> float:
+        nonlocal last_global_train_step, last_epoch_continuous
+        assert isinstance(global_train_step, int)
+        assert isinstance(epoch, int)
+        assert isinstance(epoch_continuous, (int, float))
+        assert isinstance(learning_rate, (int, float))
+        print(f"global_train_step: {global_train_step}, epoch: {epoch}, epoch_continuous: {epoch_continuous}")
+        if last_global_train_step is None:
+            assert global_train_step == 0 and epoch == 1
+        else:
+            # The call to this function could be repeated.
+            assert global_train_step in (last_global_train_step, last_global_train_step + 1)
+        if last_epoch_continuous is None:
+            assert epoch_continuous == 0
+        elif global_train_step == last_global_train_step:  # repeated call
+            assert epoch_continuous == last_epoch_continuous
+        else:
+            assert epoch_continuous > last_epoch_continuous
+            assert epoch >= epoch_continuous >= epoch - 1
+            epoch_continuous_diffs.append(epoch_continuous - last_epoch_continuous)
+        last_global_train_step = global_train_step
+        last_epoch_continuous = epoch_continuous
+        return learning_rate * epoch_continuous / num_epochs
+
+    config = Config(
+        dict(
+            task="train",
+            device="cpu",
+            extern_data={"data": {"dim": 9}, "classes": {"dim": 2, "sparse": True}},
+            get_model=TrainTestModel,
+            train_step=TrainTestModel.train_step,
+            batch_size=500,
+            optimizer={"class": "adam"},
+            dynamic_learning_rate=_dynamic_learning_rate,
+            num_epochs=num_epochs,
+        )
+    )
+    num_seqs_per_epoch = 100
+    dataset = init_dataset({"class": "Task12AXDataset", "num_seqs": num_seqs_per_epoch, "name": "train"})
+    dataset.init_seq_order(epoch=1)
+
+    with global_config_ctx(config):
+        engine = Engine(config=config)
+        engine.init_train_from_config(train_data=dataset)
+        engine.train()
+
+    assert last_epoch_continuous == num_epochs
+    assert epoch_continuous_diffs
+    print("epoch continuous diffs:", epoch_continuous_diffs)
+    # Just some sanity check. The exact number here depends on num_seqs_per_epoch, batch_size, etc.
+    assert numpy.min(epoch_continuous_diffs) >= 0.01
+    assert numpy.max(epoch_continuous_diffs) <= 0.1
+    # It's one more (non-repeated) call than num steps (first + very last),
+    # and the diffs is one less, so the length should match final global train step.
+    assert len(epoch_continuous_diffs) == engine.global_train_step
+
+
 if __name__ == "__main__":
     better_exchook.install()
     if len(sys.argv) <= 1:

returnn-1.20241017.4429/_setup_info_generated.py

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