[bugfix] Resolve PyTorch Profiling for Manual Optimization (#9316)

Co-authored-by: Carlos Mocholí <[email protected]>

Author: 2 people

CHANGELOG.md

@@ -37,6 +37,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed error handling in DDP process reconciliation when `_sync_dir` was not initialized ([#9267](https://github.com/PyTorchLightning/pytorch-lightning/pull/9267))
 
 
+- Fixed PyTorch Profiler not enabled for manual optimization ([#9316](https://github.com/PyTorchLightning/pytorch-lightning/pull/9316))
+
+
 - Fixed inspection of other args when a container is specified in `save_hyperparameters` ([#9125](https://github.com/PyTorchLightning/pytorch-lightning/pull/9125))
 
 ## [1.4.5] - 2021-08-31

pl_examples/basic_examples/profiler_example.py

@@ -31,6 +31,7 @@
 
 from pl_examples import _DATASETS_PATH, cli_lightning_logo
 from pytorch_lightning import LightningDataModule, LightningModule
+from pytorch_lightning.profiler.pytorch import PyTorchProfiler
 from pytorch_lightning.utilities.cli import LightningCLI
 
 DEFAULT_CMD_LINE = (
@@ -43,18 +44,34 @@
 
 
 class ModelToProfile(LightningModule):
-    def __init__(self, name: str = "resnet50"):
+    def __init__(self, name: str = "resnet18", automatic_optimization: bool = True):
         super().__init__()
         self.model = getattr(models, name)(pretrained=True)
         self.criterion = torch.nn.CrossEntropyLoss()
+        self.automatic_optimization = automatic_optimization
+        self.training_step = (
+            self.automatic_optimization_training_step
+            if automatic_optimization
+            else self.manual_optimization_training_step
+        )
 
-    def training_step(self, batch, batch_idx):
+    def automatic_optimization_training_step(self, batch, batch_idx):
         inputs, labels = batch
         outputs = self.model(inputs)
         loss = self.criterion(outputs, labels)
         self.log("train_loss", loss)
         return loss
 
+    def manual_optimization_training_step(self, batch, batch_idx):
+        opt = self.optimizers()
+        opt.zero_grad()
+        inputs, labels = batch
+        outputs = self.model(inputs)
+        loss = self.criterion(outputs, labels)
+        self.log("train_loss", loss)
+        self.manual_backward(loss)
+        opt.step()
+
     def validation_step(self, batch, batch_idx):
         inputs, labels = batch
         outputs = self.model(inputs)
@@ -77,18 +94,20 @@ def train_dataloader(self, *args, **kwargs):
         trainset = torchvision.datasets.CIFAR10(
             root=_DATASETS_PATH, train=True, download=True, transform=self.transform
         )
-        return torch.utils.data.DataLoader(trainset, batch_size=32, shuffle=True, num_workers=0)
+        return torch.utils.data.DataLoader(trainset, batch_size=2, shuffle=True, num_workers=0)
 
     def val_dataloader(self, *args, **kwargs):
         valset = torchvision.datasets.CIFAR10(root=_DATASETS_PATH, train=False, download=True, transform=self.transform)
-        return torch.utils.data.DataLoader(valset, batch_size=32, shuffle=True, num_workers=0)
+        return torch.utils.data.DataLoader(valset, batch_size=2, shuffle=True, num_workers=0)
 
 
 def cli_main():
     if len(sys.argv) == 1:
         sys.argv += DEFAULT_CMD_LINE
 
-    LightningCLI(ModelToProfile, CIFAR10DataModule)
+    LightningCLI(
+        ModelToProfile, CIFAR10DataModule, save_config_overwrite=True, trainer_defaults={"profiler": PyTorchProfiler()}
+    )
 
 
 if __name__ == "__main__":

pytorch_lightning/profiler/pytorch.py

@@ -15,7 +15,7 @@
 import inspect
 import logging
 import os
-from functools import partial
+from functools import lru_cache, partial
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional, Set, Type, TYPE_CHECKING, Union
 
@@ -24,9 +24,10 @@
 from torch.autograd.profiler import record_function
 
 from pytorch_lightning.profiler.base import BaseProfiler
-from pytorch_lightning.utilities import rank_zero_deprecation, rank_zero_warn
+from pytorch_lightning.utilities import rank_zero_warn
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _KINETO_AVAILABLE
+from pytorch_lightning.utilities.warnings import WarningCache
 
 if TYPE_CHECKING:
     from torch.autograd.profiler import EventList
@@ -38,6 +39,7 @@
     from torch.profiler import ProfilerAction, ProfilerActivity, tensorboard_trace_handler
 
 log = logging.getLogger(__name__)
+warning_cache = WarningCache()
 
 _PROFILER = Union[torch.autograd.profiler.profile, torch.cuda.profiler.profile, torch.autograd.profiler.emit_nvtx]
 
@@ -116,6 +118,7 @@ def pre_step(self, current_action: str) -> None:
         self._current_action = current_action
 
     def reset(self):
+        # handle properly `fast_dev_run`. PyTorch Profiler will fail otherwise.
         self._num_optimizer_step_and_closure = 0
         self._num_validation_step = 0
         self._num_test_step = 0
@@ -128,9 +131,15 @@ def reset(self):
         self._current_action: Optional[str] = None
         self._start_action_name: Optional[str] = None
 
+    @property
+    def is_training(self) -> bool:
+        return self._current_action is not None and (
+            self._current_action.startswith("optimizer_step_and_closure_") or self._current_action == "training_step"
+        )
+
     @property
     def num_step(self) -> int:
-        if self._current_action is not None and self._current_action.startswith("optimizer_step_and_closure_"):
+        if self.is_training:
             return self._num_optimizer_step_and_closure
         if self._current_action == "validation_step":
             return self._num_validation_step
@@ -141,7 +150,7 @@ def num_step(self) -> int:
         return 0
 
     def _step(self) -> None:
-        if self._current_action is not None and self._current_action.startswith("optimizer_step_and_closure_"):
+        if self.is_training:
             self._num_optimizer_step_and_closure += 1
         elif self._current_action == "validation_step":
             if self._start_action_name == "on_fit_start":
@@ -156,7 +165,7 @@ def _step(self) -> None:
 
     @property
     def has_finished(self) -> bool:
-        if self._current_action is not None and self._current_action.startswith("optimizer_step_and_closure_"):
+        if self.is_training:
             return self._optimizer_step_and_closure_reached_end
         if self._current_action == "validation_step":
             return self._validation_step_reached_end
@@ -172,9 +181,9 @@ def __call__(self, num_step: int) -> "ProfilerAction":
             return ProfilerAction.NONE
 
         self._step()
-        action = self._schedule(self.num_step)
+        action = self._schedule(max(self.num_step, 0))
         if action == ProfilerAction.RECORD_AND_SAVE:
-            if self._current_action is not None and self._current_action.startswith("optimizer_step_and_closure_"):
+            if self.is_training:
                 self._optimizer_step_and_closure_reached_end = True
             elif self._current_action == "validation_step":
                 self._validation_step_reached_end = True
@@ -196,7 +205,7 @@ class PyTorchProfiler(BaseProfiler):
         "predict_step",
     }
     RECORD_FUNCTION_PREFIX = "optimizer_step_and_closure_"
-    STEP_FUNCTIONS = {"validation_step", "test_step", "predict_step"}
+    STEP_FUNCTIONS = {"training_step", "validation_step", "test_step", "predict_step"}
     STEP_FUNCTION_PREFIX = "optimizer_step_and_closure_"
     AVAILABLE_SORT_KEYS = {
         "cpu_time",
@@ -320,6 +329,7 @@ def _init_kineto(self, profiler_kwargs: Any) -> None:
                 raise MisconfigurationException(
                     f"Schedule should return a `torch.profiler.ProfilerAction`. Found: {action}"
                 )
+        self._default_schedule()
         schedule = schedule if has_schedule else self._default_schedule()
         self._schedule = ScheduleWrapper(schedule) if schedule is not None else schedule
         self._profiler_kwargs["schedule"] = self._schedule
@@ -331,28 +341,13 @@ def _init_kineto(self, profiler_kwargs: Any) -> None:
         with_stack = profiler_kwargs.get("with_stack", False) or self._export_to_flame_graph
         self._profiler_kwargs["with_stack"] = with_stack
 
-    def __deprecation_check(
-        self, profiled_functions: Optional[List[str]], record_functions: Optional[Set[str]]
-    ) -> Set[str]:
-        if record_functions is None:
-            record_functions = set()
-
-        if profiled_functions is not None:
-            rank_zero_deprecation(
-                "`PyTorchProfiler.profiled_functions` has been renamed to"
-                " `record_functions` in v1.3 and will be removed in v1.5"
-            )
-            if not record_functions:
-                record_functions |= set(profiled_functions)
-            else:
-                raise MisconfigurationException(
-                    "You set `PytorchProfiler.profiled_functions` and `PyTorchProfiler.record_functions`."
-                    "  Please use only the later."
-                )
-
-        return record_functions
+    def _should_override_schedule(self) -> bool:
+        return (self._lightning_module is not None and self._lightning_module.trainer.limit_train_batches < 5) and (
+            self._schedule is not None and self._schedule._schedule == self._default_schedule()
+        )
 
     @staticmethod
+    @lru_cache(1)
     def _default_schedule() -> Optional[callable]:
         if _KINETO_AVAILABLE:
             # Those schedule defaults allow the profiling overhead to be negligible over training time.
@@ -393,11 +388,18 @@ def start(self, action_name: str) -> None:
             if self._register is not None:
                 self._register.__enter__()
 
+        if self._lightning_module is not None:
+            # when the model is used in automatic optimization,
+            # we use `optimizer_step_and_closure` to step the model.
+            if self._lightning_module.automatic_optimization and "training_step" in self.STEP_FUNCTIONS:
+                self.STEP_FUNCTIONS.remove("training_step")
+
         if (
             self.profiler is not None
             and (action_name in self._record_functions or action_name.startswith(self.RECORD_FUNCTION_PREFIX))
             and action_name not in self._recording_map
         ):
+
             recording = record_function(action_name)
             recording.__enter__()
             self._recording_map[action_name] = recording
@@ -413,6 +415,17 @@ def stop(self, action_name: str) -> None:
         if self.profiler is not None and (
             action_name in self.STEP_FUNCTIONS or action_name.startswith(self.STEP_FUNCTION_PREFIX)
         ):
+
+            # the default schedule requires a minimum of 5 steps to properly work: `wait=1, warmup=1, active=3`.
+            # otherwise, this will raise a `segmentation fault`.
+            if self._should_override_schedule():
+                warning_cache.warn(
+                    "The PyTorch Profiler default schedule will be overridden as there is not enough "
+                    "steps to properly record traces."
+                )
+                self._schedule = None
+                self.profiler.schedule = torch.profiler.profiler._default_schedule_fn
+
             if self._schedule is not None:
                 self._schedule.pre_step(action_name)
 

tests/helpers/__init__.py

@@ -1,2 +1,7 @@
-from tests.helpers.boring_model import BoringDataModule, BoringModel, RandomDataset  # noqa: F401
+from tests.helpers.boring_model import (  # noqa: F401
+    BoringDataModule,
+    BoringModel,
+    ManualOptimBoringModel,
+    RandomDataset,
+)
 from tests.helpers.datasets import TrialMNIST  # noqa: F401

tests/helpers/boring_model.py

@@ -184,3 +184,18 @@ def test_dataloader(self):
 
     def predict_dataloader(self):
         return DataLoader(self.random_predict)
+
+
+class ManualOptimBoringModel(BoringModel):
+    def __init__(self):
+        super().__init__()
+        self.automatic_optimization = False
+
+    def training_step(self, batch, batch_idx):
+        opt = self.optimizers()
+        output = self(batch)
+        loss = self.loss(batch, output)
+        opt.zero_grad()
+        self.manual_backward(loss)
+        opt.step()
+        return loss

tests/profiler/test_profiler.py

@@ -29,7 +29,7 @@
 from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_7
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _KINETO_AVAILABLE
-from tests.helpers import BoringModel
+from tests.helpers import BoringModel, ManualOptimBoringModel
 from tests.helpers.runif import RunIf
 
 PROFILER_OVERHEAD_MAX_TOLERANCE = 0.0005
@@ -309,50 +309,48 @@ def test_pytorch_profiler_trainer_ddp(tmpdir, pytorch_profiler):
         assert any(f"{local_rank}-validation_step" in f for f in files)
 
 
-def test_pytorch_profiler_trainer_test(tmpdir):
+@pytest.mark.parametrize("fast_dev_run", [1, 2, 3, 4, 5])
+@pytest.mark.parametrize("boring_model_cls", [ManualOptimBoringModel, BoringModel])
+def test_pytorch_profiler_trainer_fit(fast_dev_run, boring_model_cls, tmpdir):
     """Ensure that the profiler can be given to the trainer and test step are properly recorded."""
-    pytorch_profiler = PyTorchProfiler(dirpath=tmpdir, filename="profile", schedule=None)
-    model = BoringModel()
-    trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, limit_test_batches=2, profiler=pytorch_profiler)
-    trainer.test(model)
+    pytorch_profiler = PyTorchProfiler(dirpath=tmpdir, filename="profile")
+    model = boring_model_cls()
+    trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, fast_dev_run=fast_dev_run, profiler=pytorch_profiler)
+    trainer.fit(model)
 
-    assert sum(e.name == "test_step" for e in pytorch_profiler.function_events)
+    assert sum(e.name == "validation_step" for e in pytorch_profiler.function_events)
 
-    path = pytorch_profiler.dirpath / f"test-{pytorch_profiler.filename}.txt"
+    path = pytorch_profiler.dirpath / f"fit-{pytorch_profiler.filename}.txt"
     assert path.read_text("utf-8")
 
     if _KINETO_AVAILABLE:
         files = sorted(file for file in os.listdir(tmpdir) if file.endswith(".json"))
-        assert any(f"test-{pytorch_profiler.filename}" in f for f in files)
-        path = pytorch_profiler.dirpath / f"test-{pytorch_profiler.filename}.txt"
+        assert any(f"fit-{pytorch_profiler.filename}" in f for f in files)
+        path = pytorch_profiler.dirpath / f"fit-{pytorch_profiler.filename}.txt"
         assert path.read_text("utf-8")
 
 
-def test_pytorch_profiler_trainer_predict(tmpdir):
-    """Ensure that the profiler can be given to the trainer and predict function are properly recorded."""
+@pytest.mark.parametrize("fn, step_name", [("test", "test"), ("validate", "validation"), ("predict", "predict")])
+@pytest.mark.parametrize("boring_model_cls", [BoringModel, ManualOptimBoringModel])
+def test_pytorch_profiler_trainer(fn, step_name, boring_model_cls, tmpdir):
+    """Ensure that the profiler can be given to the trainer and test step are properly recorded."""
     pytorch_profiler = PyTorchProfiler(dirpath=tmpdir, filename="profile", schedule=None)
-    model = BoringModel()
+    model = boring_model_cls()
     model.predict_dataloader = model.train_dataloader
-    trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, limit_predict_batches=2, profiler=pytorch_profiler)
-    trainer.predict(model)
-
-    assert sum(e.name == "predict_step" for e in pytorch_profiler.function_events)
-    path = pytorch_profiler.dirpath / f"predict-{pytorch_profiler.filename}.txt"
-    assert path.read_text("utf-8")
-
-
-def test_pytorch_profiler_trainer_validate(tmpdir):
-    """Ensure that the profiler can be given to the trainer and validate function are properly recorded."""
-    pytorch_profiler = PyTorchProfiler(dirpath=tmpdir, filename="profile", schedule=None)
-    model = BoringModel()
-    trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, limit_val_batches=2, profiler=pytorch_profiler)
-    trainer.validate(model)
+    trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, limit_test_batches=2, profiler=pytorch_profiler)
+    getattr(trainer, fn)(model)
 
-    assert sum(e.name == "validation_step" for e in pytorch_profiler.function_events)
+    assert sum(e.name == f"{step_name}_step" for e in pytorch_profiler.function_events)
 
-    path = pytorch_profiler.dirpath / f"validate-{pytorch_profiler.filename}.txt"
+    path = pytorch_profiler.dirpath / f"{fn}-{pytorch_profiler.filename}.txt"
     assert path.read_text("utf-8")
 
+    if _KINETO_AVAILABLE:
+        files = sorted(file for file in os.listdir(tmpdir) if file.endswith(".json"))
+        assert any(f"{fn}-{pytorch_profiler.filename}" in f for f in files)
+        path = pytorch_profiler.dirpath / f"{fn}-{pytorch_profiler.filename}.txt"
+        assert path.read_text("utf-8")
+
 
 def test_pytorch_profiler_nested(tmpdir):
     """Ensure that the profiler handles nested context"""
@@ -467,7 +465,7 @@ def on_fit_end(self, trainer, *args, **kwargs) -> None:
 
     profiler = cls(dirpath=tmpdir, filename="profiler")
     model = BoringModel()
-    trainer = Trainer(default_root_dir=tmpdir, fast_dev_run=True, profiler=profiler, callbacks=[TestCallback()])
+    trainer = Trainer(default_root_dir=tmpdir, fast_dev_run=1, profiler=profiler, callbacks=[TestCallback()])
     trainer.fit(model)
 
     assert profiler._output_file is None