Use torch.nn.utils.clip_grad_norm_ and add clip_grad_by_value support for TPU

CHANGELOG.md

@@ -105,6 +105,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added new `EarlyStopping` parameters `stopping_threshold` and `divergence_threshold` ([#6868](https://github.com/PyTorchLightning/pytorch-lightning/pull/6868))
 
 
+- Added `clip_grad_by_value` support for TPUs ([#7025](https://github.com/PyTorchLightning/pytorch-lightning/pull/7025))
+
 
 ### Changed
 
@@ -135,6 +137,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - `pl.seed_everyting` will now also set the seed on the `DistributedSampler` ([#7024](https://github.com/PyTorchLightning/pytorch-lightning/pull/7024))
 
 
+- Changed `clip_grad_norm` to use `torch.nn.utils.clip_grad_norm_` ([#7025](https://github.com/PyTorchLightning/pytorch-lightning/pull/7025))
+
+
 ### Deprecated
 
 - Deprecated `TrainerTrainingTricksMixin` in favor of a separate utilities module for NaN/Inf detection for gradients and parameters ([#6834](https://github.com/PyTorchLightning/pytorch-lightning/pull/6834/))

docs/source/advanced/training_tricks.rst

@@ -26,10 +26,14 @@ The effect is a large effective batch size of size KxN.
 
 Gradient Clipping
 -----------------
-Gradient clipping may be enabled to avoid exploding gradients. By default, this will `clip the gradient norm
-<https://pytorch.org/docs/stable/nn.html#torch.nn.utils.clip_grad_norm_>`_ computed over all model parameters together.
-If ``gradient_clip_algorithm`` option is set to ``value``, which is ``norm`` by default, this will
-`clip the gradient value <https://pytorch.org/docs/stable/nn.html#torch.nn.utils.clip_grad_value_>`_ for each parameter instead.
+Gradient clipping may be enabled to avoid exploding gradients. By default, this will clip the gradient norm by calling
+:func:`torch.nn.utils.clip_grad_norm_` computed over all model parameters together.
+If the Trainer's ``gradient_clip_algorithm`` is set to ``'value'`` (``'norm'`` by default), this will use instead
+:func:`torch.nn.utils.clip_grad_norm_` for each parameter instead.
+
+.. note::
+    If using mixed precision, the ``gradient_clip_val`` does not need to be changed as the gradients are unscaled
+    before applying the clipping function.
 
 .. seealso:: :class:`~pytorch_lightning.trainer.trainer.Trainer`
 
@@ -38,11 +42,10 @@ If ``gradient_clip_algorithm`` option is set to ``value``, which is ``norm`` by
     # DEFAULT (ie: don't clip)
     trainer = Trainer(gradient_clip_val=0)
 
-    # clip gradients with norm above 0.5
-    trainer = Trainer(gradient_clip_val=0.5)
+    # clip gradients' global norm to <=0.5
+    trainer = Trainer(gradient_clip_val=0.5)  # gradient_clip_algorithm='norm' by default
 
-    # clip gradients with value above 0.5
-    # gradient_clip_algorithm types => :class:`~pytorch_lightning.utilities.enums.GradClipAlgorithmType`
+    # clip gradients' maximum magnitude to <=0.5
     trainer = Trainer(gradient_clip_val=0.5, gradient_clip_algorithm='value')
 
 ----------

pytorch_lightning/accelerators/tpu.py

@@ -11,7 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Any, Callable, Union
+from typing import Any, Callable
 
 from torch.optim import Optimizer
 
@@ -20,15 +20,16 @@
 from pytorch_lightning.plugins.precision import MixedPrecisionPlugin
 from pytorch_lightning.plugins.training_type.single_tpu import SingleTPUPlugin
 from pytorch_lightning.plugins.training_type.tpu_spawn import TPUSpawnPlugin
-from pytorch_lightning.utilities import _XLA_AVAILABLE, GradClipAlgorithmType
+from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_5, _XLA_AVAILABLE
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 
 if _XLA_AVAILABLE:
     import torch_xla.core.xla_model as xm
-    from torch_xla._patched_functions import clip_grad_norm_
 
-    # rename to mock in a test
-    xla_clip_grad_norm_ = clip_grad_norm_
+    # clip_grad_norm_ was updated to not require this patch in 1.5.0
+    if _TORCH_GREATER_EQUAL_1_5:
+        from torch_xla._patched_functions import _apply_patches
+        _apply_patches()  # patches torch.nn.utils.clip_grad_norm_
 
 
 class TPUAccelerator(Accelerator):
@@ -42,8 +43,7 @@ def setup(self, trainer: 'pl.Trainer', model: 'pl.LightningModule') -> None:
         """
         if isinstance(self.precision_plugin, MixedPrecisionPlugin):
             raise MisconfigurationException(
-                "amp + tpu is not supported. "
-                "Only bfloats are supported on TPU. Consider using TPUHalfPrecisionPlugin"
+                "amp + tpu is not supported. Only bfloats are supported on TPU. Consider using TPUHalfPrecisionPlugin"
             )
 
         if not isinstance(self.training_type_plugin, (SingleTPUPlugin, TPUSpawnPlugin)):
@@ -54,21 +54,3 @@ def run_optimizer_step(
         self, optimizer: Optimizer, optimizer_idx: int, lambda_closure: Callable, **kwargs: Any
     ) -> None:
         xm.optimizer_step(optimizer, barrier=False, optimizer_args={'closure': lambda_closure, **kwargs})
-
-    def clip_gradients(
-        self,
-        optimizer: Optimizer,
-        clip_val: Union[float, int],
-        gradient_clip_algorithm: GradClipAlgorithmType = GradClipAlgorithmType.NORM,
-    ) -> None:
-        assert gradient_clip_algorithm == GradClipAlgorithmType.NORM, \
-            "Only NORM gradient clipping is supported on TPU for now"
-
-        grad_clip_val = float(clip_val)
-        if grad_clip_val <= 0:
-            return
-
-        parameters = self.model.parameters()
-        norm_type = 2.0
-
-        xla_clip_grad_norm_(parameters, grad_clip_val, norm_type)

pytorch_lightning/plugins/precision/mixed.py

@@ -22,6 +22,5 @@
 class MixedPrecisionPlugin(PrecisionPlugin):
     """Base Class for mixed precision"""
 
-    EPSILON: float = 1e-5
     backend: 'AMPType'
     precision: Union[str, int] = "mixed"

pytorch_lightning/plugins/precision/precision_plugin.py

@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import math
 from typing import Any, Callable, List, Tuple, Union
 
 import torch
@@ -28,10 +27,9 @@
 class PrecisionPlugin(Plugin):
     """
     Base class for all plugins handling the precision-specific parts of the training.
-    The static classattributes EPSILON and precision must be overwritten in child-classes and their
-    default values reflect fp32 training.
+    The class attribute precision must be overwritten in child classes.
+    The default value reflects fp32 training.
     """
-    EPSILON: float = 1e-6
     precision: Union[str, int] = 32
 
     def master_params(self, optimizer: Optimizer) -> _PARAMETERS:
@@ -117,32 +115,14 @@ def clip_gradients(
             self.clip_grad_by_value(optimizer, clip_val)
         elif gradient_clip_algorithm == GradClipAlgorithmType.NORM:
             # TODO: there should be a mechanism to set `norm_type`
-            self.clip_grad_by_norm(optimizer, clip_val, eps=self.EPSILON)
+            self.clip_grad_by_norm(optimizer, clip_val)
 
     def clip_grad_by_value(self, optimizer: Optimizer, clip_val: Union[int, float]) -> None:
         """Clip gradients by value"""
         parameters = self.master_params(optimizer)
         torch.nn.utils.clip_grad_value_(parameters, clip_value=clip_val)
 
-    def clip_grad_by_norm(
-        self, optimizer: Optimizer, clip_val: Union[int, float], norm_type: float = 2.0, eps: float = 1e-6
-    ) -> None:
+    def clip_grad_by_norm(self, optimizer: Optimizer, clip_val: Union[int, float]) -> None:
         """Clip gradients by norm"""
         parameters = self.master_params(optimizer)
-
-        # TODO: replace this with torch.nn.clip_grad_norm_
-        parameters = list(filter(lambda p: p.grad is not None, parameters))
-        device = parameters[0].device
-
-        if norm_type == math.inf:
-            total_norm = max(p.grad.data.abs().max() for p in parameters)
-        else:
-            out = torch.empty(len(parameters), device=device)
-            for i, p in enumerate(parameters):
-                torch.norm(p.grad.data.to(device), norm_type, out=out[i])
-            total_norm = torch.norm(out, norm_type)
-
-        clip_coef = torch.tensor(clip_val, device=device) / (total_norm + eps)
-        clip_coef = torch.min(clip_coef, torch.ones_like(clip_coef))
-        for p in parameters:
-            p.grad.data.mul_(clip_coef.to(p.grad.data.device))
+        torch.nn.utils.clip_grad_norm_(parameters, clip_val)

pytorch_lightning/trainer/training_tricks.py

@@ -22,8 +22,6 @@
 from pytorch_lightning.utilities import rank_zero_deprecation
 from pytorch_lightning.utilities.finite_checks import detect_nan_parameters, print_nan_gradients
 
-EPSILON = 1e-6
-EPSILON_FP16 = 1e-5
 log = logging.getLogger(__name__)
 
 

pytorch_lightning/utilities/__init__.py

@@ -45,6 +45,7 @@
     _NATIVE_AMP_AVAILABLE,
     _OMEGACONF_AVAILABLE,
     _RPC_AVAILABLE,
+    _TORCH_GREATER_EQUAL_1_5,
     _TORCH_GREATER_EQUAL_1_6,
     _TORCH_GREATER_EQUAL_1_7,
     _TORCH_GREATER_EQUAL_1_8,

pytorch_lightning/utilities/imports.py

@@ -66,6 +66,7 @@ def _compare_version(package: str, op, version) -> bool:
 _IS_WINDOWS = platform.system() == "Windows"
 _IS_INTERACTIVE = hasattr(sys, "ps1")  # https://stackoverflow.com/a/64523765
 _TORCH_LOWER_EQUAL_1_4 = _compare_version("torch", operator.le, "1.5.0")
+_TORCH_GREATER_EQUAL_1_5 = _compare_version("torch", operator.ge, "1.5.0")
 _TORCH_GREATER_EQUAL_1_6 = _compare_version("torch", operator.ge, "1.6.0")
 _TORCH_GREATER_EQUAL_1_7 = _compare_version("torch", operator.ge, "1.7.0")
 _TORCH_GREATER_EQUAL_1_8 = _compare_version("torch", operator.ge, "1.8.0")

tests/models/test_tpu.py

@@ -224,7 +224,7 @@ def test_tpu_grad_norm(tmpdir):
 @RunIf(tpu=True)
 @pl_multi_process_test
 def test_tpu_clip_grad_by_value(tmpdir):
-    """Test if clip_gradients by value works on TPU. (It should not.)"""
+    """Test if clip_gradients by value works on TPU"""
     tutils.reset_seed()
     trainer_options = dict(
         default_root_dir=tmpdir,
@@ -238,8 +238,7 @@ def test_tpu_clip_grad_by_value(tmpdir):
     )
 
     model = BoringModel()
-    with pytest.raises(AssertionError):
-        tpipes.run_model_test(trainer_options, model, on_gpu=False, with_hpc=False)
+    tpipes.run_model_test(trainer_options, model, on_gpu=False, with_hpc=False)
 
 
 @RunIf(tpu=True)
@@ -383,7 +382,7 @@ def test_reduce(rank):
 @RunIf(tpu=True)
 @pl_multi_process_test
 @pytest.mark.parametrize("clip_val", [10])
-@mock.patch("pytorch_lightning.accelerators.tpu.xla_clip_grad_norm_")
+@mock.patch("torch.nn.utils.clip_grad_norm_")
 def test_tpu_precision_16_clip_gradients(mock_clip_grad_norm, clip_val, tmpdir):
     """
     Ensure that clip gradients is only called if the value is greater than 0.