Add AMP for validation, prediction and testing (#6565)

* Add Tests for val and test-steps

* Add native AMP

* pep8 tests

* pep8 plugin

* changelog

(cherry picked from commit 634d831)

CHANGELOG.md

@@ -92,6 +92,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Added Autocast in validation, test and predict modes for Native AMP ([#6565](https://github.com/PyTorchLightning/pytorch-lightning/pull/6565))
+
 - Made the `Plugin.reduce` method more consistent across all Plugins to reflect a mean-reduction by default ([#6011](https://github.com/PyTorchLightning/pytorch-lightning/pull/6011))
 
 

pytorch_lightning/plugins/precision/native_amp.py

@@ -93,3 +93,21 @@ def train_step_context(self) -> Generator[autocast, None, None]:
         """Enable autocast context"""
         with torch.cuda.amp.autocast():
             yield
+
+    @contextmanager
+    def val_step_context(self) -> Generator[None, None, None]:
+        """Enable autocast context"""
+        with torch.cuda.amp.autocast():
+            yield
+
+    @contextmanager
+    def test_step_context(self) -> Generator[None, None, None]:
+        """Enable autocast context"""
+        with torch.cuda.amp.autocast():
+            yield
+
+    @contextmanager
+    def predict_context(self) -> Generator[None, None, None]:
+        """Enable autocast context"""
+        with torch.cuda.amp.autocast():
+            yield

tests/models/test_amp.py

@@ -17,24 +17,43 @@
 import pytest
 import torch
 from torch import optim
+from torch.utils.data import DataLoader
 
 import tests.helpers.utils as tutils
 from pytorch_lightning import Trainer
 from pytorch_lightning.plugins.environments import SLURMEnvironment
 from pytorch_lightning.trainer.states import TrainerState
 from pytorch_lightning.utilities import _APEX_AVAILABLE
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
-from tests.helpers import BoringModel
+from tests.helpers import BoringModel, RandomDataset
 
 
 class AMPTestModel(BoringModel):
 
-    def training_step(self, batch, batch_idx):
+    def _step(self, batch, batch_idx):
         assert torch.is_autocast_enabled()
         output = self(batch)
         assert output.dtype == torch.float16
         loss = self.loss(batch, output)
-        return {"loss": loss}
+        return loss
+
+    def training_step(self, batch, batch_idx):
+        output = self._step(batch, batch_idx)
+        return {"loss": output}
+
+    def validation_step(self, batch, batch_idx):
+        output = self._step(batch, batch_idx)
+        return {"x": output}
+
+    def test_step(self, batch, batch_idx):
+        output = self._step(batch, batch_idx)
+        return {"y": output}
+
+    def predict(self, batch, batch_idx, dataloader_idx=None):
+        assert torch.is_autocast_enabled()
+        output = self(batch)
+        assert output.dtype == torch.float16
+        return output
 
 
 @pytest.mark.skip(reason='dp + amp not supported currently')  # TODO
@@ -54,6 +73,8 @@ def test_amp_single_gpu_dp(tmpdir):
     model = AMPTestModel()
     # tutils.run_model_test(trainer_options, model)
     trainer.fit(model)
+    trainer.test(model)
+    trainer.predict(model, DataLoader(RandomDataset(32, 64)))
 
     assert trainer.state == TrainerState.FINISHED, f"Training failed with {trainer.state}"
 
@@ -73,6 +94,8 @@ def test_amp_single_gpu_ddp_spawn(tmpdir):
     model = AMPTestModel()
     # tutils.run_model_test(trainer_options, model)
     trainer.fit(model)
+    trainer.test(model)
+    trainer.predict(model, DataLoader(RandomDataset(32, 64)))
     assert trainer.state == TrainerState.FINISHED, f"Training failed with {trainer.state}"
 
 
@@ -112,6 +135,8 @@ def test_amp_multi_gpu_ddp_spawn(tmpdir):
     model = AMPTestModel()
     # tutils.run_model_test(trainer_options, model)
     trainer.fit(model)
+    trainer.test(model)
+    trainer.predict(model, DataLoader(RandomDataset(32, 64)))
     assert trainer.state == TrainerState.FINISHED, f"Training failed with {trainer.state}"