Call on_before_zero_grad model hook

CHANGELOG.md

@@ -36,6 +36,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed loggers - flushing last logged metrics even before continue, e.g. `trainer.test()` results ([#1459](https://github.com/PyTorchLightning/pytorch-lightning/pull/1459))
 
+- Added a missing call to the `on_before_zero_grad` model hook ([#1493](https://github.com/PyTorchLightning/pytorch-lightning/pull/1493)). 
+
 - 
 
 

pytorch_lightning/core/lightning.py

@@ -1157,6 +1157,10 @@ def optimizer_step(self, current_epoch, batch_idx, optimizer,
                     optimizer.step()
                     optimizer.zero_grad()
 
+        Note:
+            If you also override the :meth:`~pytorch_lightning.core.hooks.ModelHooks.on_before_zero_grad`
+            model hook don't forget to add the call to it before ``optimizer.zero_grad()`` yourself.
+
         """
         if self.trainer.use_tpu and XLA_AVAILABLE:
             xm.optimizer_step(optimizer)
@@ -1165,6 +1169,9 @@ def optimizer_step(self, current_epoch, batch_idx, optimizer,
         else:
             optimizer.step()
 
+        # model hook
+        self.on_before_zero_grad(optimizer)
+
         # clear gradients
         optimizer.zero_grad()
 

tests/models/test_hooks.py

@@ -0,0 +1,39 @@
+import pytest
+
+import tests.base.utils as tutils
+from pytorch_lightning import Trainer
+from tests.base import (
+    LightTrainDataloader,
+    LightValidationMixin,
+    TestModelBase,
+    LightTestMixin)
+
+
+@pytest.mark.parametrize('max_steps', [1, 2, 3])
+def test_on_before_zero_grad_called(max_steps):
+
+    class CurrentTestModel(
+        LightTrainDataloader,
+        LightValidationMixin,
+        LightTestMixin,
+        TestModelBase,
+    ):
+        on_before_zero_grad_called = 0
+
+        def on_before_zero_grad(self, optimizer):
+            self.on_before_zero_grad_called += 1
+
+    hparams = tutils.get_default_hparams()
+    model = CurrentTestModel(hparams)
+
+    trainer = Trainer(
+        max_steps=max_steps,
+        num_sanity_val_steps=5,
+    )
+    assert 0 == model.on_before_zero_grad_called
+    trainer.fit(model)
+    assert max_steps == model.on_before_zero_grad_called
+
+    model.on_before_zero_grad_called = 0
+    trainer.test(model)
+    assert 0 == model.on_before_zero_grad_called