def _rescale_gradients(self) -> Optional[float]:
"""
Performs gradient rescaling. Is a no-op if gradient rescaling is not enabled.
"""
if self._grad_norm:
parameters_to_clip = [p for p in self._model.parameters() if p.grad is not None]
return sparse_clip_norm(parameters_to_clip, self._grad_norm)
return None
def rescale_gradients(self) -> Optional[float]:
"""
Performs gradient rescaling. Is a no-op if gradient rescaling is not enabled.
"""
if self._grad_norm:
if self._opt_level is not None:
# See: https://nvidia.github.io/apex/advanced.html#gradient-clipping
parameters_to_clip = [
p for p in amp.master_params(self.optimizer) if p.grad is not None
]
else:
parameters_to_clip = [p for p in self.model.parameters() if p.grad is not None]
return training_util.sparse_clip_norm(parameters_to_clip, self._grad_norm)
else:
return None
def test_sparse_clip_grad(self):
# create a sparse embedding layer, then take gradient
embedding = torch.nn.Embedding(100, 16, sparse=True)
embedding.zero_grad()
ids = (torch.rand(17) * 100).long()
# Set some of the ids to the same value so that the sparse gradient
# has repeated indices. This tests some additional logic.
ids[:5] = 5
loss = embedding(ids).sum()
loss.backward()
assert embedding.weight.grad.is_sparse
# Now try to clip the gradients.
_ = sparse_clip_norm([embedding.weight], 1.5)
# Final norm should be 1.5
grad = embedding.weight.grad.coalesce()
self.assertAlmostEqual(grad._values().norm(2.0).item(), 1.5, places=5)
