def test_step(self):
model = Model()
sparsifier = WeightNormSparsifier(sparsity_level=0.5)
sparsifier.prepare(model, config=[{'tensor_fqn': 'linear.weight'}])
for g in sparsifier.groups:
# Before step
module = g['module']
assert (1.0 - module.parametrizations['weight'][0].mask.mean()
) == 0 # checking sparsity level is 0
sparsifier.enable_mask_update = True
sparsifier.step()
self.assertAlmostEqual(
model.linear.parametrizations['weight'][0].mask.mean().item(),
0.5,
places=2)
for g in sparsifier.groups:
# After step
module = g['module']
assert (1.0 - module.parametrizations['weight'][0].mask.mean()
) > 0 # checking sparsity level has increased
# Test if the mask collapses to all zeros if the weights are randomized
iters_before_collapse = 1000
for _ in range(iters_before_collapse):
model.linear.weight.data = torch.randn(model.linear.weight.shape)
sparsifier.step()
for g in sparsifier.groups:
# After step
module = g['module']
assert (1.0 - module.parametrizations['weight'][0].mask.mean()
) > 0 # checking sparsity level did not collapse
def test_step(self):
model = Model()
sparsifier = WeightNormSparsifier(sparsity_level=0.5)
sparsifier.prepare(model, config=[model.linear])
sparsifier.enable_mask_update = True
sparsifier.step()
self.assertAlmostEqual(
model.linear.parametrizations['weight'][0].mask.mean().item(),
0.5,
places=2)
def test_step(self):
model = Model()
sparsifier = WeightNormSparsifier(sparsity_level=0.5)
sparsifier.prepare(model, config=[model.linear])
for g in sparsifier.module_groups:
# Before step
module = g['module']
assert (1.0 - module.parametrizations['weight'][0].mask.mean()) == 0 # checking sparsity level is 0
sparsifier.enable_mask_update = True
sparsifier.step()
self.assertAlmostEqual(model.linear.parametrizations['weight'][0].mask.mean().item(), 0.5, places=2)
for g in sparsifier.module_groups:
# After step
module = g['module']
assert (1.0 - module.parametrizations['weight'][0].mask.mean()) > 0 # checking sparsity level has increased