| # -*- coding: utf-8 -*- |
| # Owner(s): ["module: unknown"] |
| |
| import itertools |
| import logging |
| import re |
| |
| import torch |
| from torch import nn |
| from torch.ao.sparsity import BaseSparsifier, WeightNormSparsifier, FakeSparsity |
| from torch.nn.utils.parametrize import is_parametrized |
| |
| from torch.testing._internal.common_utils import TestCase |
| |
| logging.basicConfig(format='%(asctime)s - %(name)s - %(levelname)s - %(message)s', level=logging.INFO) |
| |
| class Model(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.seq = nn.Sequential( |
| nn.Linear(16, 16) |
| ) |
| self.linear = nn.Linear(16, 16) |
| self.head = nn.Linear(16, 4) |
| |
| def forward(self, x): |
| x = self.seq(x) |
| x = self.linear(x) |
| x = self.head(x) |
| return x |
| |
| |
| class ImplementedSparsifier(BaseSparsifier): |
| def __init__(self, **kwargs): |
| super().__init__(defaults=kwargs) |
| |
| def update_mask(self, layer, **kwargs): |
| layer.parametrizations.weight[0].mask[0] = 0 |
| linear_state = self.state['linear'] |
| linear_state['step_count'] = linear_state.get('step_count', 0) + 1 |
| |
| |
| class TestBaseSparsifier(TestCase): |
| def test_constructor(self): |
| # Cannot instantiate the abstract base |
| self.assertRaises(TypeError, BaseSparsifier) |
| # Can instantiate the model with no configs |
| model = Model() |
| sparsifier = ImplementedSparsifier(test=3) |
| sparsifier.prepare(model, config=None) |
| assert len(sparsifier.module_groups) == 3 |
| sparsifier.step() |
| # Can instantiate the model with configs |
| sparsifier = ImplementedSparsifier(test=3) |
| sparsifier.prepare(model, [model.linear]) |
| assert len(sparsifier.module_groups) == 1 |
| assert sparsifier.module_groups[0]['fqn'] == 'linear' |
| assert 'test' in sparsifier.module_groups[0] |
| assert sparsifier.module_groups[0]['test'] == 3 |
| |
| def test_prepare_config(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(test=3) |
| # Make sure there are no parametrizations before `prepare` |
| assert not hasattr(model.seq[0], 'parametrizations') |
| assert not hasattr(model.linear, 'parametrizations') |
| assert not hasattr(model.head, 'parametrizations') |
| sparsifier.prepare(model, config=[ |
| {'fqn': 'seq.0', 'test': 42}, |
| # No 'linear' to make sure it will be skipped in the sparsification |
| {'module': model.head, 'fqn': 'head'} |
| ]) |
| assert len(sparsifier.module_groups) == 2 |
| # Check if default argument is not assigned if explicit |
| assert sparsifier.module_groups[0]['fqn'] == 'seq.0' |
| assert sparsifier.module_groups[0]['test'] == 42 |
| # Check if FQN and module are pointing to the same location |
| assert sparsifier.module_groups[1]['fqn'] == 'head' |
| assert sparsifier.module_groups[1]['module'] == model.head |
| # Check if parameterizations are attached |
| assert hasattr(model.seq[0], 'parametrizations') |
| assert not hasattr(model.linear, 'parametrizations') |
| assert hasattr(model.head, 'parametrizations') |
| |
| def test_step(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(test=3) |
| sparsifier.enable_mask_update = True |
| sparsifier.prepare(model, [model.linear]) |
| sparsifier.step() |
| assert torch.all(model.linear.parametrizations.weight[0].mask[0] == 0) |
| |
| def test_state_dict(self): |
| step_count = 3 |
| model0 = Model() |
| sparsifier0 = ImplementedSparsifier(test=3) |
| sparsifier0.prepare(model0, [model0.linear]) |
| mask = model0.linear.parametrizations['weight'][0].mask |
| mask.data = torch.arange(mask.shape[0] * mask.shape[1]).reshape(mask.shape) |
| for step in range(step_count): |
| sparsifier0.step() |
| state_dict = sparsifier0.state_dict() |
| |
| # Check the expected keys in the state_dict |
| assert 'state' in state_dict |
| assert 'linear' in state_dict['state'] |
| assert 'step_count' in state_dict['state']['linear'] |
| assert state_dict['state']['linear']['step_count'] == 3 |
| |
| assert 'module_groups' in state_dict |
| assert 'test' in state_dict['module_groups'][0] |
| assert 'fqn' in state_dict['module_groups'][0] |
| assert state_dict['module_groups'][0]['fqn'] == 'linear' |
| |
| # Check loading static_dict creates an equivalent model |
| model1 = Model() |
| sparsifier1 = ImplementedSparsifier() |
| sparsifier1.prepare(model1, None) |
| |
| assert sparsifier0.state != sparsifier1.state |
| |
| # Make sure the masks are different in the beginning |
| for mg in sparsifier0.module_groups: |
| if mg['fqn'] == 'linear': |
| mask0 = mg['module'].parametrizations.weight[0].mask |
| for mg in sparsifier1.module_groups: |
| if mg['fqn'] == 'linear': |
| mask1 = mg['module'].parametrizations.weight[0].mask |
| self.assertNotEqual(mask0, mask1) |
| |
| sparsifier1.load_state_dict(state_dict) |
| |
| # Make sure the states are loaded, and are correct |
| assert sparsifier0.state == sparsifier1.state |
| |
| # Make sure the masks (and all dicts) are the same after loading |
| assert len(sparsifier0.module_groups) == len(sparsifier1.module_groups) |
| for idx in range(len(sparsifier0.module_groups)): |
| mg0 = sparsifier0.module_groups[idx] |
| mg1 = sparsifier1.module_groups[idx] |
| for key in mg0.keys(): |
| assert key in mg1 |
| if key == 'module': |
| # We cannot compare modules as they are different |
| param0 = mg0[key].parametrizations.weight[0] |
| param1 = mg1[key].parametrizations.weight[0] |
| assert hasattr(param0, 'mask') |
| assert hasattr(param1, 'mask') |
| self.assertEqual(param0.__dict__, param1.__dict__) |
| else: |
| assert mg0[key] == mg1[key] |
| |
| def test_mask_squash(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(test=3) |
| sparsifier.prepare(model, [model.linear]) |
| assert hasattr(model.linear.parametrizations.weight[0], 'mask') |
| assert is_parametrized(model.linear, 'weight') |
| assert not is_parametrized(model.seq[0], 'weight') |
| |
| sparsifier.squash_mask() |
| assert not is_parametrized(model.seq[0], 'weight') |
| assert not is_parametrized(model.linear, 'weight') |
| |
| def test_mask_squash_with_params1(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(foo=3, bar=2, baz=1) |
| sparsifier.prepare(model, [model.linear, model.seq[0]]) |
| sparsifier.squash_mask( |
| params_to_keep_per_layer={ |
| 'linear': ('foo', 'bar'), |
| 'seq.0': ('baz',) |
| }) |
| assert not is_parametrized(model.seq[0], 'weight') |
| assert not is_parametrized(model.linear, 'weight') |
| assert hasattr(model.seq[0], 'sparse_params') |
| assert hasattr(model.linear, 'sparse_params') |
| assert model.seq[0].sparse_params.get('foo', None) is None |
| assert model.seq[0].sparse_params.get('bar', None) is None |
| assert model.seq[0].sparse_params.get('baz', None) == 1 |
| assert model.linear.sparse_params.get('foo', None) == 3 |
| assert model.linear.sparse_params.get('bar', None) == 2 |
| assert model.linear.sparse_params.get('baz', None) is None |
| |
| def test_mask_squash_with_params2(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(foo=3, bar=2, baz=1) |
| sparsifier.prepare(model, [model.linear, model.seq[0]]) |
| sparsifier.squash_mask(params_to_keep=('foo', 'bar')) |
| assert not is_parametrized(model.seq[0], 'weight') |
| assert not is_parametrized(model.linear, 'weight') |
| assert hasattr(model.seq[0], 'sparse_params') |
| assert hasattr(model.linear, 'sparse_params') |
| assert model.seq[0].sparse_params.get('foo', None) == 3 |
| assert model.seq[0].sparse_params.get('bar', None) == 2 |
| assert model.seq[0].sparse_params.get('baz', None) is None |
| assert model.linear.sparse_params.get('foo', None) == 3 |
| assert model.linear.sparse_params.get('bar', None) == 2 |
| assert model.linear.sparse_params.get('baz', None) is None |
| |
| def test_mask_squash_with_params3(self): |
| model = Model() |
| sparsifier = ImplementedSparsifier(foo=3, bar=2, baz=1) |
| sparsifier.prepare(model, [model.linear, model.seq[0]]) |
| sparsifier.squash_mask( |
| params_to_keep=('foo', 'bar'), |
| params_to_keep_per_layer={'seq.0': ('baz',)}) |
| assert not is_parametrized(model.seq[0], 'weight') |
| assert not is_parametrized(model.linear, 'weight') |
| assert hasattr(model.seq[0], 'sparse_params') |
| assert hasattr(model.linear, 'sparse_params') |
| assert model.seq[0].sparse_params.get('foo', None) == 3 |
| assert model.seq[0].sparse_params.get('bar', None) == 2 |
| assert model.seq[0].sparse_params.get('baz', None) == 1 |
| assert model.linear.sparse_params.get('foo', None) == 3 |
| assert model.linear.sparse_params.get('bar', None) == 2 |
| assert model.linear.sparse_params.get('baz', None) is None |
| |
| |
| class TestWeightNormSparsifier(TestCase): |
| def test_constructor(self): |
| model = Model() |
| sparsifier = WeightNormSparsifier() |
| sparsifier.prepare(model, config=None) |
| for g in sparsifier.module_groups: |
| assert isinstance(g['module'], nn.Linear) |
| # The module_groups are unordered |
| assert g['fqn'] in ('seq.0', 'linear', 'head') |
| |
| 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 |
| # 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.module_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_2_of_4(self): |
| model = Model() |
| sparsifier = WeightNormSparsifier(sparsity_level=1.0, |
| sparse_block_shape=(1, 4), |
| zeros_per_block=2) |
| sparsifier.prepare(model, config=[model.linear]) |
| sparsifier.step() |
| # make sure the sparsity level is approximately 50% |
| self.assertAlmostEqual(model.linear.parametrizations['weight'][0].mask.mean().item(), 0.5, places=2) |
| # Make sure each block has exactly 50% zeros |
| module = sparsifier.module_groups[0]['module'] |
| mask = module.parametrizations['weight'][0].mask |
| for row in mask: |
| for idx in range(0, len(row), 4): |
| block = row[idx:idx + 4] |
| block, _ = block.sort() |
| assert (block[:2] == 0).all() |
| assert (block[2:] != 0).all() |
| |
| def test_prepare(self): |
| model = Model() |
| sparsifier = WeightNormSparsifier() |
| sparsifier.prepare(model, config=None) |
| for g in sparsifier.module_groups: |
| module = g['module'] |
| # Check mask exists |
| assert hasattr(module.parametrizations['weight'][0], 'mask') |
| # Check parametrization exists and is correct |
| assert is_parametrized(module, 'weight') |
| assert type(module.parametrizations.weight[0]) == FakeSparsity |
| |
| def test_mask_squash(self): |
| model = Model() |
| sparsifier = WeightNormSparsifier() |
| sparsifier.prepare(model, config=None) |
| sparsifier.squash_mask() |
| for g in sparsifier.module_groups: |
| module = g['module'] |
| assert not is_parametrized(module, 'weight') |
| assert not hasattr(module, 'mask') |
| |
| def test_sparsity_levels(self): |
| sparsity_levels = [-1.0, 0.0, 0.5, 1.0, 2.0] |
| sparse_block_shapes = [(1, 1), (1, 4), (2, 2), (4, 1)] |
| zeros_per_blocks = [0, 1, 2, 3, 4] |
| |
| testcases = itertools.tee(itertools.product(sparsity_levels, |
| sparse_block_shapes, |
| zeros_per_blocks)) |
| # Create a config and model with all the testcases |
| model = nn.Sequential() |
| sparsifier = WeightNormSparsifier() |
| |
| sparsity_per_layer_config = [] |
| p = re.compile(r'[-\.\s]') |
| for sl, sbs, zpb in testcases[0]: |
| # Make sure the number of zeros is not > values in a block |
| if zpb > sbs[0] * sbs[1]: |
| continue |
| layer_name = f'{sl}_{sbs}_{zpb}' |
| layer_name = p.sub('_', layer_name) |
| |
| layer = nn.Linear(12, 12, bias=False) |
| layer.weight = nn.Parameter(torch.ones(12, 12)) |
| model.add_module(layer_name, layer) |
| config = { |
| 'fqn': layer_name, |
| 'sparsity_level': sl, |
| 'sparse_block_shape': sbs, |
| 'zeros_per_block': zpb |
| } |
| sparsity_per_layer_config.append(config) |
| |
| sparsifier.prepare(model, sparsity_per_layer_config) |
| sparsifier.step() |
| sparsifier.squash_mask() |
| model.eval() |
| |
| for sl, sbs, zpb in testcases[1]: |
| if zpb > sbs[0] * sbs[1]: |
| continue |
| layer_name = f'{sl}_{sbs}_{zpb}' |
| layer_name = p.sub('_', layer_name) |
| layer = getattr(model, layer_name) |
| |
| # Level of sparsity is achieved |
| sparse_mask = (layer.weight == 0).float() |
| if zpb == 0: |
| assert sparse_mask.mean() == 0 |
| else: |
| # Ratio of individual zeros in the tensor |
| true_sl = min(max(sl, 0.0), 1.0) |
| true_sl = true_sl * zpb / sbs[0] / sbs[1] |
| assert sparse_mask.mean() == true_sl |
