[feat] add random_sparse_mask api (#1066)

* [feat] add random_sparse_mask api

* correct test skip

Co-authored-by: Min Xu <min.xu.public@gmail.com>

fairscale/experimental/wgit/__init__.py

@@ -21,7 +21,7 @@
 
 
 from .repo import Repo
-from .signal_sparsity import Algo, SignalSparsity
+from .signal_sparsity import Algo, SignalSparsity, random_sparse_mask
 from .signal_sparsity_profiling import EnergyConcentrationProfile
 from .version import __version_tuple__
 

fairscale/experimental/wgit/signal_sparsity.py

@@ -426,3 +426,21 @@ def lossy_compress(self, dense: Tensor) -> Tuple[Tensor, Optional[Tensor], Optio
             sst = self.dense_to_sst(dense)
             dst = self.dense_sst_to_dst(dense, sst)
             return self.sst_dst_to_dense(sst, dst), sst, dst
+
+
+def random_sparse_mask(dense: Tensor, percent: float, dim: int) -> Tensor:
+    """Get a random sparse mask
+
+    Args:
+        dense (Tensor):
+            Input dense tensor (no zeros).
+        percent (float):
+            Percent of non-zeros.
+        dim (int):
+            Dimension on which the random sparse mask is computed.
+    """
+    assert percent > 0, percent
+    rand = torch.rand_like(dense)
+    ones = torch.ones_like(dense)
+    k = _get_k_for_topk(percent, None, dense.shape[dim])
+    return _scatter_topk_to_sparse_tensor(rand, ones, k, dim)

tests/experimental/wgit/test_signal_sparsity.py

@@ -7,7 +7,7 @@
 import torch
 
 from fair_dev.testing.testing import objects_are_equal
-from fairscale.experimental.wgit.signal_sparsity import SignalSparsity
+from fairscale.experimental.wgit.signal_sparsity import SignalSparsity, random_sparse_mask
 
 # Our own tolerance
 ATOL = 1e-6
@@ -427,3 +427,19 @@ def test_dst_disabled():
     objects_are_equal(rt, result_rt, raise_exception=True, rtol=RTOL, atol=ATOL)
     objects_are_equal(sst, result_sst, raise_exception=True, rtol=RTOL, atol=ATOL)
     assert dst is None
+
+
+@pytest.mark.parametrize("device", ["cpu", "cuda"])
+def test_random_sparse_mask(device):
+    """Tests random_sparse_mask API."""
+    if device == "cuda" and not torch.cuda.is_available():
+        pytest.skip("no GPU")
+
+    dense = torch.tensor([0.5000, 0.6000, 0.7000, 0.8000, 0.9000]).to(device)
+    mask = random_sparse_mask(dense, 0.2, 0)
+    assert mask.sum() == 1
+    for d in [0, 1]:
+        dense = torch.rand(100, 100).to(device)
+        mask = random_sparse_mask(dense, 0.01, d)
+        assert objects_are_equal(mask.sum(dim=d), torch.ones(100).to(device), raise_exception=True)
+        assert mask.sum() == 100