[feat] signal sparsity profiling class (#1060)

* added a profiling class

* no more type ignore after merging main

* fixed a int/round bug

* add unit tests

* skip if no cuda for a test

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

fairscale/experimental/wgit/__init__.py

@@ -22,6 +22,7 @@
 
 from .repo import Repo
 from .signal_sparsity import Algo, SignalSparsity
+from .signal_sparsity_profiling import EnergyConcentrationProfile
 from .version import __version_tuple__
 
 __version__ = ".".join([str(x) for x in __version_tuple__])

fairscale/experimental/wgit/signal_sparsity_profiling.py

@@ -0,0 +1,62 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+#
+# This source code is licensed under the BSD license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List
+
+import torch
+from torch import Tensor
+
+
+class EnergyConcentrationProfile:
+    """Compute "energy" concentration level for a tensor
+
+    Args:
+        dim (int):
+            The dimension to measure.
+        top_k_percents (List[float]):
+            List of percentage values. For each value, the `measure`
+            function will compute and return the percentage of "energy"
+            concentrated on that top-K percent of values in the dimension
+            to measure. Note, this is the opposite of the sparsity percentage.
+    """
+
+    def __init__(self, dim: int, top_k_percents: List[float]) -> None:
+        assert isinstance(dim, int)
+        self.dim = dim
+        self.percents = []
+        last_p = 0.0
+        for p in top_k_percents:
+            assert isinstance(p, (int, float))
+            assert p > 0, p
+            assert p <= 100, p
+            assert p > last_p, f"p {p} should be larger than last_p {last_p}"
+            self.percents.append(float(p))
+            last_p = p
+
+    def measure(self, in_tensor: Tensor) -> List[Tensor]:
+        """Compute the return the results
+
+            Note, we want this function to be nonblocking and async.
+
+        Returns:
+            (List[Tensor])
+                List of tensors. Each tensor is a singleton float
+                that contains the energy measure for that top_k_percent.
+        """
+        assert in_tensor.is_floating_point(), in_tensor.dtype
+        assert self.dim < len(in_tensor.shape), f"tensor shape {in_tensor.shape} not compatible with dim {self.dim}"
+        dim_size = in_tensor.shape[self.dim]
+        abs_tensor = in_tensor.abs()
+        full_energy = abs_tensor.sum()
+        return_tensors = []
+        for p in self.percents:
+            k = max(1, round(p / 100 * dim_size))
+            abs_top_k_values, _ = abs_tensor.topk(k, dim=self.dim)
+            return_tensors.append(abs_top_k_values.sum() / full_energy)
+        return return_tensors
+
+    def measure_fft(self, in_tensor: Tensor) -> List[Tensor]:
+        """Like measure, but do it in FFT frequency domain."""
+        return self.measure(torch.fft.fft(in_tensor, dim=self.dim).real)

tests/ci_test_list_3.txt

@@ -28,3 +28,4 @@ tests/experimental/wgit/test_api.py
 tests/experimental/wgit/test_pygit.py
 tests/experimental/wgit/test_sha1_store.py
 tests/experimental/wgit/test_signal_sparsity.py
+tests/experimental/wgit/test_signal_sparsity_profiling.py

tests/experimental/wgit/test_signal_sparsity_profiling.py

@@ -0,0 +1,74 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the BSD license found in the
+# LICENSE file in the root directory of this source tree.
+
+import time
+
+import pytest
+import torch
+
+from fair_dev.testing.testing import objects_are_equal, skip_if_no_cuda
+from fairscale.experimental.wgit.signal_sparsity_profiling import EnergyConcentrationProfile as ECP
+
+# Our own tolerance
+ATOL = 1e-6
+RTOL = 1e-5
+
+# enable this for debugging.
+# torch.set_printoptions(precision=20)
+
+
+@skip_if_no_cuda
+def test_nonblocking():
+    """Tests cpu runs ahead of the GPU in the measuring process."""
+    big = torch.rand(10, 1000, 1000).cuda()
+    ecp = ECP(dim=2, top_k_percents=[1, 5, 10, 50, 90])
+    start = time.time()
+    out = ecp.measure(big)
+    out_fft = ecp.measure_fft(big)
+    cpu_time = time.time() - start
+    torch.cuda.synchronize()
+    gpu_time = time.time() - start
+    assert cpu_time * 5 < gpu_time, f"GPU time should dominate {cpu_time} vs. {gpu_time}"
+    for o in [out, out_fft]:
+        # validate the output
+        p = [x.item() for x in o]
+        for n, n1 in zip(p, p[1:]):
+            assert n <= n1 and n >= 0 and n <= 100, f"n={n} n1={n1}"
+
+
+def get_ones():
+    """Return test data with ones tensor"""
+    return (
+        0,
+        [1, 5, 10, 100],
+        torch.ones(100),
+        [torch.tensor(0.01), torch.tensor(0.05), torch.tensor(0.1), torch.tensor(1.0)],
+    )
+
+
+def get_dim_0():
+    """Test case for dim=0 for 2D input."""
+    return (
+        0,
+        [1, 3, 33, 66, 90],
+        torch.tensor([0.1, 0.2, 0.1, 0.45]).repeat(100, 1),
+        [torch.tensor(0.01), torch.tensor(0.03), torch.tensor(0.33), torch.tensor(0.66), torch.tensor(0.9)],
+    )
+
+
+@pytest.mark.parametrize(
+    "dim, percents, in_tensor, out_tensors",
+    [
+        get_ones(),
+        get_dim_0(),
+    ],
+)
+def test_expected_output(dim, percents, in_tensor, out_tensors):
+    """Test with a few expected input & outputs."""
+    ecp = ECP(dim, percents)
+    out = ecp.measure(in_tensor)
+    objects_are_equal(out, out_tensors, raise_exception=True, rtol=RTOL, atol=ATOL)
+    out_fft = ecp.measure_fft(torch.fft.ifft(in_tensor, dim=dim))
+    objects_are_equal(out_fft, out_tensors, raise_exception=True, rtol=RTOL, atol=ATOL)