Bbonev/gradcheck (#9)

* Added gradient check to test suite
* reduced size of the unit test
* switched to parametrized for unittests

.github/workflows/tests.yml

-name: Unit tests
+name: tests
 
 on: [push]
 
@@ -22,5 +22,5 @@ jobs:
         python -m pip install -e .
     - name: Test with pytest
       run: |
-        pip install pytest pytest-cov
-        pytest ./torch_harmonics/tests.py
\ No newline at end of file
+        python -m pip install pytest pytest-cov parameterized
+        python -m pytest ./torch_harmonics/tests.py
\ No newline at end of file

Changelog.md

@@ -2,6 +2,10 @@
 
 ## Versioning
 
+### v0.6.3
+
+* Adding gradient check in unit tests
+
 ### v0.6.2
 
 * Adding github CI

README.md

@@ -43,11 +43,11 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 <!-- ## What is torch-harmonics? -->
 
-`torch-harmonics` is a differentiable implementation of the Spherical Harmonic transform in PyTorch. It was originally implemented to enable Spherical Fourier Neural Operators (SFNO). It uses quadrature rules to compute the projection onto the associated Legendre polynomials and FFTs for the projection onto the harmonic basis. This algorithm tends to outperform others with better asymptotic scaling for most practical purposes.
+torch-harmonics is a differentiable implementation of the Spherical Harmonic transform in PyTorch. It was originally implemented to enable Spherical Fourier Neural Operators (SFNO). It uses quadrature rules to compute the projection onto the associated Legendre polynomials and FFTs for the projection onto the harmonic basis. This algorithm tends to outperform others with better asymptotic scaling for most practical purposes.
 
-`torch-harmonics` uses PyTorch primitives to implement these operations, making it fully differentiable. Moreover, the quadrature can be distributed onto multiple ranks making it spatially distributed.
+torch-harmonics uses PyTorch primitives to implement these operations, making it fully differentiable. Moreover, the quadrature can be distributed onto multiple ranks making it spatially distributed.
 
-`torch-harmonics` has been used to implement a variety of differentiable PDE solvers which generated the animations below. Moreover, it has enabled the development of Spherical Fourier Neural Operators (SFNOs) [1].
+torch-harmonics has been used to implement a variety of differentiable PDE solvers which generated the animations below. Moreover, it has enabled the development of Spherical Fourier Neural Operators (SFNOs) [1].
 
 
 <table border="0" cellspacing="0" cellpadding="0">

torch_harmonics/__init__.py

@@ -29,7 +29,7 @@
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #
 
-__version__ = '0.6.2'
+__version__ = '0.6.3'
 
 from .sht import RealSHT, InverseRealSHT, RealVectorSHT, InverseRealVectorSHT
 from . import quadrature

torch_harmonics/tests.py

@@ -30,8 +30,11 @@
 #
 
 import unittest
+from parameterized import parameterized
+import math
 import numpy as np
 import torch
+from torch.autograd import gradcheck
 from torch_harmonics import *
 
 # try:
@@ -44,7 +47,7 @@ tqdm = lambda x : x
 class TestLegendrePolynomials(unittest.TestCase):
 
     def setUp(self):
-        self.cml = lambda m, l : np.sqrt((2*l + 1) / 4 / np.pi) * np.sqrt(np.math.factorial(l-m) / np.math.factorial(l+m))
+        self.cml = lambda m, l : np.sqrt((2*l + 1) / 4 / np.pi) * np.sqrt(math.factorial(l-m) / math.factorial(l+m))
         self.pml = dict()
 
         # preparing associated Legendre Polynomials (These include the Condon-Shortley phase)
@@ -62,28 +65,23 @@ class TestLegendrePolynomials(unittest.TestCase):
 
         self.lmax = self.mmax = 4
 
+        self.tol = 1e-9
+
     def test_legendre(self):
         print("Testing computation of associated Legendre polynomials")
         from torch_harmonics.legendre import precompute_legpoly
 
-        TOL = 1e-9
-
         t = np.linspace(0, np.pi, 100)
         pct = precompute_legpoly(self.mmax, self.lmax, t)
 
         for l in range(self.lmax):
             for m in range(l+1):
                 diff = pct[m, l].numpy() / self.cml(m,l) - self.pml[(m,l)](np.cos(t))
-                self.assertTrue(diff.max() <= TOL)
-        print("done.")
+                self.assertTrue(diff.max() <= self.tol)
 
 
 class TestSphericalHarmonicTransform(unittest.TestCase):
 
-    def __init__(self, testname, norm="ortho"):
-        super(TestSphericalHarmonicTransform, self).__init__(testname)  # calling the super class init varies for different python versions.  This works for 2.7
-        self.norm = norm
-    
     def setUp(self):
 
         if torch.cuda.is_available():
@@ -93,76 +91,76 @@ class TestSphericalHarmonicTransform(unittest.TestCase):
             print("Running test on CPU")
             self.device = torch.device('cpu')
 
-        self.batch_size = 128
-        self.nlat = 256
-        self.nlon = 2*self.nlat
-
-    def test_sht_leggauss(self):
-        print(f"Testing real-valued SHT on Legendre-Gauss grid with {self.norm} normalization")
+    @parameterized.expand([
+        [256, 512, 32, "ortho",   "equiangular",    1e-9],
+        [256, 512, 32, "ortho",   "legendre-gauss", 1e-9],
+        [256, 512, 32, "four-pi", "equiangular",    1e-9],
+        [256, 512, 32, "four-pi", "legendre-gauss", 1e-9],
+        [256, 512, 32, "schmidt", "equiangular",    1e-9],
+        [256, 512, 32, "schmidt", "legendre-gauss", 1e-9],
+    ])
+    def test_sht(self, nlat, nlon, batch_size, norm, grid, tol):
+        print(f"Testing real-valued SHT on {nlat}x{nlon} {grid} grid with {norm} normalization")
 
-        TOL = 1e-9
         testiters = [1, 2, 4, 8, 16]
-        mmax = self.nlat
+        if grid == "equiangular":
+            mmax = nlat // 2
+        else:
+            mmax = nlat
         lmax = mmax
 
-        sht = RealSHT(self.nlat, self.nlon, mmax=mmax, lmax=lmax, grid="legendre-gauss", norm=self.norm).to(self.device)
-        isht = InverseRealSHT(self.nlat, self.nlon, mmax=mmax, lmax=lmax, grid="legendre-gauss", norm=self.norm).to(self.device)
+        sht = RealSHT(nlat, nlon, mmax=mmax, lmax=lmax, grid=grid, norm=norm).to(self.device)
+        isht = InverseRealSHT(nlat, nlon, mmax=mmax, lmax=lmax, grid=grid, norm=norm).to(self.device)
 
-        coeffs = torch.zeros(self.batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
-        coeffs[:, :lmax, :mmax] = torch.randn(self.batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
+        with torch.no_grad():
+            coeffs = torch.zeros(batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
+            coeffs[:, :lmax, :mmax] = torch.randn(batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
             signal = isht(coeffs)
         
+        # testing error accumulation
         for iter in testiters:
             with self.subTest(i = iter):
-                print(f"{iter} iterations of batchsize {self.batch_size}:")
+                print(f"{iter} iterations of batchsize {batch_size}:")
 
                 base = signal
 
                 for _ in tqdm(range(iter)):
                     base = isht(sht(base))
             
-                # err = ( torch.norm(base-self.signal, p='fro') / torch.norm(self.signal, p='fro') ).item()
-                err = torch.mean(torch.norm(base-signal, p='fro', dim=(-1,-2)) / torch.norm(signal, p='fro', dim=(-1,-2)) ).item()
-                print(f"final relative error: {err}")
-                self.assertTrue(err <= TOL)
-
-    def test_sht_equiangular(self):
-        print(f"Testing real-valued SHT on equiangular grid with {self.norm} normalization")
-
-        TOL = 1e-1
-        testiters = [1, 2, 4, 8]
-        mmax = self.nlat // 2
+                err = torch.mean(torch.norm(base-signal, p='fro', dim=(-1,-2)) / torch.norm(signal, p='fro', dim=(-1,-2)) )
+                print(f"final relative error: {err.item()}")
+                self.assertTrue(err.item() <= tol)
+
+    @parameterized.expand([
+        [12, 24, 2, "ortho",   "equiangular",    1e-5],
+        [12, 24, 2, "ortho",   "legendre-gauss", 1e-5],
+        [12, 24, 2, "four-pi", "equiangular",    1e-5],
+        [12, 24, 2, "four-pi", "legendre-gauss", 1e-5],
+        [12, 24, 2, "schmidt", "equiangular",    1e-5],
+        [12, 24, 2, "schmidt", "legendre-gauss", 1e-5],
+    ])
+    def test_sht_grad(self, nlat, nlon, batch_size, norm, grid, tol):
+        print(f"Testing gradients of real-valued SHT on {nlat}x{nlon} {grid} grid with {norm} normalization")
+
+        if grid == "equiangular":
+            mmax = nlat // 2
+        else:
+            mmax = nlat
         lmax = mmax
 
-        sht = RealSHT(self.nlat, self.nlon, mmax=mmax, lmax=lmax, grid="equiangular", norm=self.norm).to(self.device)
-        isht = InverseRealSHT(self.nlat, self.nlon, mmax=mmax, lmax=lmax, grid="equiangular", norm=self.norm).to(self.device)
+        sht = RealSHT(nlat, nlon, mmax=mmax, lmax=lmax, grid=grid, norm=norm).to(self.device)
+        isht = InverseRealSHT(nlat, nlon, mmax=mmax, lmax=lmax, grid=grid, norm=norm).to(self.device)
 
-        coeffs = torch.zeros(self.batch_size, sht.lmax, sht.mmax, device=self.device, dtype=torch.complex128)
-        coeffs[:, :lmax, :mmax] = torch.randn(self.batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
+        with torch.no_grad():
+            coeffs = torch.zeros(batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
+            coeffs[:, :lmax, :mmax] = torch.randn(batch_size, lmax, mmax, device=self.device, dtype=torch.complex128)
             signal = isht(coeffs)
         
-        for iter in testiters:
-            with self.subTest(i = iter):
-                print(f"{iter} iterations of batchsize {self.batch_size}:")
-
-                base = signal
-
-                for _ in tqdm(range(iter)):
-                    base = isht(sht(base))
-            
-                # err = ( torch.norm(base-self.signal, p='fro') / torch.norm(self.signal, p='fro') ).item()
-                err = torch.mean(torch.norm(base-signal, p='fro', dim=(-1,-2)) / torch.norm(signal, p='fro', dim=(-1,-2)) ).item()
-                print(f"final relative error: {err}")
-                self.assertTrue(err <= TOL)
+        input = torch.randn_like(signal, requires_grad=True)
+        err_handle = lambda x : torch.mean(torch.norm( isht(sht(x)) - signal , p='fro', dim=(-1,-2)) / torch.norm(signal, p='fro', dim=(-1,-2)) )
+        test_result = gradcheck(err_handle, input, eps=1e-6, atol=tol)
+        self.assertTrue(test_result)
 
 
 if __name__ == '__main__':
-    sht_test_suite = unittest.TestSuite()
-    sht_test_suite.addTest(TestLegendrePolynomials('test_legendre'))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_leggauss',    norm="ortho"))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_equiangular', norm="ortho"))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_leggauss',    norm="four-pi"))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_equiangular', norm="four-pi"))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_leggauss',    norm="schmidt"))
-    sht_test_suite.addTest(TestSphericalHarmonicTransform('test_sht_equiangular', norm="schmidt"))
-    unittest.TextTestRunner(verbosity=2).run(sht_test_suite)
\ No newline at end of file
+    unittest.main()
\ No newline at end of file