tests for matmul chain rule

Author: dance858

cvxpy/reductions/dnlp2smooth/canonicalizers/__init__.py

@@ -45,8 +45,7 @@
 from cvxpy.reductions.dnlp2smooth.canonicalizers.log_canon import log_canon
 from cvxpy.reductions.dnlp2smooth.canonicalizers.smooth_full_domain_canon import (
     exp_canon, logistic_canon, sinh_canon, asinh_canon, tanh_canon,
-    sin_canon, cos_canon, normcdf_canon, prod_canon, quad_form_canon)
-from cvxpy.reductions.dnlp2smooth.canonicalizers.multiply_canon import matmul_canon, multiply_canon
+    sin_canon, cos_canon, normcdf_canon, prod_canon, quad_form_canon, matmul_canon, multiply_canon)
 from cvxpy.reductions.dnlp2smooth.canonicalizers.pnorm_canon import pnorm_canon
 from cvxpy.reductions.dnlp2smooth.canonicalizers.power_canon import power_canon
 from cvxpy.reductions.dnlp2smooth.canonicalizers.entr_canon import entr_canon

cvxpy/reductions/dnlp2smooth/canonicalizers/multiply_canon.py

@@ -15,34 +15,15 @@
 """
 
 
-from cvxpy.expressions.variable import Variable
 
 
 # TODO: (dance858) do we need to copy things? Do we need to canonicalize at all?
 # Can we have a common canonicalizer for binary operators with full domain?
-def multiply_canon(expr, args):
-    return expr.copy(args), []
-
-def matmul_canon(expr, args):
-    t1 = args[0]
-    t2 = args[1]
-    constraints = []
-
-    # if either is constant, no canonicalization needed
-    if t1.is_constant() or t2.is_constant():
-        return expr.copy([t1, t2]), []
-
-    if not isinstance(t1, Variable):
-        t1 = Variable(t1.shape)
-        constraints += [t1 == args[0]]
-        t1.value = args[0].value
-
-    if not isinstance(t2, Variable):
-        t2 = Variable(t2.shape)
-        constraints += [t2 == args[1]]
-        t2.value = args[1].value
-
-    return expr.copy([t1, t2]), constraints
+#def multiply_canon(expr, args):
+#    return expr.copy(args), []
+#
+#def matmul_canon(expr, args):
+#    return expr.copy(args), []
 
 # ----------------------------------------------------------------------------------
 #                                Old versions
@@ -67,3 +48,24 @@ def matmul_canon(expr, args):
 #        t2.value = args[1].value
 #
 #    return expr.copy([t1, t2]), constraints
+
+#def matmul_canon(expr, args):
+#    t1 = args[0]
+#    t2 = args[1]
+#    constraints = []
+#
+#    # if either is constant, no canonicalization needed
+#    if t1.is_constant() or t2.is_constant():
+#        return expr.copy([t1, t2]), []
+#
+#    if not isinstance(t1, Variable):
+#        t1 = Variable(t1.shape)
+#        constraints += [t1 == args[0]]
+#        t1.value = args[0].value
+#
+#    if not isinstance(t2, Variable):
+#        t2 = Variable(t2.shape)
+#        constraints += [t2 == args[1]]
+#        t2.value = args[1].value
+#
+#    return expr.copy([t1, t2]), constraints

cvxpy/reductions/dnlp2smooth/canonicalizers/rel_entr_canon.py

@@ -22,7 +22,7 @@
 from cvxpy.expressions.variable import Variable
 from cvxpy.reductions.dnlp2smooth.canonicalizers.entr_canon import entr_canon
 from cvxpy.reductions.dnlp2smooth.canonicalizers.log_canon import log_canon
-from cvxpy.reductions.dnlp2smooth.canonicalizers.multiply_canon import multiply_canon
+from cvxpy.reductions.dnlp2smooth.canonicalizers.smooth_full_domain_canon import multiply_canon
 
 MIN_INIT = 1e-3
 

cvxpy/reductions/dnlp2smooth/canonicalizers/smooth_full_domain_canon.py

@@ -43,4 +43,9 @@ def smooth_full_domain_canon_chain_rule(expr, args):
 logistic_canon = smooth_full_domain_canon_chain_rule
 normcdf_canon = smooth_full_domain_canon_chain_rule
 
+# bivariate atoms with chain rule implemented in diff engine
+multiply_canon = smooth_full_domain_canon_chain_rule
+matmul_canon = smooth_full_domain_canon_chain_rule
+
+
 # TODO: do we even need the smooth full domain canon chain rule canonicalizers?

cvxpy/tests/nlp_tests/stress_tests_diff_engine/test_compositions.py

@@ -129,3 +129,48 @@ def test_quad_form_composition_two(self):
         prob.solve(solver=cp.IPOPT, nlp=True, verbose=True)
         checker = DerivativeChecker(prob)
         checker.run_and_assert()
+
+    def test_matmul_composition_one(self):
+        np.random.seed(0)
+        m, n, p = 5, 7, 11
+        X = cp.Variable((m, n), bounds=[-1, 1], name='X')
+        Y = cp.Variable((n, p), bounds=[-2, 2], name='Y')
+        Y.value = np.random.rand(n, p)
+        obj = cp.sum(cp.matmul(X, cp.cos(Y)))
+        problem = cp.Problem(cp.Minimize(obj))
+        problem.solve(solver=cp.IPOPT, nlp=True, verbose=True)
+        assert(problem.status == cp.OPTIMAL)
+        checker = DerivativeChecker(problem)
+        checker.run_and_assert()
+
+    def test_matmul_composition_two(self):
+        np.random.seed(0)
+        m, n, p = 5, 5, 5
+        X = cp.Variable((m, n), bounds=[-1, 1])
+        Y = cp.Variable((n, p), bounds=[-2, 2])
+        Y.value = np.random.rand(n, p)
+        X.value = np.random.rand(m, n)
+        obj = cp.sum(cp.matmul(cp.matmul(X, X), cp.cos(Y) + X))
+        problem = cp.Problem(cp.Minimize(obj))
+        checker = DerivativeChecker(problem)
+        checker.run_and_assert()
+        problem.solve(solver=cp.IPOPT, nlp=True, verbose=True)
+        assert(problem.status == cp.OPTIMAL)
+        checker = DerivativeChecker(problem)
+        checker.run_and_assert()
+    
+    def test_matmul_composition_three(self):
+        np.random.seed(0)
+        m, n, p = 5, 5, 5
+        X = cp.Variable((m, n), bounds=[-1, 1], name='X')
+        Y = cp.Variable((n, p), bounds=[-2, 2], name='Y')
+        Y.value = np.random.rand(n, p)
+        X.value = np.random.rand(m, n)
+        obj = cp.sum(cp.matmul(cp.matmul(X, X.T), (cp.cos(Y) + X).T))
+        problem = cp.Problem(cp.Minimize(obj))
+        checker = DerivativeChecker(problem)
+        checker.run_and_assert()
+        problem.solve(solver=cp.IPOPT, nlp=True, verbose=True)
+        assert(problem.status == cp.OPTIMAL)
+        checker = DerivativeChecker(problem)
+        checker.run_and_assert()