CP_APR: Add tests fpr sptensor, and corresponding bug fixes to improve test coverage.

Author: ntjohnson1

pyttb/cp_apr.py

@@ -15,7 +15,7 @@ def cp_apr(tensor, rank, algorithm='mu', stoptol=1e-4, stoptime=1e6, maxiters=10
            kappa=0.01, kappatol=1e-10, epsActive=1e-8, mu0=1e-5, precompinds=True, inexact=True,
            lbfgsMem=3):
     """
-    Compute nonnegative CP with alternating Poisson regression.
+    Compute non-negative CP with alternating Poisson regression.
     
     Parameters
     ----------
@@ -376,9 +376,10 @@ def tt_cp_apr_pdnr(tensor, rank, init, stoptol, stoptime, maxiters, maxinneriter
         sparseIx = []
         for n in range(N):
             num_rows = M[n].shape[0]
-            sparseIx.append(np.zeros((num_rows, 1)))
+            row_indices = []
             for jj in range(num_rows):
-                sparseIx[n][jj] = np.where(tensor.subs[:, n] == jj)[0]
+                row_indices.append(np.where(tensor.subs[:, n] == jj)[0])
+            sparseIx.append(row_indices)
 
         if printitn > 0:
             print('done\n')
@@ -426,7 +427,7 @@ def tt_cp_apr_pdnr(tensor, rank, init, stoptol, stoptime, maxiters, maxinneriter
                         M.factor_matrices[n][jj, :] = 0
                         continue
 
-                    x_row = tensor.vals(sparse_indices)
+                    x_row = tensor.vals[sparse_indices]
 
                     # Calculate just the columns of Pi needed for this row.
                     Pi = ttb.tt_calcpi_prowsubprob(tensor, M, rank, n, N, isSparse, sparse_indices)
@@ -683,9 +684,10 @@ def tt_cp_apr_pqnr(tensor, rank, init, stoptol, stoptime, maxiters, maxinneriter
         sparseIx = []
         for n in range(N):
             num_rows = M[n].shape[0]
-            sparseIx.append(np.zeros((num_rows, 1)))
+            row_indices = []
             for jj in range(num_rows):
-                sparseIx[n][jj] = np.where(tensor.subs[:, n] == jj)[0]
+                row_indices.append(np.where(tensor.subs[:, n] == jj)[0])
+            sparseIx.append(row_indices)
 
         if printitn > 0:
             print('done\n')
@@ -726,7 +728,7 @@ def tt_cp_apr_pqnr(tensor, rank, init, stoptol, stoptime, maxiters, maxinneriter
                         M.factor_matrices[n][jj, :] = 0
                         continue
 
-                    x_row = tensor.vals(sparse_indices)
+                    x_row = tensor.vals[sparse_indices]
 
                     # Calculate just the columns of Pi needed for this row.
                     Pi = ttb.tt_calcpi_prowsubprob(tensor, M, rank, n, N, isSparse, sparse_indices)
@@ -1227,7 +1229,7 @@ def tt_loglikelihood_row(isSparse, data_row, model_row, Pi):
     """
     term1 = -np.sum(model_row)
     if isSparse:
-        term2 = np.sum(data_row.transpose() * np.log(model_row.dot(Pi)))
+        term2 = np.sum(data_row.transpose() * np.log(model_row.dot(Pi.transpose())))
     else:
         b_pi = model_row.dot(Pi.transpose())
         term2 = 0

tests/test_cp_apr.py

@@ -125,6 +125,13 @@ def test_cpapr_mu(capsys):
     capsys.readouterr()
     assert output['nTotalIters'] == 2
 
+    # Edge cases
+    # Confirm timeout works
+    non_correct_answer = ktensorInstance*2
+    _ = ttb.cp_apr(tensorInstance, 2, init=non_correct_answer, stoptime=-1)
+    out, _ = capsys.readouterr()
+    assert "time limit exceeded" in out
+
 @pytest.mark.indevelopment
 def test_cpapr_pdnr(capsys):
     # Test simple case
@@ -139,6 +146,23 @@ def test_cpapr_pdnr(capsys):
     capsys.readouterr()
     assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-04).all()
 
+    # Try solve with sptensor
+    sptensorInstance = ttb.sptensor.from_tensor_type(tensorInstance)
+    np.random.seed(123)
+    M, _, _ = ttb.cp_apr(sptensorInstance, 2, algorithm="pdnr")
+    capsys.readouterr()
+    assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-04).all()
+    M, _, _ = ttb.cp_apr(sptensorInstance, 2, algorithm="pdnr", precompinds=False)
+    capsys.readouterr()
+    assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-04).all()
+
+    # Edge cases
+    # Confirm timeout works
+    non_correct_answer = ktensorInstance*2
+    _ = ttb.cp_apr(tensorInstance, 2, init=non_correct_answer, algorithm="pdnr", stoptime=-1)
+    out, _ = capsys.readouterr()
+    assert "time limit exceeded" in out
+
 @pytest.mark.indevelopment
 def test_cpapr_pqnr(capsys):
     # Test simple case
@@ -165,6 +189,22 @@ def test_cpapr_pqnr(capsys):
     capsys.readouterr()
     assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-01).all()
 
+    # Try solve with sptensor
+    sptensorInstance = ttb.sptensor.from_tensor_type(tensorInstance)
+    np.random.seed(123)
+    M, _, _ = ttb.cp_apr(sptensorInstance, 2, algorithm="pqnr")
+    capsys.readouterr()
+    assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-01).all()
+    M, _, _ = ttb.cp_apr(sptensorInstance, 2, algorithm="pqnr", precompinds=False)
+    capsys.readouterr()
+    assert np.isclose(M.full().data, ktensorInstance.full().data, rtol=1e-01).all()
+
+    # Edge cases
+    # Confirm timeout works
+    _ = ttb.cp_apr(tensorInstance, 2, algorithm="pqnr", stoptime=-1)
+    out, _ = capsys.readouterr()
+    assert "time limit exceeded" in out
+
 
 # PDNR tests below
 @pytest.mark.indevelopment
@@ -367,3 +407,22 @@ def test_getSearchDirPqnr():
     search, pred = ttb.getSearchDirPqnr(model_row, phi, 1e-6, delta_model, delta_grad, phi, 1, 5, False)
     # This only verifies that for the right shaped input nothing crashes. Doesn't verify correctness
     assert True
+
+def test_cp_apr_negative_tests():
+    dense_tensor = ttb.tensor.from_data(np.ones((2, 2, 2)))
+    bad_weights = np.array([8.])
+    bad_factors = [np.array([[1.]])] * 3
+    bad_initial_guess_shape = ttb.ktensor.from_data(bad_weights, bad_factors)
+    with pytest.raises(AssertionError):
+        ttb.cp_apr(dense_tensor, init=bad_initial_guess_shape, rank=1)
+    good_weights = np.array([8.] * 3)
+    good_factor = np.array([[1., 1., 1.], [1., 1., 1.]])
+    bad_initial_guess_factors = ttb.ktensor.from_data(good_weights, [-1. * good_factor] * 3)
+    with pytest.raises(AssertionError):
+        ttb.cp_apr(dense_tensor, init=bad_initial_guess_factors, rank=3)
+    bad_initial_guess_weight = ttb.ktensor.from_data(-1. * good_weights, [good_factor] * 3)
+    with pytest.raises(AssertionError):
+        ttb.cp_apr(dense_tensor, init=bad_initial_guess_weight, rank=3)
+
+    with pytest.raises(AssertionError):
+        ttb.cp_apr(dense_tensor, rank=1, algorithm='UNSUPPORTED_ALG')