TTENSOR: Improve test coverage and corresponding bug fixes discovered.

Author: nickj

pyttb/ttensor.py

@@ -15,6 +15,8 @@
 import textwrap
 import warnings
 
+ALT_CORE_ERROR = "TTensor doesn't support non-tensor cores yet"
+
 class ttensor(object):
     """
     TTENSOR Class for Tucker tensors (decomposed).
@@ -55,7 +57,7 @@ def from_data(cls, core, factors):
         >>> import numpy as np
 
         Set up input data
-        # Create sptensor with explicit data description
+        # Create ttensor with explicit data description
 
         >>> core_values = np.ones((2,2,2))
         >>> core = ttb.tensor.from_data(core_values)
@@ -98,6 +100,8 @@ def _validate_ttensor(self):
         """
         # Confirm all factors are matrices
         for factor_idx, factor in enumerate(self.u):
+            if not isinstance(factor, np.ndarray):
+                raise ValueError(f"Factor matrices must be numpy arrays but factor {factor_idx} was {type(factor)}")
             if len(factor.shape) != 2:
                 raise ValueError(f"Factor matrix {factor_idx} has shape {factor.shape} and is not a matrix!")
 
@@ -156,7 +160,7 @@ def full(self):
 
         # There is a small chance tensor could be sparse so ensure we cast that to dense.
         if not isinstance(recomposed_tensor, tensor):
-            recomposed_tensor = tensor(recomposed_tensor)
+            raise ValueError(ALT_CORE_ERROR)
         return recomposed_tensor
 
     def double(self):
@@ -331,7 +335,7 @@ def ttv(self, vector, dims=None):
         # Check that each multiplicand is the right size.
         for i in range(dims.size):
             if vector[vidx[i]].shape != (self.shape[dims[i]],):
-                assert False, "Multiplicand is wrong size"
+                raise ValueError("Multiplicand is wrong size")
 
         # Get remaining dimensions when we're done
         remdims = np.setdiff1d(np.arange(0, self.ndims), dims)
@@ -364,11 +368,7 @@ def mttkrp(self, U, n):
         -------
         :class:`numpy.ndarray`
         """
-
-        if n == 0:
-            R = U[1].shape[-1]
-        else:
-            R = U[0].shape[-1]
+        # NOTE: MATLAB version calculates an unused R here
 
         W = [None] * self.ndims
         for i in range(0, self.ndims):
@@ -444,10 +444,6 @@ def ttm(self, matrix, dims=None, transpose=False):
 
         # Check that each multiplicand is the right size.
         for i in range(len(dims)):
-            import logging
-            logging.warning(
-                f"Matrix shape: \n\t{matrix[vidx[i]].shape}"
-            )
             if matrix[vidx[i]].shape[size_idx] != self.shape[dims[i]]:
                 raise ValueError(f"Multiplicand {i} is wrong size")
 
@@ -482,7 +478,7 @@ def reconstruct(self, samples=None, modes=None):
         if modes is None:
             modes = np.arange(self.ndims)
         elif isinstance(modes, list):
-            modes = np.array([modes])
+            modes = np.array(modes)
         elif np.isscalar(modes):
             modes = np.array([modes])
 
@@ -500,7 +496,10 @@ def reconstruct(self, samples=None, modes=None):
 
         full_samples = [np.array([])] * self.ndims
         for sample, mode in zip(samples, modes):
-            full_samples[mode] = sample
+            if np.isscalar(sample):
+                full_samples[mode] = np.array([sample])
+            else:
+                full_samples[mode] = sample
 
         shape = self.shape
         new_u = []
@@ -510,7 +509,7 @@ def reconstruct(self, samples=None, modes=None):
                 new_u.append(self.u[k])
                 continue
             elif len(full_samples[k].shape) == 2 and full_samples[k].shape[-1] == shape[k]:
-                new_u.append(full_samples[k] * self.u[k])
+                new_u.append(full_samples[k].dot(self.u[k]))
             else:
                 new_u.append(self.u[k][full_samples[k], :])
 
@@ -540,14 +539,14 @@ def nvecs(self, n, r, flipsign = True):
         H = self.core.ttm(V)
 
         if isinstance(H, sptensor):
-            raise NotImplementedError("TTensor doesn't support sparse core yet")
+            raise NotImplementedError(ALT_CORE_ERROR)
         else:
             HnT = tenmat.from_tensor_type(H.full(), cdims=np.array([n])).double()
 
         G = self.core
 
         if isinstance(G, sptensor):
-            raise NotImplementedError("TTensor doesn't support sparse core yet")
+            raise NotImplementedError(ALT_CORE_ERROR)
         else:
             GnT = tenmat.from_tensor_type(G.full(), cdims=np.array([n])).double()
 

tests/test_ttensor.py

@@ -5,6 +5,7 @@
 import numpy as np
 import pyttb as ttb
 import pytest
+import scipy.sparse as sparse
 
 @pytest.fixture()
 def sample_ttensor():
@@ -41,6 +42,30 @@ def test_ttensor_initialization_from_data(sample_ttensor):
     assert isinstance(ttensorInstance.core, ttb.tensor)
     assert all([isinstance(a_factor, np.ndarray) for a_factor in ttensorInstance.u])
 
+    # Negative Tests
+    non_array_factor = ttensorInstance.u + [1]
+    with pytest.raises(ValueError):
+        ttb.ttensor.from_data(ttensorInstance.core, non_array_factor[1:])
+
+    non_matrix_factor = ttensorInstance.u + [np.array([1])]
+    with pytest.raises(ValueError):
+        ttb.ttensor.from_data(ttensorInstance.core, non_matrix_factor[1:])
+
+    too_few_factors = ttensorInstance.u.copy()
+    too_few_factors.pop()
+    with pytest.raises(ValueError):
+        ttb.ttensor.from_data(ttensorInstance.core, too_few_factors)
+
+    wrong_shape_factor = ttensorInstance.u.copy()
+    row, col = wrong_shape_factor[0].shape
+    wrong_shape_factor[0] = np.random.random((row+1, col+1))
+    with pytest.raises(ValueError):
+        ttb.ttensor.from_data(ttensorInstance.core, wrong_shape_factor)
+
+    # Enforce error until sptensor core/other cores supported
+    with pytest.raises(ValueError):
+        ttb.ttensor.from_data(ttb.sptensor.from_tensor_type(ttensorInstance.core), ttensorInstance.u)
+
 @pytest.mark.indevelopment
 def test_ttensor_initialization_from_tensor_type(sample_ttensor):
 
@@ -58,6 +83,17 @@ def test_ttensor_full(sample_ttensor):
     # This sanity check only works for all 1's
     assert tensor.double() == np.prod(ttensorInstance.core.shape)
 
+    # Negative tests
+    sparse_core = ttb.sptensor()
+    sparse_core.shape = ttensorInstance.core.shape
+    sparse_u = [sparse.coo_matrix(np.zeros(factor.shape)) for factor in ttensorInstance.u]
+    # We could probably make these properties to avoid this edge case but expect to eventually cover these alternate
+    # cores
+    ttensorInstance.core = sparse_core
+    ttensorInstance.u = sparse_u
+    with pytest.raises(ValueError):
+        ttensorInstance.full()
+
 @pytest.mark.indevelopment
 def test_ttensor_double(sample_ttensor):
     ttensorInstance = sample_ttensor
@@ -87,17 +123,45 @@ def test_sptensor__neg__(sample_ttensor):
     assert ttensorInstance.isequal(ttensorInstance3)
 
 @pytest.mark.indevelopment
-def test_ttensor_innerproduct(sample_ttensor):
+def test_ttensor_innerproduct(sample_ttensor, random_ttensor):
     ttensorInstance = sample_ttensor
 
     # TODO these are an overly simplistic edge case for ttensors that are a single float
 
     # ttensor innerprod ttensor
     assert ttensorInstance.innerprod(ttensorInstance) == ttensorInstance.double()**2
+    core_dim = ttensorInstance.core.shape[0] + 1
+    ndim = ttensorInstance.ndims
+    large_core_ttensor = ttb.ttensor.from_data(
+        ttb.tensor.from_data(np.ones((core_dim,)*ndim)),
+        [np.ones((1, core_dim))] * ndim
+    )
+    assert large_core_ttensor.innerprod(ttensorInstance) == ttensorInstance.full().innerprod(large_core_ttensor.full())
 
     # ttensor innerprod tensor
     assert ttensorInstance.innerprod(ttensorInstance.full()) == ttensorInstance.double() ** 2
 
+    # ttensr innerprod ktensor
+    ktensorInstance = ttb.ktensor.from_data(np.array([8.]), [np.array([[1.]])]*3)
+    assert ttensorInstance.innerprod(ktensorInstance) == ttensorInstance.double() ** 2
+
+    # ttensor innerprod tensor (shape larger than core)
+    random_ttensor.innerprod(random_ttensor.full())
+
+    # Negative Tests
+    ttensor_extra_factors = ttb.ttensor.from_tensor_type(ttensorInstance)
+    ttensor_extra_factors.u.extend(ttensorInstance.u)
+    with pytest.raises(ValueError):
+        ttensorInstance.innerprod(ttensor_extra_factors)
+
+    tensor_extra_dim = ttb.tensor.from_data(np.ones(ttensorInstance.shape + (1,)))
+    with pytest.raises(ValueError):
+        ttensorInstance.innerprod(tensor_extra_dim)
+
+    invalid_option = []
+    with pytest.raises(ValueError):
+        ttensorInstance.innerprod(invalid_option)
+
 @pytest.mark.indevelopment
 def test_ttensor__mul__(sample_ttensor):
     ttensorInstance = sample_ttensor
@@ -107,6 +171,10 @@ def test_ttensor__mul__(sample_ttensor):
     assert (ttensorInstance * mul_factor).double() == np.prod(ttensorInstance.core.shape) * mul_factor
     assert (ttensorInstance * float(2)).double() == np.prod(ttensorInstance.core.shape) * float(mul_factor)
 
+    # Negative tests
+    with pytest.raises(ValueError):
+        _ = ttensorInstance * 'some_string'
+
 @pytest.mark.indevelopment
 def test_ttensor__rmul__(sample_ttensor):
     ttensorInstance = sample_ttensor
@@ -116,6 +184,10 @@ def test_ttensor__rmul__(sample_ttensor):
     assert (mul_factor * ttensorInstance).double() == np.prod(ttensorInstance.core.shape) * mul_factor
     assert (float(2) * ttensorInstance).double() == np.prod(ttensorInstance.core.shape) * float(mul_factor)
 
+    # Negative tests
+    with pytest.raises(ValueError):
+        _ = 'some_string' * ttensorInstance
+
 @pytest.mark.indevelopment
 def test_ttensor_ttv(sample_ttensor):
     ttensorInstance = sample_ttensor
@@ -124,6 +196,17 @@ def test_ttensor_ttv(sample_ttensor):
     final_value = sample_ttensor.ttv(trivial_vectors)
     assert final_value == np.prod(ttensorInstance.core.shape)
 
+    assert np.allclose(
+        ttensorInstance.ttv(trivial_vectors[0], 0).double(),
+        ttensorInstance.full().ttv(trivial_vectors[0], 0).double()
+    )
+
+    # Negative tests
+    wrong_shape_vector = trivial_vectors.copy()
+    wrong_shape_vector[0] = np.array([mul_factor, mul_factor])
+    with pytest.raises(ValueError):
+        sample_ttensor.ttv(wrong_shape_vector)
+
 @pytest.mark.indevelopment
 def test_ttensor_mttkrp(random_ttensor):
     ttensorInstance = random_ttensor
@@ -133,23 +216,32 @@ def test_ttensor_mttkrp(random_ttensor):
     ]
     final_value = ttensorInstance.mttkrp(vectors, 2)
     full_value = ttensorInstance.full().mttkrp(vectors, 2)
-    assert np.all(np.isclose(final_value, full_value)), (
+    assert np.allclose(final_value, full_value), (
         f"TTensor value is: \n{final_value}\n\n"
         f"Full value is: \n{full_value}"
     )
 
 @pytest.mark.indevelopment
-def test_ttensor_norm(random_ttensor):
+def test_ttensor_norm(sample_ttensor, random_ttensor):
     ttensorInstance = random_ttensor
     assert np.isclose(ttensorInstance.norm(), ttensorInstance.full().norm())
 
+    # Core larger than full tensor
+    ttensorInstance = sample_ttensor
+    assert np.isclose(ttensorInstance.norm(), ttensorInstance.full().norm())
+
 @pytest.mark.indevelopment
 def test_ttensor_permute(random_ttensor):
     ttensorInstance = random_ttensor
     original_order = np.arange(0, len(ttensorInstance.core.shape))
     permuted_tensor = ttensorInstance.permute(original_order)
     assert ttensorInstance.isequal(permuted_tensor)
 
+    # Negative Tests
+    with pytest.raises(ValueError):
+        bad_permutation_order = np.arange(0, len(ttensorInstance.core.shape) + 1)
+        ttensorInstance.permute(bad_permutation_order)
+
 @pytest.mark.indevelopment
 def test_ttensor_ttm(random_ttensor):
     ttensorInstance = random_ttensor
@@ -163,19 +255,91 @@ def test_ttensor_ttm(random_ttensor):
         f"TTensor value is: \n{final_value}\n\n"
         f"Full value is: \n{reverse_value}"
     )
+    final_value = ttensorInstance.ttm(matrices)  # No dims
+    assert final_value.isequal(reverse_value)
+    final_value = ttensorInstance.ttm(matrices, list(range(len(matrices))))  # Dims as list
+    assert final_value.isequal(reverse_value)
+
+
+    single_tensor_result = ttensorInstance.ttm(matrices[0], 0)
+    single_tensor_full_result = ttensorInstance.full().ttm(matrices[0], 0)
+    assert np.allclose(single_tensor_result.double(), single_tensor_full_result.double()), (
+        f"TTensor value is: \n{single_tensor_result.full()}\n\n"
+        f"Full value is: \n{single_tensor_full_result}"
+    )
+
+    transposed_matrices = [matrix.transpose() for matrix in matrices]
+    transpose_value = ttensorInstance.ttm(transposed_matrices, np.arange(len(matrices)), transpose=True)
+    assert final_value.isequal(transpose_value)
+
+    # Negative Tests
+    big_wrong_size = 123
+    matrices[0] = np.random.random((big_wrong_size, big_wrong_size))
+    with pytest.raises(ValueError):
+        _ = ttensorInstance.ttm(matrices, np.arange(len(matrices)))
+
 
 @pytest.mark.indevelopment
 def test_ttensor_reconstruct(random_ttensor):
     ttensorInstance = random_ttensor
     # TODO: This slice drops the singleton dimension, should it? If so should ttensor squeeze during reconstruct?
     full_slice = ttensorInstance.full()[:, 1, :]
     ttensor_slice = ttensorInstance.reconstruct(1, 1)
-    assert np.all(np.isclose(full_slice.double(), ttensor_slice.squeeze().double()))
+    assert np.allclose(full_slice.double(), ttensor_slice.squeeze().double())
     assert ttensorInstance.reconstruct().isequal(ttensorInstance.full())
+    sample_all_modes = [np.array([0])] * len(ttensorInstance.shape)
+    sample_all_modes[-1] = 0  # Make raw scalar
+    reconstruct_scalar = ttensorInstance.reconstruct(sample_all_modes).full().double()
+    full_scalar = ttensorInstance.full()[tuple(sample_all_modes)]
+    assert np.isclose(reconstruct_scalar, full_scalar)
+
+    scale = np.random.random(ttensorInstance.u[1].shape).transpose()
+    _ = ttensorInstance.reconstruct(scale, 1)
+    # FIXME from the MATLAB docs wasn't totally clear how to validate this
+
+    # Negative Tests
+    with pytest.raises(ValueError):
+        _ = ttensorInstance.reconstruct(1, [0, 1])
 
 @pytest.mark.indevelopment
 def test_ttensor_nvecs(random_ttensor):
     ttensorInstance = random_ttensor
-    ttensor_eigvals = ttensorInstance.nvecs(0, 2)
-    full_eigvals = ttensorInstance.full().nvecs(0, 2)
+    n = 0
+    r = 2
+    ttensor_eigvals = ttensorInstance.nvecs(n, r)
+    full_eigvals = ttensorInstance.full().nvecs(n, r)
+    assert np.allclose(ttensor_eigvals, full_eigvals)
+
+    # Test for eig vals larger than shape-1
+    n = 1
+    r = 2
+    full_eigvals = ttensorInstance.full().nvecs(n, r)
+    with pytest.warns(Warning) as record:
+        ttensor_eigvals = ttensorInstance.nvecs(n, r)
+    assert 'Greater than or equal to tensor.shape[n] - 1 eigenvectors requires cast to dense to solve' \
+           in str(record[0].message)
     assert np.allclose(ttensor_eigvals, full_eigvals)
+
+    # Negative Tests
+    sparse_core = ttb.sptensor()
+    sparse_core.shape = ttensorInstance.core.shape
+    ttensorInstance.core = sparse_core
+
+    # Sparse core
+    with pytest.raises(NotImplementedError):
+        ttensorInstance.nvecs(0, 1)
+
+    # Sparse factors
+    sparse_u = [sparse.coo_matrix(np.zeros(factor.shape)) for factor in ttensorInstance.u]
+    ttensorInstance.u = sparse_u
+    with pytest.raises(NotImplementedError):
+        ttensorInstance.nvecs(0, 1)
+
+@pytest.mark.indevelopment
+def test_sptensor_isequal(sample_ttensor):
+    ttensorInstance = sample_ttensor
+    # Negative Tests
+    assert not ttensorInstance.isequal(ttensorInstance.full())
+    ttensor_extra_factors = ttb.ttensor.from_tensor_type(ttensorInstance)
+    ttensor_extra_factors.u.extend(ttensorInstance.u)
+    assert not ttensorInstance.isequal(ttensor_extra_factors)