Merge pull request #29 from meyer-lab/normalize

Normalize matrix weights to a mWeights in CMTF

Author: cyrillustan

tensorpack/cmtf.py

@@ -16,6 +16,8 @@
 
 def buildMat(tFac):
     """ Build the matrix in CMTF from the factors. """
+    if hasattr(tFac, 'mWeights'):
+        return tFac.factors[0] @ (tFac.mFactor * tFac.mWeights).T
     return tFac.factors[0] @ tFac.mFactor.T
 
 
@@ -78,7 +80,7 @@ def sort_factors(tFac):
 
     # Add the variance of the matrix
     if hasattr(tFac, 'mFactor'):
-        norm += np.sum(np.square(tFac.factors[0]), axis=0) * np.sum(np.square(tFac.mFactor), axis=0)
+        norm += np.sum(np.square(tFac.factors[0]), axis=0) * np.sum(np.square(tFac.mFactor), axis=0) * tFac.mWeights
 
     order = np.flip(np.argsort(norm))
     tensor.weights = tensor.weights[order]
@@ -87,6 +89,7 @@ def sort_factors(tFac):
 
     if hasattr(tFac, 'mFactor'):
         tensor.mFactor = tensor.mFactor[:, order]
+        tensor.mWeights = tensor.mWeights[order]
         np.testing.assert_allclose(buildMat(tFac), buildMat(tensor), atol=1e-9)
 
     return tensor
@@ -106,6 +109,7 @@ def delete_component(tFac, compNum):
 
     if hasattr(tFac, 'mFactor'):
         tensor.mFactor = np.delete(tensor.mFactor, compNum, axis=1)
+        tensor.mWeights = np.delete(tensor.mWeights, compNum)
 
     tensor.factors = [np.delete(fac, compNum, axis=1) for fac in tensor.factors]
     return tensor
@@ -142,7 +146,9 @@ def cp_normalize(tFac):
         scales = np.linalg.norm(factor, ord=np.inf, axis=0)
         tFac.weights *= scales
         if i == 0 and hasattr(tFac, 'mFactor'):
-            tFac.mFactor *= scales
+            mScales = np.linalg.norm(tFac.mFactor, ord=np.inf, axis=0)
+            tFac.mWeights = scales * mScales
+            tFac.mFactor /= mScales
 
         tFac.factors[i] /= scales
 
@@ -245,16 +251,16 @@ def perform_CP(tOrig, r=6, tol=1e-6):
     return tFac
 
 
-def perform_CMTF(tOrig, mOrig=None, r=9, tol=1e-6, maxiter=50):
+def perform_CMTF(tOrig, mOrig, r=9, tol=1e-6, maxiter=50):
     """ Perform CMTF decomposition. """
     assert tOrig.dtype == float
-    if mOrig is not None:
-        assert mOrig.dtype == float
+    assert mOrig.dtype == float
     tFac = initialize_cmtf(tOrig, mOrig, r)
 
     # Pre-unfold
     unfolded = np.hstack((tl.unfold(tOrig, 0), mOrig))
     missingM = np.all(np.isfinite(mOrig), axis=1)
+    assert np.sum(missingM) >= 1, "mOrig must contain at least one complete row"
     R2X = -np.inf
 
     # Precalculate the missingness patterns
@@ -279,6 +285,7 @@ def perform_CMTF(tOrig, mOrig=None, r=9, tol=1e-6, maxiter=50):
         if R2X - R2X_last < tol:
             break
 
+    assert not np.all(tFac.mFactor == 0.0)
     tFac = cp_normalize(tFac)
     tFac = reorient_factors(tFac)
     tFac = sort_factors(tFac)

tensorpack/test/test_cmtf.py

@@ -20,7 +20,7 @@ def test_cmtf_R2X():
     """ Test to ensure R2X for higher components is larger. """
     arr = []
     tensor = createCube(missing=0.2, size=(10, 20, 25))
-    matrix = createCube(missing=0.2, size=(10, 15))
+    matrix = createCube(missing=0.05, size=(10, 15))
     for i in range(1, 5):
         facT = perform_CMTF(tensor, matrix, r=i)
         assert np.all(np.isfinite(facT.factors[0]))
@@ -58,7 +58,7 @@ def test_cp():
 def test_delete():
     """ Test deleting a component results in a valid tensor. """
     tOrig = createCube(missing=0.2, size=(10, 20, 25))
-    mOrig = createCube(missing=0.2, size=(10, 15))
+    mOrig = createCube(missing=0.05, size=(10, 15))
     facT = perform_CMTF(tOrig, mOrig, r=4)
 
     fullR2X = calcR2X(facT, tOrig, mOrig)
@@ -79,6 +79,7 @@ def test_sort():
 
     tFac = random_cp(tOrig.shape, 3)
     tFac.mFactor = np.random.randn(mOrig.shape[1], 3)
+    tFac.mWeights = np.ones(3)
 
     R2X = calcR2X(tFac, tOrig, mOrig)
     tRec = tl.cp_to_tensor(tFac)