Merge pull request #19 from meyer-lab/tSVD

Use truncated SVD

Author: cyrillustan

pyproject.toml

@@ -6,11 +6,12 @@ authors = ["Your Name <[email protected]>"]
 license = "MIT"
 
 [tool.poetry.dependencies]
-python = "^3.9"
+python = ">=3.9,<3.11"
 numpy = "^1.21"
 scipy = "^1.7"
 statsmodels = "^0.13"
 tensorly = "^0.7"
+scikit-learn = "^1.0.1"
 
 [tool.poetry.dev-dependencies]
 pytest = "^6.2"

setup.py

@@ -1,7 +1,7 @@
 from setuptools import setup, find_packages
 
 setup(name='tensorpack',
-      version='0.1',
+      version='0.0.2',
       description='A collection of tensor methods from the Meyer lab.',
       url='https://github.com/meyer-lab/tensorpack',
       license='MIT',

tensorpack/cmtf.py

@@ -3,6 +3,7 @@
 """
 
 import numpy as np
+from sklearn.decomposition import TruncatedSVD
 import tensorly as tl
 from tensorly.tenalg import khatri_rao
 from copy import deepcopy
@@ -192,6 +193,8 @@ def initialize_cp(tensor: np.ndarray, rank: int):
     factors = [np.ones((tensor.shape[i], rank)) for i in range(tensor.ndim)]
     contain_missing = (np.sum(~np.isfinite(tensor)) > 0)
 
+    tsvd = TruncatedSVD(n_components=rank)
+
     # SVD init mode whose size is larger than rank
     for mode in range(tensor.ndim):
         if tensor.shape[mode] >= rank:
@@ -200,7 +203,7 @@ def initialize_cp(tensor: np.ndarray, rank: int):
                 si = SoftImpute(max_rank=rank)
                 unfold = si.fit_transform(unfold)
 
-            factors[mode] = np.linalg.svd(unfold)[0][:, :rank]
+            factors[mode] = tsvd.fit_transform(unfold)
 
     return tl.cp_tensor.CPTensor((None, factors))