Merge pull request #35 from sdpython/update

Fixes precision of many unit tests

Author: xadupre

.gitignore

@@ -278,3 +278,5 @@ _doc/sphinxdoc/source/c_ml/*.html
 _unittests/ut_graph/*.prof
 _doc/notebooks/ml/onnxruntime*.json
 _doc/notebooks/ml/tree_*.onnx
+*.out
+*.err

_unittests/ut_ml/test_matrices.py

@@ -21,7 +21,7 @@ def test_gram_schmidt(self):
 
         mat = numpy.array([[1, 0.5], [0.5, 1]], dtype=float)
         res = gram_schmidt(mat)
-        self.assertEqualArray(mat1, res @ res.T)
+        self.assertEqualArray(mat1, res @ res.T, atol=1e-10)
         res2 = gram_schmidt(mat.T).T
 
         res2, change2 = gram_schmidt(mat, change=True)
@@ -40,7 +40,7 @@ def test_gram_schmidt(self):
         mat1 = numpy.array(
             [[1, 0.5, 0], [0, 0.5, 1], [1, 0.5, 1]], dtype=float)
         res = gram_schmidt(mat1)
-        self.assertEqualArray(numpy.identity(3), res @ res.T)
+        self.assertEqualArray(numpy.identity(3), res @ res.T, atol=1e-10)
 
     def test_gram_schmidt_xx(self):
         X = numpy.array([[1, 0.5, 0], [0, 0.4, 2]], dtype=float).T
@@ -49,11 +49,11 @@ def test_gram_schmidt_xx(self):
         T = T.T
         m = P.T @ X.T
         z = m @ m.T
-        self.assertEqual(z, numpy.identity(2))
+        self.assertEqualArray(z, numpy.identity(2), atol=1e-10)
         m = X @ P
-        self.assertEqual(m, T)
+        self.assertEqualArray(m, T, atol=1e-10)
         z2 = m.T @ m
-        self.assertEqual(z2, numpy.identity(2))
+        self.assertEqualArray(z2, numpy.identity(2), atol=1e-10)
 
     def test_linear_regression(self):
         X = numpy.array([[1, 0.5, 0], [0, 0.4, 2]], dtype=float).T
@@ -118,7 +118,7 @@ def test_inner_code(self):
         self.assertEqual(Pt.shape, (X.shape[1], X.shape[1]))
         _Tt = Pt @ Xt
         self.assertEqualArray(_Tt, Tt)
-        self.assertEqualArray(Tt @ Tt.T, numpy.identity(Tt.shape[0]))
+        self.assertEqualArray(Tt @ Tt.T, numpy.identity(Tt.shape[0]), atol=1e-10)
 
         beta1 = numpy.linalg.inv(Xt @ X) @ Xt @ y
         beta2 = Tt @ y @ Pt
@@ -140,8 +140,8 @@ def test_streaming_gram_schmidt(self):
                 algo1_t.append(t)
                 algo1.append(p)
                 t_ = X[:i] @ p.T
-                self.assertEqualArray(t @ t.T, idd)
-                self.assertEqualArray(t_.T @ t_, idd)
+                self.assertEqualArray(t @ t.T, idd, atol=1e-10)
+                self.assertEqualArray(t_.T @ t_, idd, atol=1e-10)
             algo2 = []
             self.assertRaise(lambda: list(streaming_gram_schmidt(X)),  # pylint: disable=W0640
                              RuntimeError)
@@ -151,7 +151,7 @@ def test_streaming_gram_schmidt(self):
                 algo2.append(p2)
                 self.assertNotEmpty(t2)
                 self.assertNotEmpty(p2)
-                self.assertEqualArray(t2.T @ t2, idd)
+                self.assertEqualArray(t2.T @ t2, idd, atol=1e-10)
             self.assertEqual(len(algo1), len(algo2))
 
     def test_streaming_linear_regression(self):
@@ -210,17 +210,17 @@ def test_profile(self):
         res = self.profile(lambda: list(
             streaming_linear_regression_gram_schmidt(X, y)))
         if __name__ == "__main__":
-            print(res[1])
+            print("***", res[1])
         self.assertIn("streaming", res[1])
         res = self.profile(lambda: list(streaming_linear_regression(X, y)))
         if __name__ == "__main__":
-            print(res[1])
+            print("***", res[1])
         self.assertIn("streaming", res[1])
 
     def test_norm2(self):
         X = numpy.array([[1, 0.5, 0], [0, 0.4, 2]], dtype=float).T
         n2 = norm2(X)
-        print(n2)
+        self.assertNotEmpty(n2)
 
 
 if __name__ == "__main__":

_unittests/ut_ml/test_neural_tree.py

@@ -105,7 +105,7 @@ def test_convert(self):
         exp = tree.predict_proba(X)
         got = root.predict(X)
         self.assertEqual(exp.shape[0], got.shape[0])
-        self.assertEqualArray(exp, got[:, -2:])
+        self.assertEqualArray(exp, got[:, -2:], atol=1e-10)
 
     def test_dot(self):
         data = load_iris()
@@ -550,7 +550,7 @@ def test_convert_compact_skl(self):
 
         skl = NeuralTreeNetClassifier(root)
         prob = skl.predict_proba(X)
-        self.assertEqualArray(exp, prob)
+        self.assertEqualArray(exp, prob, atol=1e-10)
         lab = skl.predict(X)
         self.assertEqual(lab.shape, (X.shape[0], ))
 
@@ -564,7 +564,7 @@ def test_convert_compact_skl_fit(self):
         skl.fit(X, y)
         exp = tree.predict_proba(X)
         got = skl.predict_proba(X)
-        self.assertEqualArray(exp, got)
+        self.assertEqualArray(exp, got, atol=1e-10)
 
     def test_convert_compact_skl_onnx(self):
         from mlprodict.onnx_conv import to_onnx
@@ -578,9 +578,9 @@ def test_convert_compact_skl_onnx(self):
         skl = NeuralTreeNetClassifier(root)
         got = skl.predict_proba(X)
         exp = tree.predict_proba(X)
-        self.assertEqualArray(exp, got)
+        self.assertEqualArray(exp, got, atol=1e-10)
         dec = root.predict(X)
-        self.assertEqualArray(exp, dec[:, -2:])
+        self.assertEqualArray(exp, dec[:, -2:], atol=1e-10)
 
         x32 = X.astype(numpy.float32)
         onx = to_onnx(skl, x32, target_opset=15)
@@ -589,7 +589,7 @@ def test_convert_compact_skl_onnx(self):
         self.assertIn('Softmax(', text)
         oinf = OnnxInference(onx)
         got2 = oinf.run({'X': x32})['probabilities']
-        self.assertEqualArray(exp, got2)
+        self.assertEqualArray(exp, got2, atol=1e-10)
 
     def test_convert_reg_compact(self):
         X = numpy.arange(32).astype(numpy.float64).reshape((-1, 2))

_unittests/ut_module/test_code_style.py

@@ -1,5 +1,5 @@
 """
-@brief      test log(time=0s)
+@brief      test log(time=100s)
 """
 import os
 import unittest