You can not select more than 25 topics
			Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
		
		
		
		
		
			
		
			
				
					
					
						
							96 lines
						
					
					
						
							2.6 KiB
						
					
					
				
			
		
		
	
	
							96 lines
						
					
					
						
							2.6 KiB
						
					
					
				| import timeit
 | |
| 
 | |
| import numpy as np
 | |
| import numpy.linalg
 | |
| from scipy.linalg import cho_factor, cho_solve
 | |
| 
 | |
| # We are trying to solve the following system
 | |
| # (A.T * A) * x = A.T * b
 | |
| # Where x are the polynomial coefficients and b is are the input points
 | |
| 
 | |
| # First we build A
 | |
| deg = 3
 | |
| x = np.arange(50 * 1.0)
 | |
| A = np.vstack(tuple(x**n for n in range(deg, -1, -1))).T
 | |
| 
 | |
| # The first way to solve this is using the pseudoinverse, which can be precomputed
 | |
| # x = (A.T * A)^-1 * A^T * b = PINV b
 | |
| PINV = np.linalg.pinv(A)
 | |
| 
 | |
| # Another way is using the Cholesky decomposition
 | |
| # We can note that at (A.T * A) is always positive definite
 | |
| # By precomputing the Cholesky decomposition we can efficiently solve
 | |
| # systems of the form (A.T * A) x = c
 | |
| CHO = cho_factor(np.dot(A.T, A))
 | |
| 
 | |
| 
 | |
| def model_polyfit_old(points, deg=3):
 | |
|   A = np.vstack(tuple(x**n for n in range(deg, -1, -1))).T
 | |
|   pinv = np.linalg.pinv(A)
 | |
|   return np.dot(pinv, map(float, points))
 | |
| 
 | |
| 
 | |
| def model_polyfit(points, deg=3):
 | |
|   A = np.vander(x, deg + 1)
 | |
|   pinv = np.linalg.pinv(A)
 | |
|   return np.dot(pinv, map(float, points))
 | |
| 
 | |
| 
 | |
| def model_polyfit_cho(points, deg=3):
 | |
|   A = np.vander(x, deg + 1)
 | |
|   cho = cho_factor(np.dot(A.T, A))
 | |
|   c = np.dot(A.T, points)
 | |
|   return cho_solve(cho, c, check_finite=False)
 | |
| 
 | |
| 
 | |
| def model_polyfit_np(points, deg=3):
 | |
|   return np.polyfit(x, points, deg)
 | |
| 
 | |
| 
 | |
| def model_polyfit_lstsq(points, deg=3):
 | |
|   A = np.vander(x, deg + 1)
 | |
|   return np.linalg.lstsq(A, points, rcond=None)[0]
 | |
| 
 | |
| 
 | |
| TEST_DATA = np.linspace(0, 5, num=50) + 1.
 | |
| 
 | |
| 
 | |
| def time_pinv_old():
 | |
|   model_polyfit_old(TEST_DATA)
 | |
| 
 | |
| 
 | |
| def time_pinv():
 | |
|   model_polyfit(TEST_DATA)
 | |
| 
 | |
| 
 | |
| def time_cho():
 | |
|   model_polyfit_cho(TEST_DATA)
 | |
| 
 | |
| 
 | |
| def time_np():
 | |
|   model_polyfit_np(TEST_DATA)
 | |
| 
 | |
| 
 | |
| def time_lstsq():
 | |
|   model_polyfit_lstsq(TEST_DATA)
 | |
| 
 | |
| 
 | |
| if __name__ == "__main__":
 | |
|   # Verify correct results
 | |
|   pinv_old = model_polyfit_old(TEST_DATA)
 | |
|   pinv = model_polyfit(TEST_DATA)
 | |
|   cho = model_polyfit_cho(TEST_DATA)
 | |
|   numpy = model_polyfit_np(TEST_DATA)
 | |
|   lstsq = model_polyfit_lstsq(TEST_DATA)
 | |
| 
 | |
|   assert all(np.isclose(pinv, pinv_old))
 | |
|   assert all(np.isclose(pinv, cho))
 | |
|   assert all(np.isclose(pinv, numpy))
 | |
|   assert all(np.isclose(pinv, lstsq))
 | |
| 
 | |
|   # Run benchmark
 | |
|   print("Pseudo inverse (old)", timeit.timeit("time_pinv_old()", setup="from __main__ import time_pinv_old", number=10000))
 | |
|   print("Pseudo inverse", timeit.timeit("time_pinv()", setup="from __main__ import time_pinv", number=10000))
 | |
|   print("Cholesky", timeit.timeit("time_cho()", setup="from __main__ import time_cho", number=10000))
 | |
|   print("Numpy leastsq", timeit.timeit("time_lstsq()", setup="from __main__ import time_lstsq", number=10000))
 | |
|   print("Numpy polyfit", timeit.timeit("time_np()", setup="from __main__ import time_np", number=10000))
 | |
| 
 |