"""
Benchmarks of with vs without CSR conversion

The performances are evaluated in various input dimensions.

The input array X is a n x n square matrix.

Time is calculated by running the transform 100 times for each method.

Performance Comparison
======================
Input_Dimensions Time_With_CSR_Conversion Time_Without_Conversion
-----------------------------------------------------------------
      1000               0.0339s                  0.0321s
      2000               0.1433s                  0.1376s
      3000               0.4040s                  0.4357s
      4000               0.7765s                  0.7333s
      5000               1.3254s                  1.3391s
      6000               2.0317s                  1.9206s
      7000               2.8399s                  2.7303s
      8000               3.6815s                  3.5850s
      9000               4.8620s                  4.6881s
"""
import scipy.sparse as ss
from sklearn.decomposition import TruncatedSVD
from time import time
from sklearn.utils import check_array


def benchmark_converted(dim, trials=100):
    start_time = time()
    X = ss.random(dim, dim, format='csc')
    tsvd = TruncatedSVD().fit(X)
    for i in range(trials):
        X = check_array(X, accept_sparse='csr')
        tsvd.transform(X)
    return time() - start_time


def benchmark_direct(dim, trials=100):
    start_time = time()
    X = ss.random(dim, dim, format='csc')
    tsvd = TruncatedSVD().fit(X)
    for i in range(trials):
        tsvd.transform(X)
    return time() - start_time


if __name__ == '__main__':
    print("Performance Comparison")
    print("======================")
    headers = ("Input_Dimensions",
               "Time_With_CSR_Conversion",
               "Time_Without_Conversion")
    header_len = [len(header) for header in headers]
    print("%s %s %s"
          % headers)
    print("-" * (sum(header_len) + 2))
    for dim in range(1000, 10000, 1000):
        t1 = benchmark_converted(dim)
        t2 = benchmark_direct(dim)
        print("%s %s %s" % (("%d" % dim).center(header_len[0]),
                            ("%.4fs" % t1).center(header_len[1]),
                            ("%.4fs" % t2).center(header_len[2])))