suggestion: orders of magnitude speed-up using Numba JIT

[adrianomitre]

In my machine, the following demonstrates more than 100x speed-up.

import pandas as pd
import numpy as np
import timeit
import hampel # assumes `pip install hampel`

def much_faster_hampel(series, window_length, threshold=3):
  # Please refer to [1] for `hampel_filter_forloop_numba` implementation.
  # [1] https://gist.github.com/adrianomitre/2ced5544421eca031180782cc30798af 
  result, outlier_indices = hampel_filter_forloop_numba(series.values, window_length, threshold)
  result[outlier_indices] = np.nan
  return result

ts_len = 2**8
win_sz = 7

ts = pd.Series(np.random.rand(ts_len))

# causes JIT compilation
much_faster_hampel(ts, win_sz)

timeit.timeit(lambda: hampel.hampel(ts, win_sz), number=100)
timeit.timeit(lambda: much_faster_hampel(ts, win_sz), number=100)

Please refer to this gist for hampel_filter_forloop_numba implementation.

[dwervin]

In my machine, the following demonstrates more than 100x speed-up.

import pandas as pd
import numpy as np
import timeit
import hampel # assumes `pip install hampel`

def much_faster_hampel(series, window_length, threshold=3):
  # Please refer to [1] for `hampel_filter_forloop_numba` implementation.
  # [1] https://gist.github.com/adrianomitre/2ced5544421eca031180782cc30798af 
  result, outlier_indices = hampel_filter_forloop_numba(series.values, window_length, threshold)
  result[outlier_indices] = np.nan
  return result

ts_len = 2**8
win_sz = 7

ts = pd.Series(np.random.rand(ts_len))

# causes JIT compilation
much_faster_hampel(ts, win_sz)

timeit.timeit(lambda: hampel.hampel(ts, win_sz), number=100)
timeit.timeit(lambda: much_faster_hampel(ts, win_sz), number=100)

Please refer to this gist for hampel_filter_forloop_numba implementation.

Sorry, but you are referencing a completely different library. The library you are importing is not at all used in mine.