In [144]: ser = pd.Series(arr[:, 0])
│ │ │ │
│ │ │ │ In [145]: %timeit ser.iloc[indexer]
│ │ │ │ .....: %timeit ser.take(indexer)
│ │ │ │ .....:
│ │ │ │ -147 us +- 483 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
│ │ │ │ -129 us +- 433 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
│ │ │ │ +501 us +- 70.5 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ +388 us +- 86.5 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ We have discussed MultiIndex in the previous sections pretty extensively.
│ │ │ │ Documentation about DatetimeIndex and PeriodIndex are shown here,
│ │ │ │ ├── html2text {}
│ │ │ │ │ @@ -1256,23 +1256,23 @@
│ │ │ │ │ In [141]: indexer = np.arange(10000)
│ │ │ │ │
│ │ │ │ │ In [142]: random.shuffle(indexer)
│ │ │ │ │
│ │ │ │ │ In [143]: %timeit arr[indexer]
│ │ │ │ │ .....: %timeit arr.take(indexer, axis=0)
│ │ │ │ │ .....:
│ │ │ │ │ -209 us +- 324 ns per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ -77.8 us +- 231 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
│ │ │ │ │ +698 us +- 84.8 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ +210 us +- 43 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ In [144]: ser = pd.Series(arr[:, 0])
│ │ │ │ │
│ │ │ │ │ In [145]: %timeit ser.iloc[indexer]
│ │ │ │ │ .....: %timeit ser.take(indexer)
│ │ │ │ │ .....:
│ │ │ │ │ -147 us +- 483 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
│ │ │ │ │ -129 us +- 433 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
│ │ │ │ │ +501 us +- 70.5 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ +388 us +- 86.5 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ ********** IInnddeexx ttyyppeess_?¶ **********
│ │ │ │ │ We have discussed MultiIndex in the previous sections pretty extensively.
│ │ │ │ │ Documentation about DatetimeIndex and PeriodIndex are shown _h_e_r_e, and
│ │ │ │ │ documentation about TimedeltaIndex is found _h_e_r_e.
│ │ │ │ │ In the following sub-sections we will highlight some other index types.
│ │ │ │ │ ******** CCaatteeggoorriiccaallIInnddeexx_?¶ ********
│ │ │ │ │ _C_a_t_e_g_o_r_i_c_a_l_I_n_d_e_x is a type of index that is useful for supporting indexing with
│ │ │ ├── ./usr/share/doc/python-pandas-doc/html/user_guide/enhancingperf.html
│ │ │ │ @@ -529,31 +529,31 @@
│ │ │ │ ...: s += f(a + i * dx)
│ │ │ │ ...: return s * dx
│ │ │ │ ...:
│ │ │ │
│ │ │ │
│ │ │ │
We achieve our result by using DataFrame.apply() (row-wise):
In [5]: %timeit df.apply(lambda x: integrate_f(x["a"], x["b"], x["N"]), axis=1)
│ │ │ │ -97.1 ms +- 271 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ +150 ms +- 38.7 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ Let’s take a look and see where the time is spent during this operation │ │ │ │ using the prun ipython magic function:
│ │ │ │# most time consuming 4 calls
│ │ │ │ In [6]: %prun -l 4 df.apply(lambda x: integrate_f(x["a"], x["b"], x["N"]), axis=1) # noqa E999
│ │ │ │ - 605963 function calls (605945 primitive calls) in 0.304 seconds
│ │ │ │ + 605963 function calls (605945 primitive calls) in 0.318 seconds
│ │ │ │
│ │ │ │ Ordered by: internal time
│ │ │ │ List reduced from 167 to 4 due to restriction <4>
│ │ │ │
│ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ - 1000 0.196 0.000 0.276 0.000 <ipython-input-4-c2a74e076cf0>:1(integrate_f)
│ │ │ │ - 552423 0.080 0.000 0.080 0.000 <ipython-input-3-c138bdd570e3>:1(f)
│ │ │ │ - 3000 0.005 0.000 0.019 0.000 series.py:1016(__getitem__)
│ │ │ │ - 3000 0.003 0.000 0.008 0.000 series.py:1139(_get_value)
│ │ │ │ + 1000 0.206 0.000 0.288 0.000 <ipython-input-4-c2a74e076cf0>:1(integrate_f)
│ │ │ │ + 552423 0.082 0.000 0.082 0.000 <ipython-input-3-c138bdd570e3>:1(f)
│ │ │ │ + 3000 0.005 0.000 0.020 0.000 series.py:1016(__getitem__)
│ │ │ │ + 3000 0.003 0.000 0.009 0.000 series.py:1139(_get_value)
│ │ │ │ By far the majority of time is spend inside either integrate_f or f,
│ │ │ │ hence we’ll concentrate our efforts cythonizing these two functions.
In [9]: %timeit df.apply(lambda x: integrate_f_plain(x["a"], x["b"], x["N"]), axis=1)
│ │ │ │ -87.4 ms +- 70.6 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ +143 ms +- 36 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ This has improved the performance compared to the pure Python approach by one-third.
│ │ │ │We can annotate the function variables and return types as well as use cdef
│ │ │ │ @@ -595,36 +595,36 @@
│ │ │ │ ....: for i in range(N):
│ │ │ │ ....: s += f_typed(a + i * dx)
│ │ │ │ ....: return s * dx
│ │ │ │ ....:
│ │ │ │
│ │ │ │
│ │ │ │
In [11]: %timeit df.apply(lambda x: integrate_f_typed(x["a"], x["b"], x["N"]), axis=1)
│ │ │ │ -10.1 ms +- 12.8 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +15.7 ms +- 4.25 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ Annotating the functions with C types yields an over ten times performance improvement compared to │ │ │ │ the original Python implementation.
│ │ │ │When re-profiling, time is spent creating a Series from each row, and calling __getitem__ from both
│ │ │ │ the index and the series (three times for each row). These Python function calls are expensive and
│ │ │ │ can be improved by passing an np.ndarray.
In [12]: %prun -l 4 df.apply(lambda x: integrate_f_typed(x["a"], x["b"], x["N"]), axis=1)
│ │ │ │ - 52540 function calls (52522 primitive calls) in 0.028 seconds
│ │ │ │ + 52540 function calls (52522 primitive calls) in 0.058 seconds
│ │ │ │
│ │ │ │ Ordered by: internal time
│ │ │ │ List reduced from 165 to 4 due to restriction <4>
│ │ │ │
│ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ - 3000 0.005 0.000 0.018 0.000 series.py:1016(__getitem__)
│ │ │ │ - 3000 0.003 0.000 0.008 0.000 series.py:1139(_get_value)
│ │ │ │ - 3000 0.003 0.000 0.004 0.000 indexing.py:2678(check_dict_or_set_indexers)
│ │ │ │ - 3000 0.002 0.000 0.003 0.000 base.py:3763(get_loc)
│ │ │ │ + 3000 0.011 0.000 0.016 0.000 indexing.py:2678(check_dict_or_set_indexers)
│ │ │ │ + 1000 0.010 0.000 0.049 0.000 <string>:1(<lambda>)
│ │ │ │ + 3000 0.007 0.000 0.017 0.000 series.py:1139(_get_value)
│ │ │ │ + 3000 0.007 0.000 0.007 0.000 base.py:3763(get_loc)
│ │ │ │ In [13]: %%cython
│ │ │ │ ....: cimport numpy as np
│ │ │ │ ....: import numpy as np
│ │ │ │ ....: cdef double f_typed(double x) except? -2:
│ │ │ │ ....: return x * (x - 1)
│ │ │ │ @@ -659,15 +659,15 @@
│ │ │ │ This implementation creates an array of zeros and inserts the result
│ │ │ │ of integrate_f_typed applied over each row. Looping over an ndarray is faster
│ │ │ │ in Cython than looping over a Series object.
Since apply_integrate_f is typed to accept an np.ndarray, Series.to_numpy()
│ │ │ │ calls are needed to utilize this function.
In [14]: %timeit apply_integrate_f(df["a"].to_numpy(), df["b"].to_numpy(), df["N"].to_numpy())
│ │ │ │ -1.3 ms +- 114 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ +2.33 ms +- 495 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ Performance has improved from the prior implementation by almost ten times.
│ │ │ │ │ │ │ │The majority of the time is now spent in apply_integrate_f. Disabling Cython’s boundscheck
│ │ │ │ @@ -678,15 +678,15 @@
│ │ │ │ Ordered by: internal time
│ │ │ │ List reduced from 20 to 4 due to restriction <4>
│ │ │ │
│ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ 1 0.001 0.001 0.001 0.001 <string>:1(<module>)
│ │ │ │ 1 0.000 0.000 0.001 0.001 {built-in method builtins.exec}
│ │ │ │ 3 0.000 0.000 0.000 0.000 frame.py:3853(__getitem__)
│ │ │ │ - 3 0.000 0.000 0.000 0.000 common.py:367(apply_if_callable)
│ │ │ │ + 3 0.000 0.000 0.000 0.000 base.py:541(to_numpy)
│ │ │ │
In [16]: %%cython
│ │ │ │ ....: cimport cython
│ │ │ │ ....: cimport numpy as np
│ │ │ │ ....: import numpy as np
│ │ │ │ ....: cdef np.float64_t f_typed(np.float64_t x) except? -2:
│ │ │ │ @@ -719,15 +719,15 @@
│ │ │ │ from /build/reproducible-path/pandas-2.1.4+dfsg/buildtmp/.cache/ipython/cython/_cython_magic_dbd2f2be48d2a3825f8cbd16b561de5210461988.c:1251:
│ │ │ │ /usr/lib/python3/dist-packages/numpy/core/include/numpy/npy_1_7_deprecated_api.h:17:2: warning: #warning "Using deprecated NumPy API, disable it with " "#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" [-Wcpp]
│ │ │ │ 17 | #warning "Using deprecated NumPy API, disable it with " \
│ │ │ │ | ^~~~~~~
│ │ │ │ In [17]: %timeit apply_integrate_f_wrap(df["a"].to_numpy(), df["b"].to_numpy(), df["N"].to_numpy())
│ │ │ │ -1.13 ms +- 238 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ +1.26 ms +- 286 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ However, a loop indexer i accessing an invalid location in an array would cause a segfault because memory access isn’t checked.
│ │ │ │ For more about boundscheck and wraparound, see the Cython docs on
│ │ │ │ compiler directives.
DataFrame. This engine requires the
│ │ │ │ optional dependency numexpr to be installed.
│ │ │ │ The 'python' engine is generally not useful except for testing
│ │ │ │ other evaluation engines against it. You will achieve no performance
│ │ │ │ benefits using eval() with engine='python' and may
│ │ │ │ incur a performance hit.
In [40]: %timeit df1 + df2 + df3 + df4
│ │ │ │ -19.5 ms +- 5.81 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +25.2 ms +- 1.55 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ In [41]: %timeit pd.eval("df1 + df2 + df3 + df4", engine="python")
│ │ │ │ -19 ms +- 3.08 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +28.7 ms +- 1.76 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ DataFrame.eval() method¶In addition to the top level pandas.eval() function you can also
│ │ │ │ evaluate an expression in the “context” of a DataFrame.
In [58]: nrows, ncols = 20000, 100
│ │ │ │
│ │ │ │ In [59]: df1, df2, df3, df4 = [pd.DataFrame(np.random.randn(nrows, ncols)) for _ in range(4)]
│ │ │ │ DataFrame arithmetic:
In [60]: %timeit df1 + df2 + df3 + df4
│ │ │ │ -16.3 ms +- 4.91 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +26.2 ms +- 987 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ In [61]: %timeit pd.eval("df1 + df2 + df3 + df4")
│ │ │ │ -12.4 ms +- 1.4 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +12.7 ms +- 399 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ DataFrame comparison:
In [62]: %timeit (df1 > 0) & (df2 > 0) & (df3 > 0) & (df4 > 0)
│ │ │ │ -23.3 ms +- 6.91 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ +37.8 ms +- 3.11 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ In [63]: %timeit pd.eval("(df1 > 0) & (df2 > 0) & (df3 > 0) & (df4 > 0)")
│ │ │ │ -21.9 ms +- 1.46 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ +19.7 ms +- 578 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ DataFrame arithmetic with unaligned axes.
In [64]: s = pd.Series(np.random.randn(50))
│ │ │ │
│ │ │ │ In [65]: %timeit df1 + df2 + df3 + df4 + s
│ │ │ │ -34.7 ms +- 2.36 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ +43.9 ms +- 2.3 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ In [66]: %timeit pd.eval("df1 + df2 + df3 + df4 + s")
│ │ │ │ -15.7 ms +- 2.67 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ +14.1 ms +- 538 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ Note
│ │ │ │Operations such as
│ │ │ │1 and 2 # would parse to 1 & 2, but should evaluate to 2
│ │ │ │ 3 or 4 # would parse to 3 | 4, but should evaluate to 3
│ │ │ │ ├── html2text {}
│ │ │ │ │ @@ -113,32 +113,32 @@
│ │ │ │ │ ...: dx = (b - a) / N
│ │ │ │ │ ...: for i in range(N):
│ │ │ │ │ ...: s += f(a + i * dx)
│ │ │ │ │ ...: return s * dx
│ │ │ │ │ ...:
│ │ │ │ │ We achieve our result by using _D_a_t_a_F_r_a_m_e_._a_p_p_l_y_(_) (row-wise):
│ │ │ │ │ In [5]: %timeit df.apply(lambda x: integrate_f(x["a"], x["b"], x["N"]), axis=1)
│ │ │ │ │ -97.1 ms +- 271 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ +150 ms +- 38.7 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ Let’s take a look and see where the time is spent during this operation using
│ │ │ │ │ the _p_r_u_n_ _i_p_y_t_h_o_n_ _m_a_g_i_c_ _f_u_n_c_t_i_o_n:
│ │ │ │ │ # most time consuming 4 calls
│ │ │ │ │ In [6]: %prun -l 4 df.apply(lambda x: integrate_f(x["a"], x["b"], x["N"]),
│ │ │ │ │ axis=1) # noqa E999
│ │ │ │ │ - 605963 function calls (605945 primitive calls) in 0.304 seconds
│ │ │ │ │ + 605963 function calls (605945 primitive calls) in 0.318 seconds
│ │ │ │ │
│ │ │ │ │ Ordered by: internal time
│ │ │ │ │ List reduced from 167 to 4 due to restriction <4>
│ │ │ │ │
│ │ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ │ - 1000 0.196 0.000 0.276 0.000 :1
│ │ │ │ │ + 1000 0.206 0.000 0.288 0.000 :1
│ │ │ │ │ (integrate_f)
│ │ │ │ │ - 552423 0.080 0.000 0.080 0.000 :1
│ │ │ │ │ + 552423 0.082 0.000 0.082 0.000 :1
│ │ │ │ │ (f)
│ │ │ │ │ - 3000 0.005 0.000 0.019 0.000 series.py:1016(__getitem__)
│ │ │ │ │ - 3000 0.003 0.000 0.008 0.000 series.py:1139(_get_value)
│ │ │ │ │ + 3000 0.005 0.000 0.020 0.000 series.py:1016(__getitem__)
│ │ │ │ │ + 3000 0.003 0.000 0.009 0.000 series.py:1139(_get_value)
│ │ │ │ │ By far the majority of time is spend inside either integrate_f or f, hence
│ │ │ │ │ we’ll concentrate our efforts cythonizing these two functions.
│ │ │ │ │ ******** PPllaaiinn CCyytthhoonn_?¶ ********
│ │ │ │ │ First we’re going to need to import the Cython magic function to IPython:
│ │ │ │ │ In [7]: %load_ext Cython
│ │ │ │ │ Now, let’s simply copy our functions over to Cython:
│ │ │ │ │ In [8]: %%cython
│ │ │ │ │ @@ -149,15 +149,15 @@
│ │ │ │ │ ...: dx = (b - a) / N
│ │ │ │ │ ...: for i in range(N):
│ │ │ │ │ ...: s += f_plain(a + i * dx)
│ │ │ │ │ ...: return s * dx
│ │ │ │ │ ...:
│ │ │ │ │ In [9]: %timeit df.apply(lambda x: integrate_f_plain(x["a"], x["b"], x["N"]),
│ │ │ │ │ axis=1)
│ │ │ │ │ -87.4 ms +- 70.6 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ +143 ms +- 36 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ This has improved the performance compared to the pure Python approach by one-
│ │ │ │ │ third.
│ │ │ │ │ ******** DDeeccllaarriinngg CC ttyyppeess_?¶ ********
│ │ │ │ │ We can annotate the function variables and return types as well as use cdef and
│ │ │ │ │ cpdef to improve performance:
│ │ │ │ │ In [10]: %%cython
│ │ │ │ │ ....: cdef double f_typed(double x) except? -2:
│ │ │ │ │ @@ -169,35 +169,35 @@
│ │ │ │ │ ....: dx = (b - a) / N
│ │ │ │ │ ....: for i in range(N):
│ │ │ │ │ ....: s += f_typed(a + i * dx)
│ │ │ │ │ ....: return s * dx
│ │ │ │ │ ....:
│ │ │ │ │ In [11]: %timeit df.apply(lambda x: integrate_f_typed(x["a"], x["b"], x["N"]),
│ │ │ │ │ axis=1)
│ │ │ │ │ -10.1 ms +- 12.8 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +15.7 ms +- 4.25 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ Annotating the functions with C types yields an over ten times performance
│ │ │ │ │ improvement compared to the original Python implementation.
│ │ │ │ │ ******** UUssiinngg nnddaarrrraayy_?¶ ********
│ │ │ │ │ When re-profiling, time is spent creating a _S_e_r_i_e_s from each row, and calling
│ │ │ │ │ __getitem__ from both the index and the series (three times for each row).
│ │ │ │ │ These Python function calls are expensive and can be improved by passing an
│ │ │ │ │ np.ndarray.
│ │ │ │ │ In [12]: %prun -l 4 df.apply(lambda x: integrate_f_typed(x["a"], x["b"], x
│ │ │ │ │ ["N"]), axis=1)
│ │ │ │ │ - 52540 function calls (52522 primitive calls) in 0.028 seconds
│ │ │ │ │ + 52540 function calls (52522 primitive calls) in 0.058 seconds
│ │ │ │ │
│ │ │ │ │ Ordered by: internal time
│ │ │ │ │ List reduced from 165 to 4 due to restriction <4>
│ │ │ │ │
│ │ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ │ - 3000 0.005 0.000 0.018 0.000 series.py:1016(__getitem__)
│ │ │ │ │ - 3000 0.003 0.000 0.008 0.000 series.py:1139(_get_value)
│ │ │ │ │ - 3000 0.003 0.000 0.004 0.000 indexing.py:2678
│ │ │ │ │ + 3000 0.011 0.000 0.016 0.000 indexing.py:2678
│ │ │ │ │ (check_dict_or_set_indexers)
│ │ │ │ │ - 3000 0.002 0.000 0.003 0.000 base.py:3763(get_loc)
│ │ │ │ │ + 1000 0.010 0.000 0.049 0.000 :1()
│ │ │ │ │ + 3000 0.007 0.000 0.017 0.000 series.py:1139(_get_value)
│ │ │ │ │ + 3000 0.007 0.000 0.007 0.000 base.py:3763(get_loc)
│ │ │ │ │ In [13]: %%cython
│ │ │ │ │ ....: cimport numpy as np
│ │ │ │ │ ....: import numpy as np
│ │ │ │ │ ....: cdef double f_typed(double x) except? -2:
│ │ │ │ │ ....: return x * (x - 1)
│ │ │ │ │ ....: cpdef double integrate_f_typed(double a, double b, int N):
│ │ │ │ │ ....: cdef int i
│ │ │ │ │ @@ -238,15 +238,15 @@
│ │ │ │ │ This implementation creates an array of zeros and inserts the result of
│ │ │ │ │ integrate_f_typed applied over each row. Looping over an ndarray is faster in
│ │ │ │ │ Cython than looping over a _S_e_r_i_e_s object.
│ │ │ │ │ Since apply_integrate_f is typed to accept an np.ndarray, _S_e_r_i_e_s_._t_o___n_u_m_p_y_(_)
│ │ │ │ │ calls are needed to utilize this function.
│ │ │ │ │ In [14]: %timeit apply_integrate_f(df["a"].to_numpy(), df["b"].to_numpy(), df
│ │ │ │ │ ["N"].to_numpy())
│ │ │ │ │ -1.3 ms +- 114 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ +2.33 ms +- 495 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ Performance has improved from the prior implementation by almost ten times.
│ │ │ │ │ ******** DDiissaabblliinngg ccoommppiilleerr ddiirreeccttiivveess_?¶ ********
│ │ │ │ │ The majority of the time is now spent in apply_integrate_f. Disabling Cython’s
│ │ │ │ │ boundscheck and wraparound checks can yield more performance.
│ │ │ │ │ In [15]: %prun -l 4 apply_integrate_f(df["a"].to_numpy(), df["b"].to_numpy(),
│ │ │ │ │ df["N"].to_numpy())
│ │ │ │ │ 75 function calls in 0.001 seconds
│ │ │ │ │ @@ -254,15 +254,15 @@
│ │ │ │ │ Ordered by: internal time
│ │ │ │ │ List reduced from 20 to 4 due to restriction <4>
│ │ │ │ │
│ │ │ │ │ ncalls tottime percall cumtime percall filename:lineno(function)
│ │ │ │ │ 1 0.001 0.001 0.001 0.001 :1()
│ │ │ │ │ 1 0.000 0.000 0.001 0.001 {built-in method builtins.exec}
│ │ │ │ │ 3 0.000 0.000 0.000 0.000 frame.py:3853(__getitem__)
│ │ │ │ │ - 3 0.000 0.000 0.000 0.000 common.py:367(apply_if_callable)
│ │ │ │ │ + 3 0.000 0.000 0.000 0.000 base.py:541(to_numpy)
│ │ │ │ │ In [16]: %%cython
│ │ │ │ │ ....: cimport cython
│ │ │ │ │ ....: cimport numpy as np
│ │ │ │ │ ....: import numpy as np
│ │ │ │ │ ....: cdef np.float64_t f_typed(np.float64_t x) except? -2:
│ │ │ │ │ ....: return x * (x - 1)
│ │ │ │ │ ....: cpdef np.float64_t integrate_f_typed(np.float64_t a, np.float64_t b,
│ │ │ │ │ @@ -300,15 +300,15 @@
│ │ │ │ │ /usr/lib/python3/dist-packages/numpy/core/include/numpy/
│ │ │ │ │ npy_1_7_deprecated_api.h:17:2: warning: #warning "Using deprecated NumPy API,
│ │ │ │ │ disable it with " "#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" [-Wcpp]
│ │ │ │ │ 17 | #warning "Using deprecated NumPy API, disable it with " \
│ │ │ │ │ | ^~~~~~~
│ │ │ │ │ In [17]: %timeit apply_integrate_f_wrap(df["a"].to_numpy(), df["b"].to_numpy(),
│ │ │ │ │ df["N"].to_numpy())
│ │ │ │ │ -1.13 ms +- 238 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ +1.26 ms +- 286 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
│ │ │ │ │ However, a loop indexer i accessing an invalid location in an array would cause
│ │ │ │ │ a segfault because memory access isn’t checked. For more about boundscheck and
│ │ │ │ │ wraparound, see the Cython docs on _c_o_m_p_i_l_e_r_ _d_i_r_e_c_t_i_v_e_s.
│ │ │ │ │ ********** NNuummbbaa ((JJIITT ccoommppiillaattiioonn))_?¶ **********
│ │ │ │ │ An alternative to statically compiling Cython code is to use a dynamic just-in-
│ │ │ │ │ time (JIT) compiler with _N_u_m_b_a.
│ │ │ │ │ Numba allows you to write a pure Python function which can be JIT compiled to
│ │ │ │ │ @@ -611,17 +611,17 @@
│ │ │ │ │ The 'numexpr' engine is the more performant engine that can yield performance
│ │ │ │ │ improvements compared to standard Python syntax for large _D_a_t_a_F_r_a_m_e. This
│ │ │ │ │ engine requires the optional dependency numexpr to be installed.
│ │ │ │ │ The 'python' engine is generally nnoott useful except for testing other evaluation
│ │ │ │ │ engines against it. You will achieve nnoo performance benefits using _e_v_a_l_(_) with
│ │ │ │ │ engine='python' and may incur a performance hit.
│ │ │ │ │ In [40]: %timeit df1 + df2 + df3 + df4
│ │ │ │ │ -19.5 ms +- 5.81 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +25.2 ms +- 1.55 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ In [41]: %timeit pd.eval("df1 + df2 + df3 + df4", engine="python")
│ │ │ │ │ -19 ms +- 3.08 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +28.7 ms +- 1.76 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ ******** TThhee _DD_aa_tt_aa_FF_rr_aa_mm_ee_.._ee_vv_aa_ll_((_)) mmeetthhoodd_?¶ ********
│ │ │ │ │ In addition to the top level _p_a_n_d_a_s_._e_v_a_l_(_) function you can also evaluate an
│ │ │ │ │ expression in the “context” of a _D_a_t_a_F_r_a_m_e.
│ │ │ │ │ In [42]: df = pd.DataFrame(np.random.randn(5, 2), columns=["a", "b"])
│ │ │ │ │
│ │ │ │ │ In [43]: df.eval("a + b")
│ │ │ │ │ Out[43]:
│ │ │ │ │ @@ -718,29 +718,29 @@
│ │ │ │ │ _p_a_n_d_a_s_._e_v_a_l_(_) works well with expressions containing large arrays.
│ │ │ │ │ In [58]: nrows, ncols = 20000, 100
│ │ │ │ │
│ │ │ │ │ In [59]: df1, df2, df3, df4 = [pd.DataFrame(np.random.randn(nrows, ncols)) for
│ │ │ │ │ _ in range(4)]
│ │ │ │ │ _D_a_t_a_F_r_a_m_e arithmetic:
│ │ │ │ │ In [60]: %timeit df1 + df2 + df3 + df4
│ │ │ │ │ -16.3 ms +- 4.91 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +26.2 ms +- 987 us per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ In [61]: %timeit pd.eval("df1 + df2 + df3 + df4")
│ │ │ │ │ -12.4 ms +- 1.4 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +12.7 ms +- 399 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ _D_a_t_a_F_r_a_m_e comparison:
│ │ │ │ │ In [62]: %timeit (df1 > 0) & (df2 > 0) & (df3 > 0) & (df4 > 0)
│ │ │ │ │ -23.3 ms +- 6.91 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ +37.8 ms +- 3.11 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ In [63]: %timeit pd.eval("(df1 > 0) & (df2 > 0) & (df3 > 0) & (df4 > 0)")
│ │ │ │ │ -21.9 ms +- 1.46 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ +19.7 ms +- 578 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ _D_a_t_a_F_r_a_m_e arithmetic with unaligned axes.
│ │ │ │ │ In [64]: s = pd.Series(np.random.randn(50))
│ │ │ │ │
│ │ │ │ │ In [65]: %timeit df1 + df2 + df3 + df4 + s
│ │ │ │ │ -34.7 ms +- 2.36 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ +43.9 ms +- 2.3 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
│ │ │ │ │ In [66]: %timeit pd.eval("df1 + df2 + df3 + df4 + s")
│ │ │ │ │ -15.7 ms +- 2.67 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ +14.1 ms +- 538 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
│ │ │ │ │ Note
│ │ │ │ │ Operations such as
│ │ │ │ │ 1 and 2 # would parse to 1 & 2, but should evaluate to 2
│ │ │ │ │ 3 or 4 # would parse to 3 | 4, but should evaluate to 3
│ │ │ │ │ ~1 # this is okay, but slower when using eval
│ │ │ │ │ should be performed in Python. An exception will be raised if you try to
│ │ │ │ │ perform any boolean/bitwise operations with scalar operands that are not of
│ │ │ ├── ./usr/share/doc/python-pandas-doc/html/user_guide/scale.html
│ │ │ │ @@ -889,16 +889,16 @@
│ │ │ │ ....: files = pathlib.Path("data/timeseries/").glob("ts*.parquet")
│ │ │ │ ....: counts = pd.Series(dtype=int)
│ │ │ │ ....: for path in files:
│ │ │ │ ....: df = pd.read_parquet(path)
│ │ │ │ ....: counts = counts.add(df["name"].value_counts(), fill_value=0)
│ │ │ │ ....: counts.astype(int)
│ │ │ │ ....:
│ │ │ │ -CPU times: user 691 us, sys: 0 ns, total: 691 us
│ │ │ │ -Wall time: 702 us
│ │ │ │ +CPU times: user 604 us, sys: 35 us, total: 639 us
│ │ │ │ +Wall time: 646 us
│ │ │ │ Out[32]: Series([], dtype: int64)
│ │ │ │ Some readers, like pandas.read_csv(), offer parameters to control the
│ │ │ │ chunksize when reading a single file.
Manually chunking is an OK option for workflows that don’t
│ │ │ │ require too sophisticated of operations. Some operations, like pandas.DataFrame.groupby(), are
│ │ │ │ ├── html2text {}
│ │ │ │ │ @@ -600,16 +600,16 @@
│ │ │ │ │ ....: files = pathlib.Path("data/timeseries/").glob("ts*.parquet")
│ │ │ │ │ ....: counts = pd.Series(dtype=int)
│ │ │ │ │ ....: for path in files:
│ │ │ │ │ ....: df = pd.read_parquet(path)
│ │ │ │ │ ....: counts = counts.add(df["name"].value_counts(), fill_value=0)
│ │ │ │ │ ....: counts.astype(int)
│ │ │ │ │ ....:
│ │ │ │ │ -CPU times: user 691 us, sys: 0 ns, total: 691 us
│ │ │ │ │ -Wall time: 702 us
│ │ │ │ │ +CPU times: user 604 us, sys: 35 us, total: 639 us
│ │ │ │ │ +Wall time: 646 us
│ │ │ │ │ Out[32]: Series([], dtype: int64)
│ │ │ │ │ Some readers, like _p_a_n_d_a_s_._r_e_a_d___c_s_v_(_), offer parameters to control the chunksize
│ │ │ │ │ when reading a single file.
│ │ │ │ │ Manually chunking is an OK option for workflows that don’t require too
│ │ │ │ │ sophisticated of operations. Some operations, like _p_a_n_d_a_s_._D_a_t_a_F_r_a_m_e_._g_r_o_u_p_b_y_(_),
│ │ │ │ │ are much harder to do chunkwise. In these cases, you may be better switching to
│ │ │ │ │ a different library that implements these out-of-core algorithms for you.
│ │ │ ├── ./usr/share/doc/python-pandas-doc/html/user_guide/style.ipynb.gz
│ │ │ │ ├── style.ipynb
│ │ │ │ │ ├── Pretty-printed
│ │ │ │ │ │┄ Similarity: 0.9985610875706213%
│ │ │ │ │ │┄ Differences: {"'cells'": "{1: {'metadata': {'execution': {'iopub.execute_input': '2024-05-26T11:39:24.088609Z', "
│ │ │ │ │ │┄ "'iopub.status.busy': '2024-05-26T11:39:24.088289Z', 'iopub.status.idle': "
│ │ │ │ │ │┄ "'2024-05-26T11:39:24.711465Z', 'shell.execute_reply': "
│ │ │ │ │ │┄ "'2024-05-26T11:39:24.698031Z'}}}, 3: {'metadata': {'execution': "
│ │ │ │ │ │┄ "{'iopub.execute_input': '2024-05-26T11:39:24.720678Z', 'iopub.status.busy': "
│ │ │ │ │ │┄ "'2024-05-26T11:39:24.720278Z', 'iopub.status.idle': '2024-05-26T11:39:2 […]
│ │ │ │ │ │ @@ -39,18 +39,18 @@
│ │ │ │ │ │ ]
│ │ │ │ │ │ },
│ │ │ │ │ │ {
│ │ │ │ │ │ "cell_type": "code",
│ │ │ │ │ │ "execution_count": 1,
│ │ │ │ │ │ "metadata": {
│ │ │ │ │ │ "execution": {
│ │ │ │ │ │ - "iopub.execute_input": "2025-06-28T15:31:40.161290Z",
│ │ │ │ │ │ - "iopub.status.busy": "2025-06-28T15:31:40.160976Z",
│ │ │ │ │ │ - "iopub.status.idle": "2025-06-28T15:31:40.689868Z",
│ │ │ │ │ │ - "shell.execute_reply": "2025-06-28T15:31:40.689028Z"
│ │ │ │ │ │ + "iopub.execute_input": "2024-05-26T11:39:24.088609Z",
│ │ │ │ │ │ + "iopub.status.busy": "2024-05-26T11:39:24.088289Z",
│ │ │ │ │ │ + "iopub.status.idle": "2024-05-26T11:39:24.711465Z",
│ │ │ │ │ │ + "shell.execute_reply": "2024-05-26T11:39:24.698031Z"
│ │ │ │ │ │ },
│ │ │ │ │ │ "nbsphinx": "hidden"
│ │ │ │ │ │ },
│ │ │ │ │ │ "outputs": [],
│ │ │ │ │ │ "source": [
│ │ │ │ │ │ "import matplotlib.pyplot\n",
│ │ │ │ │ │ "# We have this here to trigger matplotlib's font cache stuff.\n",
│ │ │ │ │ │ @@ -77,36 +77,36 @@
│ │ │ │ │ │ ]
│ │ │ │ │ │ },
│ │ │ │ │ │ {
│ │ │ │ │ │ "cell_type": "code",
│ │ │ │ │ │ "execution_count": 2,
│ │ │ │ │ │ "metadata": {
│ │ │ │ │ │ "execution": {
│ │ │ │ │ │ - "iopub.execute_input": "2025-06-28T15:31:40.698338Z",
│ │ │ │ │ │ - "iopub.status.busy": "2025-06-28T15:31:40.697895Z",
│ │ │ │ │ │ - "iopub.status.idle": "2025-06-28T15:31:40.956094Z",
│ │ │ │ │ │ - "shell.execute_reply": "2025-06-28T15:31:40.954947Z"
│ │ │ │ │ │ + "iopub.execute_input": "2024-05-26T11:39:24.720678Z",
│ │ │ │ │ │ + "iopub.status.busy": "2024-05-26T11:39:24.720278Z",
│ │ │ │ │ │ + "iopub.status.idle": "2024-05-26T11:39:25.235484Z",
│ │ │ │ │ │ + "shell.execute_reply": "2024-05-26T11:39:25.219471Z"
│ │ │ │ │ │ }
│ │ │ │ │ │ },
│ │ │ │ │ │ "outputs": [],
│ │ │ │ │ │ "source": [
│ │ │ │ │ │ "import pandas as pd\n",
│ │ │ │ │ │ "import numpy as np\n",
│ │ │ │ │ │ "import matplotlib as mpl\n"
│ │ │ │ │ │ ]
│ │ │ │ │ │ },
│ │ │ │ │ │ {
│ │ │ │ │ │ "cell_type": "code",
│ │ │ │ │ │ "execution_count": 3,
│ │ │ │ │ │ "metadata": {
│ │ │ │ │ │ "execution": {
│ │ │ │ │ │ - "iopub.execute_input": "2025-06-28T15:31:40.960392Z",
│ │ │ │ │ │ - "iopub.status.busy": "2025-06-28T15:31:40.959999Z",
│ │ │ │ │ │ - "iopub.status.idle": "2025-06-28T15:31:41.101927Z",
│ │ │ │ │ │ - "shell.execute_reply": "2025-06-28T15:31:41.100736Z"
│ │ │ │ │ │ + "iopub.execute_input": "2024-05-26T11:39:25.244647Z",
│ │ │ │ │ │ + "iopub.status.busy": "2024-05-26T11:39:25.244230Z",
│ │ │ │ │ │ + "iopub.status.idle": "2024-05-26T11:39:25.536260Z",
│ │ │ │ │ │ + "shell.execute_reply": "2024-05-26T11:39:25.535440Z"
│ │ │ │ │ │ },
│ │ │ │ │ │ "nbsphinx": "hidden"
│ │ │ │ │ │ },
│ │ │ │ │ │ "outputs": [],
│ │ │ │ │ │ "source": [
│ │ │ │ │ │ "# For reproducibility - this doesn't respect uuid_len or positionally-passed uuid but the places here that use that coincidentally bypass this anyway\n",
│ │ │ │ │ │ "from pandas.io.formats.style import Styler\n",
│ │ │ │ │ │ @@ -123,18 +123,18 @@
│ │ │ │ │ │ ]
│ │ │ │ │ │ },
│ │ │ │ │ │ {
│ │ │ │ │ │ "cell_type": "code",
│ │ │ │ │ │ "execution_count": 4,
│ │ │ │ │ │ "metadata": {
│ │ │ │ │ │ "execution": {
│ │ │ │ │ │ - "iopub.execute_input": "2025-06-28T15:31:41.106060Z",
│ │ │ │ │ │ - "iopub.status.busy": "2025-06-28T15:31:41.105636Z",
│ │ │ │ │ │ - "iopub.status.idle": "2025-06-28T15:31:41.125830Z",
│ │ │ │ │ │ - "shell.execute_reply": "2025-06-28T15:31:41.125046Z"
│ │ │ │ │ │ + "iopub.execute_input": "2024-05-26T11:39:25.544612Z",
│ │ │ │ │ │ + "iopub.status.busy": "2024-05-26T11:39:25.544212Z",
│ │ │ │ │ │ + "iopub.status.idle": "2024-05-26T11:39:25.568236Z",
│ │ │ │ │ │ + "shell.execute_reply": "2024-05-26T11:39:25.567449Z"
│ │ │ │ │ │ }
│ │ │ │ │ │ },
│ │ │ │ │ │ "outputs": [
│ │ │ │ │ │ {
│ │ │ │ │ │ "data": {
│ │ │ │ │ │ "text/html": [
│ │ │ │ │ │ "