Consider the following lists short_list and long_list
short_list = list('aaabaaacaaadaaac')
np.random.seed([3,1415])
long_list = pd.DataFrame(
np.random.choice(list(ascii_letters),
(10000, 2))
).sum(1).tolist()
How do I calculate the cumulative count by unique value?
I want to use numpy and do it in linear time. I want this to compare timings with my other methods. It may be easiest to illustrate with my first proposed solution
def pir1(l):
s = pd.Series(l)
return s.groupby(s).cumcount().tolist()
print(np.array(short_list))
print(pir1(short_list))
['a' 'a' 'a' 'b' 'a' 'a' 'a' 'c' 'a' 'a' 'a' 'd' 'a' 'a' 'a' 'c']
[0, 1, 2, 0, 3, 4, 5, 0, 6, 7, 8, 0, 9, 10, 11, 1]
I've tortured myself trying to use np.unique because it returns a counts array, an inverse array, and an index array. I was sure I could these to get at a solution. The best I got is in pir4 below which scales in quadratic time. Also note that I don't care if counts start at 1 or zero as we can simply add or subtract 1.
Below are some of my attempts (none of which answer my question)
%%cython
from collections import defaultdict
def get_generator(l):
counter = defaultdict(lambda: -1)
for i in l:
counter[i] += 1
yield counter[i]
def pir2(l):
return [i for i in get_generator(l)]
def pir3(l):
return [i for i in get_generator(l)]
def pir4(l):
unq, inv = np.unique(l, 0, 1, 0)
a = np.arange(len(unq))
matches = a[:, None] == inv
return (matches * matches.cumsum(1)).sum(0).tolist()
setup
short_list = np.array(list('aaabaaacaaadaaac'))
functions
dfill takes an array and returns the positions where the array changes and repeats that index position until the next change.
# dfill
#
# Example with short_list
#
# 0 0 0 3 4 4 4 7 8 8 8 11 12 12 12 15
# [ a a a b a a a c a a a d a a a c]
#
# Example with short_list after sorting
#
# 0 0 0 0 0 0 0 0 0 0 0 0 12 13 13 15
# [ a a a a a a a a a a a a b c c d]
argunsort returns the permutation necessary to undo a sort given the argsort array. The existence of this method became know to me via this post.. With this, I can get the argsort array and sort my array with it. Then I can undo the sort without the overhead of sorting again.
cumcount will take an array sort it, find the dfill array. An np.arange less dfill will give me cumulative count. Then I un-sort
# cumcount
#
# Example with short_list
#
# short_list:
# [ a a a b a a a c a a a d a a a c]
#
# short_list.argsort():
# [ 0 1 2 4 5 6 8 9 10 12 13 14 3 7 15 11]
#
# Example with short_list after sorting
#
# short_list[short_list.argsort()]:
# [ a a a a a a a a a a a a b c c d]
#
# dfill(short_list[short_list.argsort()]):
# [ 0 0 0 0 0 0 0 0 0 0 0 0 12 13 13 15]
#
# np.range(short_list.size):
# [ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15]
#
# np.range(short_list.size) -
# dfill(short_list[short_list.argsort()]):
# [ 0 1 2 3 4 5 6 7 8 9 10 11 0 0 1 0]
#
# unsorted:
# [ 0 1 2 0 3 4 5 0 6 7 8 0 9 10 11 1]
foo function recommended by #hpaulj using defaultdict
div function recommended by #Divakar (old, I'm sure he'd update it)
code
def dfill(a):
n = a.size
b = np.concatenate([[0], np.where(a[:-1] != a[1:])[0] + 1, [n]])
return np.arange(n)[b[:-1]].repeat(np.diff(b))
def argunsort(s):
n = s.size
u = np.empty(n, dtype=np.int64)
u[s] = np.arange(n)
return u
def cumcount(a):
n = a.size
s = a.argsort(kind='mergesort')
i = argunsort(s)
b = a[s]
return (np.arange(n) - dfill(b))[i]
def foo(l):
n = len(l)
r = np.empty(n, dtype=np.int64)
counter = defaultdict(int)
for i in range(n):
counter[l[i]] += 1
r[i] = counter[l[i]]
return r - 1
def div(l):
a = np.unique(l, return_counts=1)[1]
idx = a.cumsum()
id_arr = np.ones(idx[-1],dtype=int)
id_arr[0] = 0
id_arr[idx[:-1]] = -a[:-1]+1
rng = id_arr.cumsum()
return rng[argunsort(np.argsort(l))]
demonstration
cumcount(short_list)
array([ 0, 1, 2, 0, 3, 4, 5, 0, 6, 7, 8, 0, 9, 10, 11, 1])
time testing
code
functions = pd.Index(['cumcount', 'foo', 'foo2', 'div'], name='function')
lengths = pd.RangeIndex(100, 1100, 100, 'array length')
results = pd.DataFrame(index=lengths, columns=functions)
from string import ascii_letters
for i in lengths:
a = np.random.choice(list(ascii_letters), i)
for j in functions:
results.set_value(
i, j,
timeit(
'{}(a)'.format(j),
'from __main__ import a, {}'.format(j),
number=1000
)
)
results.plot()
Here's a vectorized approach using custom grouped range creating function and np.unique for getting the counts -
def grp_range(a):
idx = a.cumsum()
id_arr = np.ones(idx[-1],dtype=int)
id_arr[0] = 0
id_arr[idx[:-1]] = -a[:-1]+1
return id_arr.cumsum()
count = np.unique(A,return_counts=1)[1]
out = grp_range(count)[np.argsort(A).argsort()]
Sample run -
In [117]: A = list('aaabaaacaaadaaac')
In [118]: count = np.unique(A,return_counts=1)[1]
...: out = grp_range(count)[np.argsort(A).argsort()]
...:
In [119]: out
Out[119]: array([ 0, 1, 2, 0, 3, 4, 5, 0, 6, 7, 8, 0, 9, 10, 11, 1])
For getting the count, few other alternatives could be proposed with focus on performance -
np.bincount(np.unique(A,return_inverse=1)[1])
np.bincount(np.fromstring('aaabaaacaaadaaac',dtype=np.uint8)-97)
Additionally, with A containing single-letter characters, we could get the count simply with -
np.bincount(np.array(A).view('uint8')-97)
Besides defaultdict there are a couple of other counters. Testing a slightly simpler case:
In [298]: from collections import defaultdict
In [299]: from collections import defaultdict, Counter
In [300]: def foo(l):
...: counter = defaultdict(int)
...: for i in l:
...: counter[i] += 1
...: return counter
...:
In [301]: short_list = list('aaabaaacaaadaaac')
In [302]: foo(short_list)
Out[302]: defaultdict(int, {'a': 12, 'b': 1, 'c': 2, 'd': 1})
In [303]: Counter(short_list)
Out[303]: Counter({'a': 12, 'b': 1, 'c': 2, 'd': 1})
In [304]: arr=[ord(i)-ord('a') for i in short_list]
In [305]: np.bincount(arr)
Out[305]: array([12, 1, 2, 1], dtype=int32)
I constructed arr because bincount only works with ints.
In [306]: timeit np.bincount(arr)
The slowest run took 82.46 times longer than the fastest. This could mean that an intermediate result is being cached.
100000 loops, best of 3: 5.63 µs per loop
In [307]: timeit Counter(arr)
100000 loops, best of 3: 13.6 µs per loop
In [308]: timeit foo(arr)
100000 loops, best of 3: 6.49 µs per loop
I'm guessing it would hard to improve on pir2 based on default_dict.
Searching and counting like this are not a strong area for numpy.
In the following picture, I have DataFrame that renders zero after each cycle of operation (the cycle has random length). I want to calculate the average (or perform other operations) for each patch. For example, the average of [0.762, 0.766] alone, and [0.66, 1.37, 2.11, 2.29] alone and so forth till the end of the DataFrame.
So I worked with this data :
random_value
0 0
1 0
2 1
3 2
4 3
5 0
6 4
7 4
8 0
9 1
There is probably a way better solution, but here is what I came with :
def avg_function(df):
avg_list = []
value_list = list(df["random_value"])
temp_list = []
for i in range(len(value_list)):
if value_list[i] == 0:
if temp_list:
avg_list.append(sum(temp_list) / len(temp_list))
temp_list = []
else:
temp_list.append(value_list[i])
if temp_list: # for the last values
avg_list.append(sum(temp_list) / len(temp_list))
return avg_list
test_list = avg_function(df=df)
test_list
[Out] : [2.0, 4.0, 1.0]
Edit: since requested in the comments, here is a way to add the means to the dataframe. I dont know if there is a way to do that with pandas (and there might be!), but I came up with this :
def add_mean(df, mean_list):
temp_mean_list = []
list_index = 0 # will be the index for the value of mean_list
df["random_value_shifted"] = df["random_value"].shift(1).fillna(0)
random_value = list(df["random_value"])
random_value_shifted = list(df["random_value_shifted"])
for i in range(df.shape[0]):
if random_value[i] == 0 and random_value_shifted[i] == 0:
temp_mean_list.append(0)
elif random_value[i] == 0 and random_value_shifted[i] != 0:
temp_mean_list.append(0)
list_index += 1
else:
temp_mean_list.append(mean_list[list_index])
df = df.drop(["random_value_shifted"], axis=1)
df["mean"] = temp_mean_list
return df
df = add_mean(df=df, mean_list=mean_list
Which gave me :
df
[Out] :
random_value mean
0 0 0
1 0 0
2 1 2
3 2 2
4 3 2
5 0 0
6 4 4
7 4 4
8 0 0
9 1 1
I have a pandas data frame which is shown below:
>>> x = [[1,2,3,4,5],[1,2,4,4,3],[2,4,5,6,7]]
>>> columns = ['a','b','c','d','e']
>>> df = pd.DataFrame(data = x, columns = columns)
>>> df
a b c d e
0 1 2 3 4 5
1 1 2 4 4 3
2 2 4 5 6 7
I have an array of objects (conditions) as shown below:
[
{
'header' : 'a',
'condition' : '==',
'values' : [1]
},
{
'header' : 'b',
'condition' : '==',
'values' : [2]
},
...
]
and an assignHeader which is:
assignHeader = decision
now I want to do an operation which builds up all the conditions from the conditions array by looping through it, for example something like this:
pConditions = []
for eachCondition in conditions:
header = eachCondition['header']
values = eachCondition['values']
if eachCondition['condition'] == "==":
pConditions.append(df[header].isin(values))
else:
pConditions.append(~df[header].isin(values))
df[assignHeader ] = and(pConditions)
I was thinking of using all operator in pandas but am unable to crack the right syntax to do so. The list I shared can go big and dynamic and so I want to use this nested approach and check for the equality. Does anyone know a way to do so?
Final Output:
conditons = [df['a']==1,df['b']==2]
>>> df['decision'] = (df['a']==1) & (df['b']==2)
>>> df
a b c d e decision
0 1 2 3 4 5 True
1 1 2 4 4 3 True
2 2 4 5 6 7 False
Here conditions array will be variable. And I want to have a function which takes df, 'newheadernameandconditions` as input and returns the output as shown below:
>>> df
a b c d e decision
0 1 2 3 4 5 True
1 1 2 4 4 3 True
2 2 4 5 6 7 False
where newheadername = 'decision'
I was able to solve the problem using the code shown below. I am not sure if this is kind of fast way of getting things done, but would love to know your inputs in case you have any specific thing to point out.
def andMerging(conditions, mergeHeader, df):
if len(conditions) != 0:
df[mergeHeader] = pd.concat(conditions, axis = 1).all(axis = 1)
return df
where conditions are an array of pd.Series with boolean values.
And conditions are formatted as shown below:
def prepareForConditionMerging(conditionsArray, df):
conditions = []
for prop in conditionsArray:
condition = prop['condition']
values = prop['values']
header = prop['header']
if type(values) == str:
values = [values]
if condition=="==":
conditions.append(df[header].isin(values))
else:
conditions.append(~df[header].isin(values))
# Here we can add more conditions such as greater than less than etc.
return conditions
Sorry if this question has been asked before, but I did not find it here nor somewhere else:
I want to fill some of the fields of a column with tuples. Currently I would have to resort to:
import pandas as pd
df = pd.DataFrame({'a': [1,2,3,4]})
df['b'] = ''
df['b'] = df['b'].astype(object)
mytuple = ('x','y')
for l in df[df.a % 2 == 0].index:
df.set_value(l, 'b', mytuple)
with df being (which is what I want)
a b
0 1
1 2 (x, y)
2 3
3 4 (x, y)
This does not look very elegant to me and probably not very efficient. Instead of the loop, I would prefer something like
df.loc[df.a % 2 == 0, 'b'] = np.array([mytuple] * sum(df.a % 2 == 0), dtype=tuple)
which (of course) does not work. How can I improve my above method by using slicing?
In [57]: df.loc[df.a % 2 == 0, 'b'] = pd.Series([mytuple] * len(df.loc[df.a % 2 == 0])).values
In [58]: df
Out[58]:
a b
0 1
1 2 (x, y)
2 3
3 4 (x, y)
I have a pandas data frame, sample, with one of the columns called PR to which am applying a lambda function as follows:
sample['PR'] = sample['PR'].apply(lambda x: NaN if x < 90)
I then get the following syntax error message:
sample['PR'] = sample['PR'].apply(lambda x: NaN if x < 90)
^
SyntaxError: invalid syntax
What am I doing wrong?
You need mask:
sample['PR'] = sample['PR'].mask(sample['PR'] < 90, np.nan)
Another solution with loc and boolean indexing:
sample.loc[sample['PR'] < 90, 'PR'] = np.nan
Sample:
import pandas as pd
import numpy as np
sample = pd.DataFrame({'PR':[10,100,40] })
print (sample)
PR
0 10
1 100
2 40
sample['PR'] = sample['PR'].mask(sample['PR'] < 90, np.nan)
print (sample)
PR
0 NaN
1 100.0
2 NaN
sample.loc[sample['PR'] < 90, 'PR'] = np.nan
print (sample)
PR
0 NaN
1 100.0
2 NaN
EDIT:
Solution with apply:
sample['PR'] = sample['PR'].apply(lambda x: np.nan if x < 90 else x)
Timings len(df)=300k:
sample = pd.concat([sample]*100000).reset_index(drop=True)
In [853]: %timeit sample['PR'].apply(lambda x: np.nan if x < 90 else x)
10 loops, best of 3: 102 ms per loop
In [854]: %timeit sample['PR'].mask(sample['PR'] < 90, np.nan)
The slowest run took 4.28 times longer than the fastest. This could mean that an intermediate result is being cached.
100 loops, best of 3: 3.71 ms per loop
You need to add else in your lambda function. Because you are telling what to do in case your condition(here x < 90) is met, but you are not telling what to do in case the condition is not met.
sample['PR'] = sample['PR'].apply(lambda x: 'NaN' if x < 90 else x)