Calculate mean value by interval coordinates in pandas - python-3.x

I have a dataframe such as :
Name Position Value
A 1 10
A 2 11
A 3 10
A 4 8
A 5 6
A 6 12
A 7 10
A 8 9
A 9 9
A 10 9
A 11 9
A 12 9
and I woulde like for each interval of 3 position, to calculate the mean of Values.
And create a new df with start and end coordinates (of length 3 then), with the Mean_value column.
Name Start End Mean_value
A 1 3 10.33 <---- here this is (10+11+10)/3 = 10.33
A 4 6 8.7
A 7 9 9.3
A 10 13 9
Does someone have an idea using pandas please ?

Solution for get each 3 rows (if exist) per Name groups - first get counter by GroupBy.cumcount with integer division and pass it to named aggregations:
g = df.groupby('Name').cumcount() // 3
df = df.groupby(['Name',g]).agg(Start=('Position','first'),
End=('Position','last'),
Value=('Value','mean')).droplevel(1).reset_index()
print (df)
Name Start End Value
0 A 1 3 10.333333
1 A 4 6 8.666667
2 A 7 9 9.333333
3 A 10 12 9.000000

Related

For and if loop combination takes lot of time in Pandas (Data manipulation)

I have two datasets, each about half a million observations. I am writing the below code and it seems the code never seems to stop executing. I would like to know if there is a better way of doing it. Appreciate inputs.
Below are sample formats of my dataframes. Both dataframes share a set of 'sid' values , meaning all the 'sid' values in 'df2' will have a match in 'df1' 'sid' values. The 'tid' values and consequently the 'rid' values (which are a combination of 'sid' and 'tid' values) may not appear in both sets.
The task is simple. I would like to create the 'tv' column in df2. Wherever the 'rid' in df2 matches with the 'rid' in 'df1', the 'tv' column in df2 takes the corresponding 'tv' value from df1. If it does not match, the 'tv' value in 'df2' will be the median 'tv' value for the matching 'sid' subset in 'df1'.
In fact my original task includes creating a few more similar columns like 'tv' in df2 (based on their values in 'df1' ; these columns exist in 'df1').
I believe as my code contains for loop combined with if else statement and multiple value assignment statements, it is taking forever to execute. Appreciate any inputs.
df1
sid tid rid tv
0 0 0 0-0 9
1 0 1 0-1 8
2 0 3 0-3 4
3 1 5 1-5 2
4 1 7 1-7 3
5 1 9 1-9 14
6 1 10 1-10 24
7 1 11 1-11 13
8 2 14 2-14 2
9 2 16 2-16 5
10 3 17 3-17 6
11 3 18 3-18 8
12 3 20 3-20 5
13 3 21 3-21 11
14 4 23 4-23 6
df2
sid tid rid
0 0 0 0-0
1 0 2 0-2
2 1 3 1-3
3 1 6 1-6
4 1 9 1-9
5 2 10 2-10
6 2 12 2-12
7 3 1 3-1
8 3 15 3-15
9 3 1 3-1
10 4 19 4-19
11 4 22 4-22
rids = [rid.split('-') for rid in df1.rid]
for r in df2.rid:
s,t = r.split('-')
if [s,t] in rids:
df2.loc[df2.rid== r,'tv'] = df1.loc[df1.rid == r,'tv']
else:
df2.loc[df2.rid== r,'tv'] = df1.loc[df1.sid == int(s),'tv'].median()
The expected df2 shall be as follows:
sid tid rid tv
0 0 0 0-0 9.0
1 0 2 0-2 8.0
2 1 3 1-3 13.0
3 1 6 1-6 13.0
4 1 9 1-9 14.0
5 2 10 2-10 3.5
6 2 12 2-12 3.5
7 3 1 3-1 7.0
8 3 15 3-15 7.0
9 3 1 3-1 7.0
10 4 19 4-19 6.0
11 4 22 4-22 6.0
You can left merge on df2 with a subset(because you need only tv column you can also pass the df1 without any subset) of df1 on 'rid' then calculate median and fill values:
out=df2.merge(df1[['rid','tv']],on='rid',how='left')
out['tv']=out['tv_y'].fillna(out['sid'].map(df1.groupby('sid')['tv'].median()))
out= out.drop(['tv_x','tid_y','tv_y'], axis=1)
out = out.rename(columns = {'tid_x': 'tid'})
out
OR
Since you said that:
all the 'sid' values in 'df2' will have a match in 'df1' 'sid' values
So you can also left merge them on ['sid','rid'] and then fillna() value of tv with the median of df1 'tv' column by mapping values using map() method:
out=df2.merge(df1,on=['sid','rid'],how='left')
out['tv']=out['tv_y'].fillna(out['sid'].map(df1.groupby('sid')['tv'].median()))
out= out.drop(['tv_x','tv_y'], axis=1)
out
output of out:
sid tid rid tv
0 0 0 0-0 9.0
1 0 2 0-2 8.0
2 1 3 1-3 13.0
3 1 6 1-6 13.0
4 1 9 1-9 14.0
5 2 10 2-10 3.5
6 2 12 2-12 3.5
7 3 1 3-1 7.0
8 3 15 3-15 7.0
9 3 1 3-1 7.0
10 4 19 4-19 6.0
11 4 22 4-22 6.0
Here is a suggestion without any loops, based on dictionaries:
matching_values = dict(zip(df1['rid'][df1['rid'].isin(df2['rid'])], df1['tv'][df1['rid'].isin(df2['rid'])]))
df2[df2['rid'].isin(df1['rid'])]['tv'] = df2[df2['rid'].isin(df1['rid'])]['rid']
df2[df2['rid'].isin(df1['rid'])]['tv'].replace(matching_values)
median_values = df2[(~df2['rid'].isin(df1['rid']) & (df2['sid'].isin(df1['sid'])].groupby('sid')['tv'].median().to_dict()
df2[(~df2['rid'].isin(df1['rid']) & (df2['sid'].isin(df1['sid'])]['tv'] = df2[(~df2['rid'].isin(df1['rid']) & (df2['sid'].isin(df1['sid'])]['sid']
df2[(~df2['rid'].isin(df1['rid']) & (df2['sid'].isin(df1['sid'])]['tv'].replace(median_values)
This should do the trick. The logic here is that we first create a dictionary, in which the "rid and "sid" values are the keys and the median and matching "tv" values are the dictionary values. Next, we replace the "tv" values in df2 with the rid and sid keys, respectively, (because they are the dictionary keys) which can thus easily be replaced by the correct tv values by calling .replace().
Don't use for loops in pandas, that is known to be slow. That way you don't get to benefit from all the internal optimizations that have been made.
Try to use the split-apply-combine pattern:
split df1 into sid to calculate the median: df1.groupby('sid')['tv'].median()
join df2 on df1: df2.join(df1.set_index('rid'), on='rid')
fill the NaN values with the median calculated in step 1.
(Haven't tested the code).

pandas groupby extract top percent n data(descending)

I have some data like this
df = pd.DataFrame({'class':['a','a','b','b','a','a','b','c','c'],'score':[3,5,6,7,8,9,10,11,14]})
df
class score
0 a 3
1 a 5
2 b 6
3 b 7
4 a 8
5 a 9
6 b 10
7 c 11
8 c 14
I want to use groupby function extract top n% data(descending by score),i know the nlargest can make it,but the number of every group is different,so i don't know how to do it
I tried this function
top_n = 0.5
g = df.groupby(['class'])['score'].apply(lambda x:x.nlargest(int(round(top_n*len(x))),keep='all')).reset_index()
g
class level_1 score
0 a 5 9
1 a 4 8
2 b 6 10
3 b 3 7
4 c 8 14
but it can not deal with big data(more than 10 million),it is very slow,how do i speed it,thank you!

How to select last 5 rows of each unique records in pandas

Using python 3 am trying for each uniqe row in the column 'Name' to get the last 5 records from the column 'Number'. How exactly can this be done in python?
My df looks like this:
Name Number
a 5
a 6
b 7
b 8
a 9
a 10
b 11
b 12
a 9
b 8
I saw same exmples(like this one Get sum of last 5 rows for each unique id ) in SQL but that is time consuming and I would like to learn how to do it in python.
My expected output df would be like this:
Name 1 2 3 4 5
a 5 6 9 10 9
b 7 8 11 12 8
I think you need something like this:
df_out = df.groupby('Name').tail(5)
df_out.set_index(['Name', df_out.groupby('Name').cumcount() +1])['Number'].unstack()
Output:
1 2 3 4 5
Name
a 5 6 9 10 9
b 7 8 11 12 8
Looks like you need pivot after a groupby.cumcount()
df1=df.groupby('Name').tail(5)
final=(df1.assign(k=df1.groupby('Name').cumcount()+1)
.pivot(index='Name', columns='k', values='Number')
.reset_index().rename_axis(None, axis=1))
print(final)
Name 1 2 3 4 5
0 a 5 6 9 10 9
1 b 7 8 11 12 8

Variable string formatting in python 3

Input is a number, e.g. 9 and I want to print decimal, octal, hex and binary value from 1 to 9 like:
1 1 1 1
2 2 2 10
3 3 3 11
4 4 4 100
5 5 5 101
6 6 6 110
7 7 7 111
8 10 8 1000
9 11 9 1001
How can I achieve this in python3 using syntax like
dm, oc, hx, bn = len(str(9)), len(bin(9)[2:]), ...
print("{:dm%d} {:oc%s}" % (i, oct(i[2:]))
I mean if number is 999 so I want decimal 10 to be printed like ' 10' and binary equivalent of 999 is 1111100111 so I want 10 like ' 1010'.
You can use str.format() and its mini-language to do the whole thing for you:
for i in range(1, 10):
print("{v} {v:>6o} {v:>6x} {v:>6b}".format(v=i))
Which will print:
1 1 1 1
2 2 2 10
3 3 3 11
4 4 4 100
5 5 5 101
6 6 6 110
7 7 7 111
8 10 8 1000
9 11 9 1001
UPDATE: To define field 'widths' in a variable you can use a format-within-format structure:
w = 5 # field width, i.e. offset to the right for all octal/hex/binary values
for i in range(1, 10):
print("{v} {v:>{w}o} {v:>{w}x} {v:>{w}b}".format(v=i, w=w))
Or define a different width variable for each field type if you want them non-uniformly spaced.
Btw. since you've tagged your question with python-3.x, if you're using Python 3.6 or newer, you can use Literal String Interpolation to simplify it even more:
w = 5 # field width, i.e. offset to the right for all octal/hex/binary values
for v in range(1, 10):
print(f"{v} {v:>{w}o} {v:>{w}x} {v:>{w}b}")

Average of multiple files with unequal row sizes in Shell

I have 15 datafiles with unequal row sizes, but number of columns in each file is same. e.g.
ifile1.dat ifile2.dat ifile3.dat and so on ............
0 0 0 0 1 6
1 2 5 3 2 7
2 5 6 10 4 6
5 2 8 9 5 9
10 2 10 3 8 2
In each file 1st column represents the index number.
I would like to compute average of all these files for each index number in column 1. i.e.
ofile.txt
0 0 [This is computed as (0+0)/2]
1 4 [This is computed as (2+6)/2]
2 6 [This is computed as (5+7)/2]
3 [no value]
4 6 [This is computed as (6)/1]
5 4.66 [This is computed as (2+3+9)/3]
6 10
7
8 5.5
9
10 2.5
I can't think of any simple method to do it. I was thinking of a method, but seems very lengthy. Taking the average after converting all the files with same row size, .e.g.
ifile1.dat ifile2.dat ifile3.dat and so on ............
0 0 0 0 0 0
1 2 1 1 6
2 5 2 2 7
3 3 3
4 4 4 6
5 2 5 3 5 9
6 6 10 6
7 7 7
8 8 9 8 2
9 9 9
10 2 10 3 10
$ awk '{s[$1]+=$2; c[$1]++;} END{for (i in s) print i,s[i]/c[i];}' ifile*.dat
0 0
1 4
2 6
4 6
5 4.66667
6 10
8 5.5
10 2.5
In the above code, there are two arrays, s and c. s[i] is the sum of all entries with index i and c[i] is the number of entries with index i. After we have read all the files, we print the average, s[i]/c[i], for each index i.

Resources