Develop Reference

Pandas dataframe: creating a new column that is a custom function using 2 other columns

Consider the following data set stored in a pandas DataFrame dfX:
A B
1 2
4 6
7 9
I have a function that is:
def someThingSpecial(x,y)
# z = do something special with x,y
return z
I now want to create a new column in df that bears the computed z value
Looking at other SO examples, I've tried several variants including:
dfX['C'] = dfX.apply(lambda x: someThingSpecial(x=x['A'], y=x['B']), axis=1)
Which returns errors. What is the right way to do this?

This seems to work for me on v0.21. Take a look -
df
A B
0 1 2
1 4 6
2 7 9
def someThingSpecial(x,y):
return x + y
df.apply(lambda x: someThingSpecial(x.A, x.B), 1)
0 3
1 10
2 16
dtype: int64
You might want to try upgrading your pandas version to the latest stable release (0.21 as of now).
Here's another option. You can vectorise your function.
v = np.vectorize(someThingSpecial)
v now accepts arrays, but operates on each pair of elements individually. Note that this just hides the loop, as apply does, but is much cleaner. Now, you can compute C as so -
df['C'] = v(df.A, df.B)

if your function only needs one column's value, then do this instead of coldspeed's answer:
dfX['A'].apply(your_func)
to store it:
dfX['C'] = dfX['A'].apply(your_func)

How do I sort a dataframe by an array not in the dataframe

I've answered this question several times in the guise of different contexts and I realized that there isn't a good canonical approach specified anywhere.
So, to set up a simple problem:
Problem
df = pd.DataFrame(dict(A=range(6), B=[1, 2] * 3))
print(df)
A B
0 0 1
1 1 2
2 2 1
3 3 2
4 4 1
5 5 2
Question:
How do I sort by the product of columns 'A' and 'B'?
Here is an approach where I add a temporary column to the dataframe, use it to sort_values then drop it.
df.assign(P=df.prod(1)).sort_values('P').drop('P', 1)
A B
0 0 1
1 1 2
2 2 1
4 4 1
3 3 2
5 5 2
Is there a better, more concise, clearer, more consistent approach?

TL;DR
iloc + argsort
We can approach this using iloc where we can take an array of ordinal positions and return the dataframe reordered by these positions.
With the power of iloc, we can sort with any array that specifies the order.
Now, all we need to do is identify a method for getting this ordering. Turns out there is a method called argsort which does exactly this. By passing the results of argsort to iloc, we can get our dataframe sorted out.
Example 1
Using the specified problem above
df.iloc[df.prod(1).argsort()]
Same results as above
A B
0 0 1
1 1 2
2 2 1
4 4 1
3 3 2
5 5 2
That was for simplicity. We could take this further if performance is an issue and focus on numpy
v = df.values
a = v.prod(1).argsort()
pd.DataFrame(v[a], df.index[a], df.columns)
How fast are these solutions?
We can see that pd_ext_sort is the most concise but does not scale as well as the others.
np_ext_sort gives the best performance at the expense of transparency. Though, I'd argue that it's still very clear what is going on.
backtest setup
def add_drop():
return df.assign(P=df.prod(1)).sort_values('P').drop('P', 1)
def pd_ext_sort():
return df.iloc[df.prod(1).argsort()]
def np_ext_sort():
v = df.values
a = v.prod(1).argsort()
return pd.DataFrame(v[a], df.index[a], df.columns)
results = pd.DataFrame(
index=pd.Index([10, 100, 1000, 10000], name='Size'),
columns=pd.Index(['add_drop', 'pd_ext_sort', 'np_ext_sort'], name='method')
)
for i in results.index:
df = pd.DataFrame(np.random.rand(i, 2), columns=['A', 'B'])
for j in results.columns:
stmt = '{}()'.format(j)
setup = 'from __main__ import df, {}'.format(j)
results.set_value(i, j, timeit(stmt, setup, number=100))
results.plot()
Example 2
Suppose I have a column of negative and positive values. I want to sort by increasing magnitude... however, I want the negatives to come first.
Suppose I have dataframe df
df = pd.DataFrame(dict(A=range(-2, 3)))
print(df)
A
0 -2
1 -1
2 0
3 1
4 2
I'll set up 3 versions again. This time I'll use np.lexsort which returns the same type of array as argsort. Meaning, I can use it to reorder the dataframe.
Caveat: np.lexsort sorts by the last array in its list first. \shurg
def add_drop():
return df.assign(P=df.A >= 0, M=df.A.abs()).sort_values(['P', 'M']).drop(['P', 'M'], 1)
def pd_ext_sort():
v = df.A.values
return df.iloc[np.lexsort([np.abs(v), v >= 0])]
def np_ext_sort():
v = df.A.values
a = np.lexsort([np.abs(v), v >= 0])
return pd.DataFrame(v[a, None], df.index[a], df.columns)
All of which return
A
1 -1
0 -2
2 0
3 1
4 2
How fast this time?
In this example, both pd_ext_sort and np_ext_sort outperformed add_drop.
backtest setup
results = pd.DataFrame(
index=pd.Index([10, 100, 1000, 10000], name='Size'),
columns=pd.Index(['add_drop', 'pd_ext_sort', 'np_ext_sort'], name='method')
)
for i in results.index:
df = pd.DataFrame(np.random.randn(i, 1), columns=['A'])
for j in results.columns:
stmt = '{}()'.format(j)
setup = 'from __main__ import df, {}'.format(j)
results.set_value(i, j, timeit(stmt, setup, number=100))
results.plot(figsize=(15, 6))

Comparing rows of two pandas dataframes?

This is a continuation of my question. Fastest way to compare rows of two pandas dataframes?
I have two dataframes A and B:
A is 1000 rows x 500 columns, filled with binary values indicating either presence or absence.
For a condensed example:
A B C D E
0 0 0 0 1 0
1 1 1 1 1 0
2 1 0 0 1 1
3 0 1 1 1 0
B is 1024 rows x 10 columns, and is a full iteration from 0 to 1023 in binary form.
Example:
0 1 2
0 0 0 0
1 0 0 1
2 0 1 0
3 0 1 1
4 1 0 0
5 1 0 1
6 1 1 0
7 1 1 1
I am trying to find which rows in A, at a particular 10 columns of A, correspond with each row of B.
Each row of A[My_Columns_List] is guaranteed to be somewhere in B, but not every row of B will match up with a row in A[My_Columns_List]
For example, I want to show that for columns [B,D,E] of A,
rows [1,3] of A match up with row [6] of B,
row [0] of A matches up with row [2] of B,
row [2] of A matches up with row [3] of B.
I have tried using:
pd.merge(B.reset_index(), A.reset_index(),
left_on = B.columns.tolist(),
right_on =A.columns[My_Columns_List].tolist(),
suffixes = ('_B','_A')))
This works, but I was hoping that this method would be faster:
S = 2**np.arange(10)
A_ID = np.dot(A[My_Columns_List],S)
B_ID = np.dot(B,S)
out_row_idx = np.where(np.in1d(A_ID,B_ID))[0]
But when I do this, out_row_idx returns an array containing all the indices of A, which doesn't tell me anything.
I think this method will be faster, but I don't know why it returns an array from 0 to 999.
Any input would be appreciated!
Also, credit goes to #jezrael and #Divakar for these methods.

I'll stick by my initial answer but maybe explain better.
You are asking to compare 2 pandas dataframes. Because of that, I'm going to build dataframes. I may use numpy, but my inputs and outputs will be dataframes.
Setup
You said we have a a 1000 x 500 array of ones and zeros. Let's build that.
A_init = pd.DataFrame(np.random.binomial(1, .5, (1000, 500)))
A_init.columns = pd.MultiIndex.from_product([range(A_init.shape[1]/10), range(10)])
A = A_init
In addition, I gave A a MultiIndex to easily group by columns of 10.
Solution
This is very similar to #Divakar's answer with one minor difference that I'll point out.
For one group of 10 ones and zeros, we can treat it as a bit array of length 8. We can then calculate what it's integer value is by taking the dot product with an array of powers of 2.
twos = 2 ** np.arange(10)
I can execute this for every group of 10 ones and zeros in one go like this
AtB = A.stack(0).dot(twos).unstack()
I stack to get a row of 50 groups of 10 into columns in order to do the dot product more elegantly. I then brought it back with the unstack.
I now have a 1000 x 50 dataframe of numbers that range from 0-1023.
Assume B is a dataframe with each row one of 1024 unique combinations of ones and zeros. B should be sorted like B = B.sort_values().reset_index(drop=True).
This is the part I think I failed at explaining last time. Look at
AtB.loc[:2, :2]
That value in the (0, 0) position, 951 means that the first group of 10 ones and zeros in the first row of A matches the row in B with the index 951. That's what you want!!! Funny thing is, I never looked at B. You know why, B is irrelevant!!! It's just a goofy way of representing the numbers from 0 to 1023. This is the difference with my answer, I'm ignoring B. Ignoring this useless step should save time.
These are all functions that take two dataframes A and B and returns a dataframe of indices where A matches B. Spoiler alert, I'll ignore B completely.
def FindAinB(A, B):
assert A.shape[1] % 10 == 0, 'Number of columns in A is not a multiple of 10'
rng = np.arange(A.shape[1])
A.columns = pd.MultiIndex.from_product([range(A.shape[1]/10), range(10)])
twos = 2 ** np.arange(10)
return A.stack(0).dot(twos).unstack()
def FindAinB2(A, B):
assert A.shape[1] % 10 == 0, 'Number of columns in A is not a multiple of 10'
rng = np.arange(A.shape[1])
A.columns = pd.MultiIndex.from_product([range(A.shape[1]/10), range(10)])
# use clever bit shifting instead of dot product with powers
# questionable improvement
return (A.stack(0) << np.arange(10)).sum(1).unstack()
I'm channelling my inner #Divakar (read, this is stuff I've learned from Divakar)
def FindAinB3(A, B):
assert A.shape[1] % 10 == 0, 'Number of columns in A is not a multiple of 10'
a = A.values.reshape(-1, 10)
a = np.einsum('ij->i', a << np.arange(10))
return pd.DataFrame(a.reshape(A.shape[0], -1), A.index)
Minimalist One Liner
f = lambda A: pd.DataFrame(np.einsum('ij->i', A.values.reshape(-1, 10) << np.arange(10)).reshape(A.shape[0], -1), A.index)
Use it like
f(A)
Timing
FindAinB3 is an order of magnitude faster

Fastest way to compare rows of two pandas dataframes?

So I have two pandas dataframes, A and B.
A is 1000 rows x 500 columns, filled with binary values indicating either presence or absence.
B is 1024 rows x 10 columns, and is a full iteration of 0's and 1's, hence having 1024 rows.
I am trying to find which rows in A, at a particular 10 columns of A, correspond with a given row in B. I need the whole row to match up, rather than element by element.
For example, I would want
A[(A.ix[:,(1,2,3,4,5,6,7,8,9,10)==(1,0,1,0,1,0,0,1,0,0)).all(axis=1)]
To return something that rows (3,5,8,11,15) in A match up with that (1,0,1,0,1,0,0,1,0,0) row of B at those particular columns (1,2,3,4,5,6,7,8,9,10)
And I want to do this over every row in B.
The best way I could figure out to do this was:
import numpy as np
for i in B:
B_array = np.array(i)
Matching_Rows = A[(A.ix[:,(1,2,3,4,5,6,7,8,9,10)] == B_array).all(axis=1)]
Matching_Rows_Index = Matching_Rows.index
This isn't terrible for one instance, but I use it in a while loop that runs around 20,000 times; therefore, it slows it down quite a bit.
I have been messing around with DataFrame.apply to no avail. Could map work better?
I was just hoping someone saw something obviously more efficient as I am fairly new to python.
Thanks and best regards!

We can abuse the fact that both dataframes have binary values 0 or 1 by collapsing the relevant columns from A and all columns from B into 1D arrays each, when considering each row as a sequence of binary numbers that could be converted to decimal number equivalents. This should reduce the problem set considerably, which would help with performance. Now, after getting those 1D arrays, we can use np.in1d to look for matches from B in A and finally np.where on it to get the matching indices.
Thus, we would have an implementation like so -
# Setup 1D arrays corresponding to selected cols from A and entire B
S = 2**np.arange(10)
A_ID = np.dot(A[range(1,11)],S)
B_ID = np.dot(B,S)
# Look for matches that exist from B_ID in A_ID, whose indices
# would be desired row indices that have matched from B
out_row_idx = np.where(np.in1d(A_ID,B_ID))[0]
Sample run -
In [157]: # Setup dataframes A and B with rows 0, 4 in A having matches from B
...: A_arr = np.random.randint(0,2,(10,14))
...: B_arr = np.random.randint(0,2,(7,10))
...:
...: B_arr[2] = A_arr[4,1:11]
...: B_arr[4] = A_arr[4,1:11]
...: B_arr[5] = A_arr[0,1:11]
...:
...: A = pd.DataFrame(A_arr)
...: B = pd.DataFrame(B_arr)
...:
In [158]: S = 2**np.arange(10)
...: A_ID = np.dot(A[range(1,11)],S)
...: B_ID = np.dot(B,S)
...: out_row_idx = np.where(np.in1d(A_ID,B_ID))[0]
...:
In [159]: out_row_idx
Out[159]: array([0, 4])

You can use merge with reset_index - output are indexes of B which are equal in A in custom columns:
A = pd.DataFrame({'A':[1,0,1,1],
'B':[0,0,1,1],
'C':[1,0,1,1],
'D':[1,1,1,0],
'E':[1,1,0,1]})
print (A)
A B C D E
0 1 0 1 1 1
1 0 0 0 1 1
2 1 1 1 1 0
3 1 1 1 0 1
B = pd.DataFrame({'0':[1,0,1],
'1':[1,0,1],
'2':[1,0,0]})
print (B)
0 1 2
0 1 1 1
1 0 0 0
2 1 1 0
print (pd.merge(B.reset_index(),
A.reset_index(),
left_on=B.columns.tolist(),
right_on=A.columns[[0,1,2]].tolist(),
suffixes=('_B','_A')))
index_B 0 1 2 index_A A B C D E
0 0 1 1 1 2 1 1 1 1 0
1 0 1 1 1 3 1 1 1 0 1
2 1 0 0 0 1 0 0 0 1 1
print (pd.merge(B.reset_index(),
A.reset_index(),
left_on=B.columns.tolist(),
right_on=A.columns[[0,1,2]].tolist(),
suffixes=('_B','_A'))[['index_B','index_A']])
index_B index_A
0 0 2
1 0 3
2 1 1

You can do it in pandas by using loc or ix and telling it to find the rows where the ten columns are all equal. Like this:
A.loc[(A[1]==B[1]) & (A[2]==B[2]) & (A[3]==B[3]) & A[4]==B[4]) & (A[5]==B[5]) & (A[6]==B[6]) & (A[7]==B[7]) & (A[8]==B[8]) & (A[9]==B[9]) & (A[10]==B[10])]
This is quite ugly in my opinion but it will work and gets rid of the loop so it should be significantly faster. I wouldn't be surprised if someone could come up with a more elegant way of coding the same operation.

In this special case, your rows of 10 zeros and ones can be interpreted as 10 digit binaries. If B is in order, then it can be interpreted as a range from 0 to 1023. In this case, all we need to do is take A's rows in 10 column chunks and calculate what its binary equivalent is.
I'll start by defining a range of powers of two so I can do matrix multiplication with it.
twos = pd.Series(np.power(2, np.arange(10)))
Next, I'll relabel A's columns into a MultiIndex and stack to get my chunks of 10.
A = pd.DataFrame(np.random.binomial(1, .5, (1000, 500)))
A.columns = pd.MultiIndex.from_tuples(zip((A.columns / 10).tolist(), (A.columns % 10).tolist()))
A_ = A.stack(0)
A_.head()
Finally, I'll multiply A_ with twos to get integer representation of each row and unstack.
A_.dot(twos).unstack()
This is now a 1000 x 50 DataFrame where each cell represents which of B's rows we matched for that particular 10 column chunk for that particular row of A. There isn't even a need for B.

cogroup like operation for pandas

I was trying to using pandas to analysis a fairly large data set (~5GB). I wanted to divide the data sets into groups, then perform a Cartesian product on each group, and then aggregate the result.
The apply operation of pandas is quite expressive, I could first group, and then do the Cartesian product on each group using apply, and then aggregate the result using sum. The problem with this approach, however, is that apply is not lazy, it will compute all the intermediate results before the aggregation, and the intermediate results (Cartesian production on each group) is very large.
I was looking at Apache Spark and found one very interesting operator called cogroup. The definition is here:
When called on datasets of type (K, V) and (K, W), returns a dataset of (K, Iterable, Iterable) tuples. This operation is also called groupWith.
This seems to be exactly what I want. If I could first cogroup and then do a sum, then the intermediate results won't be expanded (assuming cogroup works in the same lazy fashion as group).
Is there operation similar to cogroup in pandas, or how to achieve my goal efficiently?
Here is my example:
I want to group the data by id, and then do a Cartesian product for each group, and then group by cluster_x and cluster_y and aggregate the count_x and count_y using sum. The following code works, but is extremely slow and consumes too much memory.
# add dummy_key to do Cartesian product by merge
df['dummy_key'] = 1
def join_group(g):
return pandas.merge(g, g, on='dummy_key')\
[['cache_cluster_x', 'count_x', 'cache_cluster_y', 'count_y']]
df_count_stats = df.groupby(['id'], as_index=True).apply(join_group).\
groupby(['cache_cluster_x', 'cache_cluster_y'], as_index=False)\
[['count_x', 'count_y']].sum()
A toy data set
id cluster count
0 i1 A 2
1 i1 B 3
2 i2 A 1
3 i2 B 4
Intermediate result after the apply (can be large)
cluster_x count_x cluster_y count_y
id
i1 0 A 2 A 2
1 A 2 B 3
2 B 3 A 2
3 B 3 B 3
i2 0 A 1 A 1
1 A 1 B 4
2 B 4 A 1
3 B 4 B 4
The desired final result
cluster_x cluster_y count_x count_y
0 A A 3 3
1 A B 3 7
2 B A 7 3
3 B B 7 7

My first attempt failed, sort of: while I was able to limit the memory use (by summing over the Cartesian product within each group), it was considerably slower than the original. But for your particular desired output, I think we can simplify the problem considerably:
import numpy as np, pandas as pd
def fake_data(nids, nclusters, ntile):
ids = ["i{}".format(i) for i in range(1,nids+1)]
clusters = ["A{}".format(i) for i in range(nclusters)]
df = pd.DataFrame(index=pd.MultiIndex.from_product([ids, clusters], names=["id", "cluster"]))
df = df.reset_index()
df = pd.concat([df]*ntile)
df["count"] = np.random.randint(0, 10, size=len(df))
return df
def join_group(g):
m= pd.merge(g, g, on='dummy_key')
return m[['cluster_x', 'count_x', 'cluster_y', 'count_y']]
def old_method(df):
df["dummy_key"] = 1
h1 = df.groupby(['id'], as_index=True).apply(join_group)
h2 = h1.groupby(['cluster_x', 'cluster_y'], as_index=False)
h3 = h2[['count_x', 'count_y']].sum()
return h3
def new_method1(df):
m1 = df.groupby("cluster", as_index=False)["count"].sum()
m1["dummy_key"] = 1
m2 = m1.merge(m1, on="dummy_key")
m2 = m2.sort_index(axis=1).drop(["dummy_key"], axis=1)
return m2
which gives (with df as your toy frame):
>>> new_method1(df)
cluster_x cluster_y count_x count_y
0 A A 3 3
1 A B 3 7
2 B A 7 3
3 B B 7 7
>>> df2 = fake_data(100, 100, 1)
>>> %timeit old_method(df2)
1 loops, best of 3: 954 ms per loop
>>> %timeit new_method1(df2)
100 loops, best of 3: 8.58 ms per loop
>>> (old_method(df2) == new_method1(df2)).all().all()
True
and even
>>> df2 = fake_data(100, 100, 100)
>>> %timeit new_method1(df2)
10 loops, best of 3: 88.8 ms per loop
Whether this will be enough of an improvement to handle your actual case, I'm not sure.