Develop Reference

Create dummy variabel and fill based on condition

I have a dataframe consisting of online reviews. I have assigned topics (topic 1-5; and 0 meaning no topic is assigned) and labels (positive or negative) in each instance. I want to create a dummy variable for each topic and label. This is what my data looks like...
reviewId
topic
label
01
2
negative
02
2
positive
03
0
negative
04
5
negative
05
1
positive
What should I do to make my data look like this? (1 meaning assigned, 0 meaning not assigned)
reviewId
topic
label
T1pos
T1neg
T2pos
T2neg
T3pos
T3neg
T4pos
T4neg
T5pos
T5neg
01
2
negative
0
0
0
1
0
0
0
0
0
0
02
2
positive
0
0
1
0
0
0
0
0
0
0
03
0
negative
0
0
0
0
0
0
0
0
0
0
04
5
negative
0
0
0
0
0
0
0
0
0
1
05
1
positive
1
0
0
0
0
0
0
0
0
0

You can create your own encoding by converting the two columns to a power of two and get its binary representation:
# I used 'p' as 'pos' and 'n' as 'neg' to save space
MAX_TOPIC = df['topic'].max()
mi = pd.MultiIndex.from_product([range(1, MAX_TOPIC+1), ['p', 'n']])
mi = [f'T{t}{l}' for t, l in mi]
# >> 2 to remove T0n and T0p
num = np.array(2**(df['topic']*2+df['label'].eq('negative'))) >> 2
hot = (((n[:, None] & (1 << np.arange(MAX_TOPIC*2)))) > 0).astype(int)
out = pd.concat([df, pd.DataFrame(hot, columns=mi, index=df.index)], axis=1)
Output:
>>> out
reviewId topic label T1p T1n T2p T2n T3p T3n T4p T4n T5p T5n
0 1 2 negative 0 0 0 1 0 0 0 0 0 0
1 2 2 positive 0 0 1 0 0 0 0 0 0 0
2 3 0 negative 0 0 0 0 0 0 0 0 0 0
3 4 5 negative 0 0 0 0 0 0 0 0 0 1
4 5 1 positive 1 0 0 0 0 0 0 0 0 0
>>> num
array([ 8, 4, 0, 512, 1])
The binary representation comes from Convert integer to binary array with suitable padding

Someone can probably come up with a more elegant solution, but this works:
import numpy as np
import pandas as pd
# recreate your DataFrame:
df = pd.DataFrame({
'reviewid': ['01', '02', '03', '04', '05'],
'topic': [2, 2, 0, 5, 1],
'label': ['neg', 'pos', 'neg', 'neg', 'pos']})
# Add dummy columns initialized to 0:
dummies = [
f'T{t}{lab}' for t in sorted(df.topic.unique()) if t != 0
for lab in sorted(df.label.unique())]
dummy_df = pd.DataFrame(
np.zeros((len(df), len(dummies)), dtype=int),
columns=dummies,
index=df.index)
df = pd.concat([df, dummy_df], axis=1)
# Fill in the dummy columns
for i, (t, lab) in enumerate(zip(df.topic, df.label)):
if t != 0:
df.loc[i, f'T{t}{lab}'] = 1
df # view result

Rename the row and columns of a dataframe

I have extracted a NC file using python and after processing the data the final output is an array with the of (199, 314). I convert the array to data frame, but the rows and columns names (index) start from zero to 199 and 314, respectively.
from netCDF4 import Dataset
import numpy as np
import pandas as pd
data = Dataset('GolestanM.nc', 'r')
dims = data.dimensions
ndims = len(dims)
vars = data.variables
nvars = len(vars)
attrs = data.ncattrs
lon = data.variables['lon'][:]
lat = data.variables['lat'][:]
t = data.variables['time'][496]
fire = data.variables['FireMask'][496,:,:]
dataset = pd.DataFrame(fire)
However, I want to rename these index using the following format:
Columns: first name 53.7042 and then +0.0083 until the name reach 56.3208
[0-->53.7042, 1-->53.7.25, ... , 314-->53.3208]
Rows: first name 38.1125 and then -0.0083 until the name reach 36.4625
[0-->38.1125, 1-->38.1042, ... , 199-->36.4625]
to do this I have the code bellow:
dataset = dataset.rename(index={0: "38.1125"})
dataset = dataset.rename(columns={0: "53.7042"})
dataset = dataset.rename(index = lambda x: x + (0.0083),
columns = lambda x: x + (0.0083))
However doin this give me the following error:
TypeError: can only concatenate str (not "float") to str
CAn any one help me with the probem.

Idea is multiple x (columns or index names) in lambda function:
#sample data
dataset = pd.DataFrame(0, index=range(10), columns=range(10))
dataset = dataset.rename(index = lambda x: 38.1125 - 0.0083 * x ,
columns = lambda x: 53.7042 + 0.0083* x)
print (dataset)
53.7042 53.7125 53.7208 53.7291 53.7374 53.7457 53.7540 \
38.1125 0 0 0 0 0 0 0
38.1042 0 0 0 0 0 0 0
38.0959 0 0 0 0 0 0 0
38.0876 0 0 0 0 0 0 0
38.0793 0 0 0 0 0 0 0
38.0710 0 0 0 0 0 0 0
38.0627 0 0 0 0 0 0 0
38.0544 0 0 0 0 0 0 0
38.0461 0 0 0 0 0 0 0
38.0378 0 0 0 0 0 0 0
53.7623 53.7706 53.7789
38.1125 0 0 0
38.1042 0 0 0
38.0959 0 0 0
38.0876 0 0 0
38.0793 0 0 0
38.0710 0 0 0
38.0627 0 0 0
38.0544 0 0 0
38.0461 0 0 0
38.0378 0 0 0

How to split a list using two nested conditions

Basically I have list of 0s and 1s. Each value in the list represents a data sample from an hour. Thus, if there are 24 0s and 1s in the list that means there are 24 hours, or a single day. I want to capture the first time the data cycles from 0s to 1s back to 0s in a span of 24 hours (or vice versa from 1s to 0s back to 1s).
signal = [1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1,0,0,0,0,0,0,0,1]
expected output:
# D
signal = [1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1,0,0,0,0,0,0,0,1,1,0,0,0]
output = [0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0]
# ^ cycle.1:day.1 |dayline ^cycle.1:day.2
In the output list, when there is 1 that means 1 cycle is completed at that position of the signal list and at rest of the position there are 0. There should only 1 cycle in a days that's why only 1 is there.
I don't how to split this list according to that so can someone please help?

It seams to me like what you are trying to do is split your data first into blocks of 24, and then to find either the first rising edge, or the first falling edge depending on the first hour in that block.
Below I have tried to distill my understanding of what you are trying to accomplish into the following function. It takes in a numpy.array containing zeros and ones, as in your example. It checks to see what the first hour in the day is, and decides what type of edge to look for.
it detects an edge by using np.diff. This gives us an array containing -1's, 0's, and 1's. We then look for the first index of either a -1 falling edge, or 1 rising edge. The function returns that index, or if no edges were found it returns the index of the last element, or nothing.
For more info see the docs for descriptions on numpy features used here np.diff, np.array.nonzero, np.array_split
import numpy as np
def get_cycle_index(day):
'''
returns the first index of a cycle defined by nipun vats
if no cycle is found returns nothing
'''
first_hour = day[0]
if first_hour == 0:
edgetype = -1
else:
edgetype = 1
edges = np.diff(np.r_[day, day[-1]])
if (edges == edgetype).any():
return (edges == edgetype).nonzero()[0][0]
elif (day.sum() == day.size) or day.sum() == 0:
return
else:
return day.size - 1
Below is an example of how you might use this function in your case.
import numpy as np
_data = [1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
#_data = np.random.randint(0,2,280, dtype='int')
data = np.array(_data, 'int')
#split the data into a set of 'day' blocks
blocks = np.array_split(data, np.arange(24,data.size, 24))
_output = []
for i, day in enumerate(blocks):
print(f'day {i}')
buffer = np.zeros(day.size, dtype='int')
print('\tsignal:', *day, sep = ' ')
cycle_index = get_cycle_index(day)
if cycle_index:
buffer[cycle_index] = 1
print('\toutput:', *buffer, sep=' ')
_output.append(buffer)
output = np.concatenate(_output)
print('\nfinal output:\n', *output, sep=' ')
this yeilds the following output:
day 0
signal: 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 1 1 1 1 1 0
output: 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
day 1
signal: 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
output: 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
day 2
signal: 0 0 0 0 0 0
output: 0 0 0 0 0 0
final output:
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

How to multiply every column of one Pandas Dataframe with every column of another Dataframe efficiently?

I'm trying to multiply two pandas dataframes with each other. Specifically, I want to multiply every column with every column of the other df.
The dataframes are one-hot encoded, so they look like this:
col_1, col_2, col_3, ...
0 1 0
1 0 0
0 0 1
...
I could just iterate through each of the columns using a for loop, but in python that is computationally expensive, and I'm hoping there's an easier way.
One of the dataframes has 500 columns, the other has 100 columns.
This is the fastest version that I've been able to write so far:
interact_pd = pd.DataFrame(index=df_1.index)
df1_columns = [column for column in df_1]
for column in df_2:
col_pd = df_1[df1_columns].multiply(df_2[column], axis="index")
interact_pd = interact_pd.join(col_pd, lsuffix='_' + column)
I iterate over each column in df_2 and multiply all of df_1 by that column, then I append the result to interact_pd. I would rather not do it using a for loop however, as this is very computationally costly. Is there a faster way of doing it?
EDIT: example
df_1:
1col_1, 1col_2, 1col_3
0 1 0
1 0 0
0 0 1
df_2:
2col_1, 2col_2
0 1
1 0
0 0
interact_pd:
1col_1_2col_1, 1col_2_2col_1,1col_3_2col_1, 1col_1_2col_2, 1col_2_2col_2,1col_3_2col_2
0 0 0 0 1 0
1 0 0 0 0 0
0 0 0 0 0 0

# use numpy to get a pair of indices that map out every
# combination of columns from df_1 and columns of df_2
pidx = np.indices((df_1.shape[1], df_2.shape[1])).reshape(2, -1)
# use pandas MultiIndex to create a nice MultiIndex for
# the final output
lcol = pd.MultiIndex.from_product([df_1.columns, df_2.columns],
names=[df_1.columns.name, df_2.columns.name])
# df_1.values[:, pidx[0]] slices df_1 values for every combination
# like wise with df_2.values[:, pidx[1]]
# finally, I marry up the product of arrays with the MultiIndex
pd.DataFrame(df_1.values[:, pidx[0]] * df_2.values[:, pidx[1]],
columns=lcol)
Timing
code
from string import ascii_letters
df_1 = pd.DataFrame(np.random.randint(0, 2, (1000, 26)), columns=list(ascii_letters[:26]))
df_2 = pd.DataFrame(np.random.randint(0, 2, (1000, 52)), columns=list(ascii_letters))
def pir1(df_1, df_2):
pidx = np.indices((df_1.shape[1], df_2.shape[1])).reshape(2, -1)
lcol = pd.MultiIndex.from_product([df_1.columns, df_2.columns],
names=[df_1.columns.name, df_2.columns.name])
return pd.DataFrame(df_1.values[:, pidx[0]] * df_2.values[:, pidx[1]],
columns=lcol)
def Test2(DA,DB):
MA = DA.as_matrix()
MB = DB.as_matrix()
MM = np.zeros((len(MA),len(MA[0])*len(MB[0])))
Col = []
for i in range(len(MB[0])):
for j in range(len(MA[0])):
MM[:,i*len(MA[0])+j] = MA[:,j]*MB[:,i]
Col.append('1col_'+str(i+1)+'_2col_'+str(j+1))
return pd.DataFrame(MM,dtype=int,columns=Col)
results

You can multiply along the index axis your first df with each column of the second df, this is the fastest method for big datasets (see below):
df = pd.concat([df_1.mul(col[1], axis="index") for col in df_2.iteritems()], axis=1)
# Change the name of the columns
df.columns = ["_".join([i, j]) for j in df_2.columns for i in df_1.columns]
df
1col_1_2col_1 1col_2_2col_1 1col_3_2col_1 1col_1_2col_2 \
0 0 0 0 0
1 1 0 0 0
2 0 0 0 0
1col_2_2col_2 1col_3_2col_2
0 1 0
1 0 0
2 0 0
--> See benchmark for comparisons with other answers to choose the best option for your dataset.
Benchmark
Functions:
def Test2(DA,DB):
MA = DA.as_matrix()
MB = DB.as_matrix()
MM = np.zeros((len(MA),len(MA[0])*len(MB[0])))
Col = []
for i in range(len(MB[0])):
for j in range(len(MA[0])):
MM[:,i*len(MA[0])+j] = MA[:,j]*MB[:,i]
Col.append('1col_'+str(i+1)+'_2col_'+str(j+1))
return pd.DataFrame(MM,dtype=int,columns=Col)
def Test3(df_1, df_2):
df = pd.concat([df_1.mul(i[1], axis="index") for i in df_2.iteritems()], axis=1)
df.columns = ["_".join([i,j]) for j in df_2.columns for i in df_1.columns]
return df
def Test4(df_1,df_2):
pidx = np.indices((df_1.shape[1], df_2.shape[1])).reshape(2, -1)
lcol = pd.MultiIndex.from_product([df_1.columns, df_2.columns],
names=[df_1.columns.name, df_2.columns.name])
return pd.DataFrame(df_1.values[:, pidx[0]] * df_2.values[:, pidx[1]],
columns=lcol)
def jeanrjc_imp(df_1, df_2):
df = pd.concat([df_1.mul(‌​i[1], axis="index") for i in df_2.iteritems()], axis=1, keys=df_2.columns)
return df
Code:
Sorry, ugly code, the plot at the end matters :
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
df_1 = pd.DataFrame(np.random.randint(0, 2, (1000, 600)))
df_2 = pd.DataFrame(np.random.randint(0, 2, (1000, 600)))
df_1.columns = ["1col_"+str(i) for i in range(len(df_1.columns))]
df_2.columns = ["2col_"+str(i) for i in range(len(df_2.columns))]
resa = {}
resb = {}
resc = {}
for f, r in zip([Test2, Test3, Test4, jeanrjc_imp], ["T2", "T3", "T4", "T3bis"]):
resa[r] = []
resb[r] = []
resc[r] = []
for i in [5, 10, 30, 50, 150, 200]:
a = %timeit -o f(df_1.iloc[:,:i], df_2.iloc[:, :10])
b = %timeit -o f(df_1.iloc[:,:i], df_2.iloc[:, :50])
c = %timeit -o f(df_1.iloc[:,:i], df_2.iloc[:, :200])
resa[r].append(a.best)
resb[r].append(b.best)
resc[r].append(c.best)
X = [5, 10, 30, 50, 150, 200]
fig, ax = plt.subplots(1, 3, figsize=[16,5])
for j, (a, r) in enumerate(zip(ax, [resa, resb, resc])):
for i in r:
a.plot(X, r[i], label=i)
a.set_xlabel("df_1 columns #")
a.set_title("df_2 columns # = {}".format(["10", "50", "200"][j]))
ax[0].set_ylabel("time(s)")
plt.legend(loc=0)
plt.tight_layout()
With T3b <=> jeanrjc_imp. Which is a bit faster that Test3.
Conclusion:
Depending on your dataset size, pick the right function, between Test4 and Test3(b). Given the OP's dataset, Test3 or jeanrjc_imp should be the fastest, and also the shortest to write!
HTH

You can use numpy.
Consider this example code, I did modify the variable names, but Test1() is essentially your code. I didn't bother create the correct column names in that function though:
import pandas as pd
import numpy as np
A = [[1,0,1,1],[0,1,1,0],[0,1,0,1]]
B = [[0,0,1,0],[1,0,1,0],[1,1,0,0],[1,0,0,1],[1,0,0,0]]
DA = pd.DataFrame(A).T
DB = pd.DataFrame(B).T
def Test1(DA,DB):
E = pd.DataFrame(index=DA.index)
DAC = [column for column in DA]
for column in DB:
C = DA[DAC].multiply(DB[column], axis="index")
E = E.join(C, lsuffix='_' + str(column))
return E
def Test2(DA,DB):
MA = DA.as_matrix()
MB = DB.as_matrix()
MM = np.zeros((len(MA),len(MA[0])*len(MB[0])))
Col = []
for i in range(len(MB[0])):
for j in range(len(MA[0])):
MM[:,i*len(MA[0])+j] = MA[:,j]*MB[:,i]
Col.append('1col_'+str(i+1)+'_2col_'+str(j+1))
return pd.DataFrame(MM,dtype=int,columns=Col)
print Test1(DA,DB)
print Test2(DA,DB)
Output:
0_1 1_1 2_1 0 1 2 0_3 1_3 2_3 0 1 2 0 1 2
0 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0
1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0
2 1 1 0 1 1 0 0 0 0 0 0 0 0 0 0
3 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0
1col_1_2col_1 1col_1_2col_2 1col_1_2col_3 1col_2_2col_1 1col_2_2col_2 \
0 0 0 0 1 0
1 0 0 0 0 0
2 1 1 0 1 1
3 0 0 0 0 0
1col_2_2col_3 1col_3_2col_1 1col_3_2col_2 1col_3_2col_3 1col_4_2col_1 \
0 0 1 0 0 1
1 0 0 1 1 0
2 0 0 0 0 0
3 0 0 0 0 1
1col_4_2col_2 1col_4_2col_3 1col_5_2col_1 1col_5_2col_2 1col_5_2col_3
0 0 0 1 0 0
1 0 0 0 0 0
2 0 0 0 0 0
3 0 1 0 0 0
Performance of your function:
%timeit(Test1(DA,DB))
100 loops, best of 3: 11.1 ms per loop
Performance of my function:
%timeit(Test2(DA,DB))
1000 loops, best of 3: 464 µs per loop
It's not beautiful, but it's efficient.

python, read '.dat' file with differents columns for each lines

I need to extract some data from .dat file which I usually do with
import numpy as np
file = np.loadtxt('blablabla.dat')
Here my data are not separated by a specific delimiter but have predefined length (digits) and some lines don't have any values for some columns.
Here an sample to be clear :
3 0 36 0 0 0 0 0 0 0 99.
-2 0 0 0 0 0 0 0 0 0 99.
2 0 0 0 0 0 0 0 0 0 .LA.0?. 3.
5 0 0 0 0 2 4 0 0 0 .SAS7?. 99.
-5 0 0 0 0 0 0 0 0 0 99.
99 0 0 0 0 0 0 0 0 0 .S..3*. 3.5
My little code above get the error :
# Convert each value according to its column and store
ValueError: Wrong number of columns at line 3
Does someone have an idea about how to collect this kind of data?

numpy.genfromtxt seems to be what you want; it you can specify field widths for each column and treats missing data as NaNs.
For this case:
import numpy as np
data = np.genfromtxt('blablabla.dat',delimiter=[2,3,4,3,3,2,3,4,5,3,8,5])
If you want to keep information in the string part of the file, you could read twice and specify the usecols parameter:
import numpy as np
number_data = np.genfromtxt('blablabla.dat',delimiter=[2,3,4,3,3,2,3,4,5,3,8,5],\
usecols=(0,1,2,3,4,5,6,7,8,9,11))
string_data = np.genfromtxt('blablabla.dat',delimiter=[2,3,4,3,3,2,3,4,5,3,8,5],\
usecols=(10),dtype=str)

What you essentially need is to get list of empty "columns" position that serve as delimiters
That will get you started
In [108]: table = ''' 3 0 36 0 0 0 0 0 0 0 99.
.....: -2 0 0 0 0 0 0 0 0 0 99.
.....: 2 0 0 0 0 0 0 0 0 0 .LA.0?. 3.
.....: 5 0 0 0 0 2 4 0 0 0 .SAS7?. 99.
.....: -5 0 0 0 0 0 0 0 0 0 99.
.....: 99 0 0 0 0 0 0 0 0 0 .S..3*. 3.5'''.split('\n')
In [110]: max_row_len = max(len(row) for row in table)
In [117]: spaces = reduce(lambda res, row: res.intersection(idx for idx, c in enumerate(row) if c == ' '), table, set(range(max_row_len)))
This code builds set of character positions in the longest row - and reduce leaves only set of positions that have spaces in all rows