I have a 4D array containing daily time-series of gridded data for different years with shape (year, day, x-coordinate, y-coordinate). The actual shape of my array is (19, 133, 288, 620), so I have 19 years of data with 133 days per year over a 288 x 620 grid. I want to take the weekly average of each grid cell over the period of record. The shape of the weekly averaged array should be (19, 19, 288, 620), or (year, week, x-coordinate, y-coordinate). I would like to use numpy to achieve this.
Here I construct some dummy data to work with and an array of what the solution should be:
import numpy as np
a1 = np.arange(1, 10).reshape(3, 3)
a1days = np.repeat(a1[np.newaxis, ...], 7, axis=0)
b1 = np.arange(10, 19).reshape(3, 3)
b1days = np.repeat(b1[np.newaxis, ...], 7, axis=0)
c1year = np.concatenate((a1days, b1days), axis=0)
a2 = np.arange(19, 28).reshape(3, 3)
a2days = np.repeat(a2[np.newaxis, ...], 7, axis=0)
b2 = np.arange(29, 38).reshape(3, 3)
b2days = np.repeat(b2[np.newaxis, ...], 7, axis=0)
c2year = np.concatenate((a2days, b2days), axis=0)
dummy_data = np.concatenate((c1year, c2year), axis=0).reshape(2, 14, 3, 3)
solution = np.concatenate((a1, b1, a2, b2), axis=0).reshape(2, 2, 3, 3)
The shape of the dummy_data is (2, 14, 3, 3). Per the dummy data, I have two years of data, 14 days per year, over a 3 X 3 grid. I want to return the weekly average of the grid for both years, resulting in a solution with shape (2, 2, 3, 3).
You can reshape and take mean:
week_mean = dummy_data.reshape(2,-1,7,3,3).mean(axis=2)
# in your case .reshape(year, -1, 7, x_coord, y_coord)
# check:
(dummy_data.reshape(2,2,7,3,3).mean(axis=2) == solution).all()
# True
How do I create an empty DataFrame, then add rows, one by one?
I created an empty DataFrame:
df = pd.DataFrame(columns=('lib', 'qty1', 'qty2'))
Then I can add a new row at the end and fill a single field with:
df = df._set_value(index=len(df), col='qty1', value=10.0)
It works for only one field at a time. What is a better way to add new row to df?
You can use df.loc[i], where the row with index i will be what you specify it to be in the dataframe.
>>> import pandas as pd
>>> from numpy.random import randint
>>> df = pd.DataFrame(columns=['lib', 'qty1', 'qty2'])
>>> for i in range(5):
>>> df.loc[i] = ['name' + str(i)] + list(randint(10, size=2))
>>> df
lib qty1 qty2
0 name0 3 3
1 name1 2 4
2 name2 2 8
3 name3 2 1
4 name4 9 6
In case you can get all data for the data frame upfront, there is a much faster approach than appending to a data frame:
Create a list of dictionaries in which each dictionary corresponds to an input data row.
Create a data frame from this list.
I had a similar task for which appending to a data frame row by row took 30 min, and creating a data frame from a list of dictionaries completed within seconds.
rows_list = []
for row in input_rows:
dict1 = {}
# get input row in dictionary format
# key = col_name
dict1.update(blah..)
rows_list.append(dict1)
df = pd.DataFrame(rows_list)
In the case of adding a lot of rows to dataframe, I am interested in performance. So I tried the four most popular methods and checked their speed.
Performance
Using .append (NPE's answer)
Using .loc (fred's answer)
Using .loc with preallocating (FooBar's answer)
Using dict and create DataFrame in the end (ShikharDua's answer)
Runtime results (in seconds):
Approach
1000 rows
5000 rows
10 000 rows
.append
0.69
3.39
6.78
.loc without prealloc
0.74
3.90
8.35
.loc with prealloc
0.24
2.58
8.70
dict
0.012
0.046
0.084
So I use addition through the dictionary for myself.
Code:
import pandas as pd
import numpy as np
import time
del df1, df2, df3, df4
numOfRows = 1000
# append
startTime = time.perf_counter()
df1 = pd.DataFrame(np.random.randint(100, size=(5,5)), columns=['A', 'B', 'C', 'D', 'E'])
for i in range( 1,numOfRows-4):
df1 = df1.append( dict( (a,np.random.randint(100)) for a in ['A','B','C','D','E']), ignore_index=True)
print('Elapsed time: {:6.3f} seconds for {:d} rows'.format(time.perf_counter() - startTime, numOfRows))
print(df1.shape)
# .loc w/o prealloc
startTime = time.perf_counter()
df2 = pd.DataFrame(np.random.randint(100, size=(5,5)), columns=['A', 'B', 'C', 'D', 'E'])
for i in range( 1,numOfRows):
df2.loc[i] = np.random.randint(100, size=(1,5))[0]
print('Elapsed time: {:6.3f} seconds for {:d} rows'.format(time.perf_counter() - startTime, numOfRows))
print(df2.shape)
# .loc with prealloc
df3 = pd.DataFrame(index=np.arange(0, numOfRows), columns=['A', 'B', 'C', 'D', 'E'] )
startTime = time.perf_counter()
for i in range( 1,numOfRows):
df3.loc[i] = np.random.randint(100, size=(1,5))[0]
print('Elapsed time: {:6.3f} seconds for {:d} rows'.format(time.perf_counter() - startTime, numOfRows))
print(df3.shape)
# dict
startTime = time.perf_counter()
row_list = []
for i in range (0,5):
row_list.append(dict( (a,np.random.randint(100)) for a in ['A','B','C','D','E']))
for i in range( 1,numOfRows-4):
dict1 = dict( (a,np.random.randint(100)) for a in ['A','B','C','D','E'])
row_list.append(dict1)
df4 = pd.DataFrame(row_list, columns=['A','B','C','D','E'])
print('Elapsed time: {:6.3f} seconds for {:d} rows'.format(time.perf_counter() - startTime, numOfRows))
print(df4.shape)
P.S.: I believe my realization isn't perfect, and maybe there is some optimization that could be done.
You could use pandas.concat(). For details and examples, see Merge, join, and concatenate.
For example:
def append_row(df, row):
return pd.concat([
df,
pd.DataFrame([row], columns=row.index)]
).reset_index(drop=True)
df = pd.DataFrame(columns=('lib', 'qty1', 'qty2'))
new_row = pd.Series({'lib':'A', 'qty1':1, 'qty2': 2})
df = append_row(df, new_row)
NEVER grow a DataFrame!
Yes, people have already explained that you should NEVER grow a DataFrame, and that you should append your data to a list and convert it to a DataFrame once at the end. But do you understand why?
Here are the most important reasons, taken from my post here.
It is always cheaper/faster to append to a list and create a DataFrame in one go.
Lists take up less memory and are a much lighter data structure to work with, append, and remove.
dtypes are automatically inferred for your data. On the flip side, creating an empty frame of NaNs will automatically make them object, which is bad.
An index is automatically created for you, instead of you having to take care to assign the correct index to the row you are appending.
This is The Right Way™ to accumulate your data
data = []
for a, b, c in some_function_that_yields_data():
data.append([a, b, c])
df = pd.DataFrame(data, columns=['A', 'B', 'C'])
These options are horrible
append or concat inside a loop
append and concat aren't inherently bad in isolation. The
problem starts when you iteratively call them inside a loop - this
results in quadratic memory usage.
# Creates empty DataFrame and appends
df = pd.DataFrame(columns=['A', 'B', 'C'])
for a, b, c in some_function_that_yields_data():
df = df.append({'A': i, 'B': b, 'C': c}, ignore_index=True)
# This is equally bad:
# df = pd.concat(
# [df, pd.Series({'A': i, 'B': b, 'C': c})],
# ignore_index=True)
Empty DataFrame of NaNs
Never create a DataFrame of NaNs as the columns are initialized with
object (slow, un-vectorizable dtype).
# Creates DataFrame of NaNs and overwrites values.
df = pd.DataFrame(columns=['A', 'B', 'C'], index=range(5))
for a, b, c in some_function_that_yields_data():
df.loc[len(df)] = [a, b, c]
The Proof is in the Pudding
Timing these methods is the fastest way to see just how much they differ in terms of their memory and utility.
Benchmarking code for reference.
It's posts like this that remind me why I'm a part of this community. People understand the importance of teaching folks getting the right answer with the right code, not the right answer with wrong code. Now you might argue that it is not an issue to use loc or append if you're only adding a single row to your DataFrame. However, people often look to this question to add more than just one row - often the requirement is to iteratively add a row inside a loop using data that comes from a function (see related question). In that case it is important to understand that iteratively growing a DataFrame is not a good idea.
If you know the number of entries ex ante, you should preallocate the space by also providing the index (taking the data example from a different answer):
import pandas as pd
import numpy as np
# we know we're gonna have 5 rows of data
numberOfRows = 5
# create dataframe
df = pd.DataFrame(index=np.arange(0, numberOfRows), columns=('lib', 'qty1', 'qty2') )
# now fill it up row by row
for x in np.arange(0, numberOfRows):
#loc or iloc both work here since the index is natural numbers
df.loc[x] = [np.random.randint(-1,1) for n in range(3)]
In[23]: df
Out[23]:
lib qty1 qty2
0 -1 -1 -1
1 0 0 0
2 -1 0 -1
3 0 -1 0
4 -1 0 0
Speed comparison
In[30]: %timeit tryThis() # function wrapper for this answer
In[31]: %timeit tryOther() # function wrapper without index (see, for example, #fred)
1000 loops, best of 3: 1.23 ms per loop
100 loops, best of 3: 2.31 ms per loop
And - as from the comments - with a size of 6000, the speed difference becomes even larger:
Increasing the size of the array (12) and the number of rows (500) makes
the speed difference more striking: 313ms vs 2.29s
mycolumns = ['A', 'B']
df = pd.DataFrame(columns=mycolumns)
rows = [[1,2],[3,4],[5,6]]
for row in rows:
df.loc[len(df)] = row
You can append a single row as a dictionary using the ignore_index option.
>>> f = pandas.DataFrame(data = {'Animal':['cow','horse'], 'Color':['blue', 'red']})
>>> f
Animal Color
0 cow blue
1 horse red
>>> f.append({'Animal':'mouse', 'Color':'black'}, ignore_index=True)
Animal Color
0 cow blue
1 horse red
2 mouse black
For efficient appending, see How to add an extra row to a pandas dataframe and Setting With Enlargement.
Add rows through loc/ix on non existing key index data. For example:
In [1]: se = pd.Series([1,2,3])
In [2]: se
Out[2]:
0 1
1 2
2 3
dtype: int64
In [3]: se[5] = 5.
In [4]: se
Out[4]:
0 1.0
1 2.0
2 3.0
5 5.0
dtype: float64
Or:
In [1]: dfi = pd.DataFrame(np.arange(6).reshape(3,2),
.....: columns=['A','B'])
.....:
In [2]: dfi
Out[2]:
A B
0 0 1
1 2 3
2 4 5
In [3]: dfi.loc[:,'C'] = dfi.loc[:,'A']
In [4]: dfi
Out[4]:
A B C
0 0 1 0
1 2 3 2
2 4 5 4
In [5]: dfi.loc[3] = 5
In [6]: dfi
Out[6]:
A B C
0 0 1 0
1 2 3 2
2 4 5 4
3 5 5 5
For the sake of a Pythonic way:
res = pd.DataFrame(columns=('lib', 'qty1', 'qty2'))
res = res.append([{'qty1':10.0}], ignore_index=True)
print(res.head())
lib qty1 qty2
0 NaN 10.0 NaN
You can also build up a list of lists and convert it to a dataframe -
import pandas as pd
columns = ['i','double','square']
rows = []
for i in range(6):
row = [i, i*2, i*i]
rows.append(row)
df = pd.DataFrame(rows, columns=columns)
giving
i double square
0 0 0 0
1 1 2 1
2 2 4 4
3 3 6 9
4 4 8 16
5 5 10 25
If you always want to add a new row at the end, use this:
df.loc[len(df)] = ['name5', 9, 0]
I figured out a simple and nice way:
>>> df
A B C
one 1 2 3
>>> df.loc["two"] = [4,5,6]
>>> df
A B C
one 1 2 3
two 4 5 6
Note the caveat with performance as noted in the comments.
This is not an answer to the OP question, but a toy example to illustrate ShikharDua's answer which I found very useful.
While this fragment is trivial, in the actual data I had 1,000s of rows, and many columns, and I wished to be able to group by different columns and then perform the statistics below for more than one target column. So having a reliable method for building the data frame one row at a time was a great convenience. Thank you ShikharDua!
import pandas as pd
BaseData = pd.DataFrame({ 'Customer' : ['Acme','Mega','Acme','Acme','Mega','Acme'],
'Territory' : ['West','East','South','West','East','South'],
'Product' : ['Econ','Luxe','Econ','Std','Std','Econ']})
BaseData
columns = ['Customer','Num Unique Products', 'List Unique Products']
rows_list=[]
for name, group in BaseData.groupby('Customer'):
RecordtoAdd={} #initialise an empty dict
RecordtoAdd.update({'Customer' : name}) #
RecordtoAdd.update({'Num Unique Products' : len(pd.unique(group['Product']))})
RecordtoAdd.update({'List Unique Products' : pd.unique(group['Product'])})
rows_list.append(RecordtoAdd)
AnalysedData = pd.DataFrame(rows_list)
print('Base Data : \n',BaseData,'\n\n Analysed Data : \n',AnalysedData)
You can use a generator object to create a Dataframe, which will be more memory efficient over the list.
num = 10
# Generator function to generate generator object
def numgen_func(num):
for i in range(num):
yield ('name_{}'.format(i), (i*i), (i*i*i))
# Generator expression to generate generator object (Only once data get populated, can not be re used)
numgen_expression = (('name_{}'.format(i), (i*i), (i*i*i)) for i in range(num) )
df = pd.DataFrame(data=numgen_func(num), columns=('lib', 'qty1', 'qty2'))
To add raw to existing DataFrame you can use append method.
df = df.append([{ 'lib': "name_20", 'qty1': 20, 'qty2': 400 }])
Instead of a list of dictionaries as in ShikharDua's answer (row-based), we can also represent our table as a dictionary of lists (column-based), where each list stores one column in row-order, given we know our columns beforehand. At the end we construct our DataFrame once.
In both cases, the dictionary keys are always the column names. Row order is stored implicitly as order in a list. For c columns and n rows, this uses one dictionary of c lists, versus one list of n dictionaries. The list-of-dictionaries method has each dictionary storing all keys redundantly and requires creating a new dictionary for every row. Here we only append to lists, which overall is the same time complexity (adding entries to list and dictionary are both amortized constant time) but may have less overhead due to being a simple operation.
# Current data
data = {"Animal":["cow", "horse"], "Color":["blue", "red"]}
# Adding a new row (be careful to ensure every column gets another value)
data["Animal"].append("mouse")
data["Color"].append("black")
# At the end, construct our DataFrame
df = pd.DataFrame(data)
# Animal Color
# 0 cow blue
# 1 horse red
# 2 mouse black
Create a new record (data frame) and add to old_data_frame.
Pass a list of values and the corresponding column names to create a new_record (data_frame):
new_record = pd.DataFrame([[0, 'abcd', 0, 1, 123]], columns=['a', 'b', 'c', 'd', 'e'])
old_data_frame = pd.concat([old_data_frame, new_record])
Here is the way to add/append a row in a Pandas DataFrame:
def add_row(df, row):
df.loc[-1] = row
df.index = df.index + 1
return df.sort_index()
add_row(df, [1,2,3])
It can be used to insert/append a row in an empty or populated Pandas DataFrame.
If you want to add a row at the end, append it as a list:
valuestoappend = [va1, val2, val3]
res = res.append(pd.Series(valuestoappend, index = ['lib', 'qty1', 'qty2']), ignore_index = True)
Another way to do it (probably not very performant):
# add a row
def add_row(df, row):
colnames = list(df.columns)
ncol = len(colnames)
assert ncol == len(row), "Length of row must be the same as width of DataFrame: %s" % row
return df.append(pd.DataFrame([row], columns=colnames))
You can also enhance the DataFrame class like this:
import pandas as pd
def add_row(self, row):
self.loc[len(self.index)] = row
pd.DataFrame.add_row = add_row
All you need is loc[df.shape[0]] or loc[len(df)]
# Assuming your df has 4 columns (str, int, str, bool)
df.loc[df.shape[0]] = ['col1Value', 100, 'col3Value', False]
or
df.loc[len(df)] = ['col1Value', 100, 'col3Value', False]
You can concatenate two DataFrames for this. I basically came across this problem to add a new row to an existing DataFrame with a character index (not numeric).
So, I input the data for a new row in a duct() and index in a list.
new_dict = {put input for new row here}
new_list = [put your index here]
new_df = pd.DataFrame(data=new_dict, index=new_list)
df = pd.concat([existing_df, new_df])
initial_data = {'lib': np.array([1,2,3,4]), 'qty1': [1,2,3,4], 'qty2': [1,2,3,4]}
df = pd.DataFrame(initial_data)
df
lib qty1 qty2
0 1 1 1
1 2 2 2
2 3 3 3
3 4 4 4
val_1 = [10]
val_2 = [14]
val_3 = [20]
df.append(pd.DataFrame({'lib': val_1, 'qty1': val_2, 'qty2': val_3}))
lib qty1 qty2
0 1 1 1
1 2 2 2
2 3 3 3
3 4 4 4
0 10 14 20
You can use a for loop to iterate through values or can add arrays of values.
val_1 = [10, 11, 12, 13]
val_2 = [14, 15, 16, 17]
val_3 = [20, 21, 22, 43]
df.append(pd.DataFrame({'lib': val_1, 'qty1': val_2, 'qty2': val_3}))
lib qty1 qty2
0 1 1 1
1 2 2 2
2 3 3 3
3 4 4 4
0 10 14 20
1 11 15 21
2 12 16 22
3 13 17 43
Make it simple. By taking a list as input which will be appended as a row in the data-frame:
import pandas as pd
res = pd.DataFrame(columns=('lib', 'qty1', 'qty2'))
for i in range(5):
res_list = list(map(int, input().split()))
res = res.append(pd.Series(res_list, index=['lib', 'qty1', 'qty2']), ignore_index=True)
pandas.DataFrame.append
DataFrame.append(self, other, ignore_index=False, verify_integrity=False, sort=False) → 'DataFrame'
Code
df = pd.DataFrame([[1, 2], [3, 4]], columns=list('AB'))
df2 = pd.DataFrame([[5, 6], [7, 8]], columns=list('AB'))
df.append(df2)
With ignore_index set to True:
df.append(df2, ignore_index=True)
If you have a data frame df and want to add a list new_list as a new row to df, you can simply do:
df.loc[len(df)] = new_list
If you want to add a new data frame new_df under data frame df, then you can use:
df.append(new_df)
We often see the construct df.loc[subscript] = … to assign to one DataFrame row. Mikhail_Sam posted benchmarks containing, among others, this construct as well as the method using dict and create DataFrame in the end. He found the latter to be the fastest by far.
But if we replace the df3.loc[i] = … (with preallocated DataFrame) in his code with df3.values[i] = …, the outcome changes significantly, in that that method performs similar to the one using dict. So we should more often take the use of df.values[subscript] = … into consideration. However note that .values takes a zero-based subscript, which may be different from the DataFrame.index.
Before going to add a row, we have to convert the dataframe to a dictionary. There you can see the keys as columns in the dataframe and the values of the columns are again stored in the dictionary, but there the key for every column is the index number in the dataframe.
That idea makes me to write the below code.
df2 = df.to_dict()
values = ["s_101", "hyderabad", 10, 20, 16, 13, 15, 12, 12, 13, 25, 26, 25, 27, "good", "bad"] # This is the total row that we are going to add
i = 0
for x in df.columns: # Here df.columns gives us the main dictionary key
df2[x][101] = values[i] # Here the 101 is our index number. It is also the key of the sub dictionary
i += 1
If all data in your Dataframe has the same dtype you might use a NumPy array. You can write rows directly into the predefined array and convert it to a dataframe at the end.
It seems to be even faster than converting a list of dicts.
import pandas as pd
import numpy as np
from string import ascii_uppercase
startTime = time.perf_counter()
numcols, numrows = 5, 10000
npdf = np.ones((numrows, numcols))
for row in range(numrows):
npdf[row, 0:] = np.random.randint(0, 100, (1, numcols))
df5 = pd.DataFrame(npdf, columns=list(ascii_uppercase[:numcols]))
print('Elapsed time: {:6.3f} seconds for {:d} rows'.format(time.perf_counter() - startTime, numOfRows))
print(df5.shape)
This code snippet uses a list of dictionaries to update the data frame. It adds on to ShikharDua's and Mikhail_Sam's answers.
import pandas as pd
colour = ["red", "big", "tasty"]
fruits = ["apple", "banana", "cherry"]
dict1={}
feat_list=[]
for x in colour:
for y in fruits:
# print(x, y)
dict1 = dict([('x',x),('y',y)])
# print(f'dict 1 {dict1}')
feat_list.append(dict1)
# print(f'feat_list {feat_list}')
feat_df=pd.DataFrame(feat_list)
feat_df.to_csv('feat1.csv')
I have the following problem: I have a matrix opened with pandas module, where each cell has a number between -1 and 1. What I wanted to find is the maximum "posible" value in a row that is also not the maximum value in another row.
If for example 2 rows has their maximum value at the same column, I compare both values and take the bigger one, then for the row that has its maximum value smaller that the other row, I took the second maximum value (and do the same analysis again and again).
To explain myself better consider my code
import pandas as pd
matrix = pd.read_csv("matrix.csv")
# this matrix has an id (or name) for each column
# ... and the firt column has the id of each row
results = pd.DataFrame(np.empty((len(matrix),3),dtype=pd.Timestamp),columns=['id1','id2','max_pos'])
l = len(matrix.col[[0]]) # number of columns
while next = 1:
next = 0
for i in range(0, len(matrix)):
max_column = str(0)
for j in range(1, l): # 1 because the first column is an id
if matrix[max_column][i] < matrix[str(j)][i]:
max_column = str(j)
results['id1'][i] = str(i) # I coul put here also matrix['0'][i]
results['id2'][i] = max_column
results['max_pos'][i] = matrix[max_column][i]
for i in range(0, len(results)): #now I will check if two or more rows have the same max column
for ii in range(0, len(results)):
# if two id1 has their max in the same column, I keep it with the biggest
# ... max value and chage the other to "-1" to iterate again
if (results['id2'][i] == results['id2'][ii]) and (results['max_pos'][i] < results['max_pos'][ii]):
matrix[results['id2'][i]][i] = -1
next = 1
Putting an example:
#consider
pd.DataFrame({'a':[1, 2, 5, 0], 'b':[4, 5, 1, 0], 'c':[3, 3, 4, 2], 'd':[1, 0, 0, 1]})
a b c d
0 1 4 3 1
1 2 5 3 0
2 5 1 4 0
3 0 0 2 1
#at the first iterarion I will have the following result
0 b 4 # this means that the row 0 has its maximum at column 'b' and its value is 4
1 b 5
2 a 5
3 c 2
#the problem is that column b is the maximum of row 0 and 1, but I know that the maximum of row 1 is bigger than row 0, so I take the second maximum of row 0, then:
0 c 3
1 b 5
2 a 5
3 c 2
#now I solved the problem for row 0 and 1, but I have that the column c is the maximum of row 0 and 3, so I compare them and take the second maximum in row 3
0 c 3
1 b 5
2 a 5
3 d 1
#now I'm done. In the case that two rows have the same column as maximum and also the same number, nothing happens and I keep with that values.
#what if the matrix would be
pd.DataFrame({'a':[1, 2, 5, 0], 'b':[5, 5, 1, 0], 'c':[3, 3, 4, 2], 'd':[1, 0, 0, 1]})
a b c d
0 1 5 3 1
1 2 5 3 0
2 5 1 4 0
3 0 0 2 1
#then, at the first itetarion the result will be:
0 b 5
1 b 5
2 a 5
3 c 2
#then, given that the max value of row 0 and 1 is at the same column, I should compare the maximum values
# ... but in this case the values are the same (both are 5), this would be the end of iterating
# ... because I can't choose between row 0 and 1 and the other rows have their maximum at different columns...
This code works perfect to me if I have a matrix of 100x100 for example. But, if the matrix size goes to 50,000x50,000 the code takes to much time in finish it. I now that my code could be the most inneficient way to do it, but I don't know how to deal with this.
I have been reading about threads in python that could help but it doesn't help if I put 50,000 threads because my computer doesn't use more CPU. I also tried to use some functions as .max() but I'm not able to get column of the max an compare it with the other max ...
If anyone could help me of give me a piece of advice to make this more efficient I would be very grateful.
Going to need more information on this. What are you trying to accomplish here?
This will help you get some of the way, but in order to fully achieve what you're doing I need more context.
We'll import numpy, random, and Counter from collections:
import numpy as np
import random
from collections import Counter
We'll create a random 50k x 50k matrix of numbers between -10M and +10M
mat = np.random.randint(-10000000,10000000,(50000,50000))
Now to get the maximums for each row we can just do the following list comprehension:
maximums = [max(mat[x,:]) for x in range(len(mat))]
Now we want to find out which ones are not maximums in any other rows. We can use Counter on our maximums list to find out how many of each there are. Counter returns a counter object that is like a dictionary with the maximum as the key, and the # of times it appears as the value.
We then do dictionary comprehension where the value is == to 1. That will give us the maximums that only show up once. we use the .keys() function to grab the numbers themselves, and then turn it into a list.
c = Counter(maximums)
{9999117: 15,
9998584: 2,
9998352: 2,
9999226: 22,
9999697: 59,
9999534: 32,
9998775: 8,
9999288: 18,
9998956: 9,
9998119: 1,
...}
k = list( {x: c[x] for x in c if c[x] == 1}.keys() )
[9998253,
9998139,
9998091,
9997788,
9998166,
9998552,
9997711,
9998230,
9998000,
...]
Lastly we can do the following list comprehension to iterate through the original maximums list to get the indicies of where these rows are.
indices = [i for i, x in enumerate(maximums) if x in k]
Depending on what else you're looking to do we can go from here.
Its not the speediest program but finding the maximums, the counter, and the indicies takes 182 seconds on a 50,000 by 50,000 matrix that is already loaded.