I find my self trying to analyse a data set and find how some variables correlate.
I need to add a loop that adds a logical test to the if statement:
Edited:
Example:
Take this data frame as example
In [11]: df
Out[11]:
INPUT1 INPUT2 INPUT3 ... OUTPUT
0 8 5 6 ... 1
1 3 2 5 ... 0
2 3 1 5 ... 1
3 1 2 5 ... 0
4 4 3 5 ... 0
I'm testing the combinations of inputs to check how they match the output
def greater_than(a,b):
return a > b
def greater_equal_than(a,b):
return a >= b
def lower_equal_than(a,b):
return a <= b
def lower_than(a,b):
return a < b
def equal(a,b):
return a == b
operation = { '>': greater_than, '>=': greater_equal_than, '<=': lower_equal_than, '<': lower_than }
escenario = pd.DataFrame(columns=['esc','pf'])
for i in range(len(names)):
for j in names[i+1:]:
for op in operation:
escenario['esc'] = df.apply(lambda x : 1 if operation[op]( names[i], j ) else 0, axis=1)
escenario['pf'] = df['OUTPUT']
match = escenario.apply(lambda x : 1 if x['pf'] == 1 and x['pf'] == x['esc'] else 0, axis=1 )
percent_match = (100 * match.sum())/escenario['pf'].sum()
percent_no_match = (100 *(escenario['esc'].sum() - match.sum())) / escenario['esc'].sum()
print( f"{names[i]} {op} {j} -> { percent_match } / {percent_no_match} " )
I need to check all the combinations of input combinations that keeps percent_match closer to a 100% and percent_no_match closer to 0%
for example:
first iteration:
INPUT2 < INPUT3
SECOND INTERATION
INPUT2 < INPUT3 and INPUT1 > INPUT2
Right now I'm running the code, sorting the print and getting the couple where the match is closer to 100 and the modifying the code to add the match, Example:
First run better output is INPUT2 < INPUT3
Then I modify this line:
escenario['esc'] = df.apply(lambda x : 1 if operation[op]( names[i], j ) else 0, axis=1)
to add the first output, like:
escenario['esc'] = df.apply(lambda x : 1 if df['INPUT2'] < DF['INPUT3'] and operation[op]( names[i], j ) else 0, axis=1)
and check again...
This last part is the one I want to automate through a loop.
Thanks
I found self modifying python script that fits perfectly into my need.
It allows to recreate a function about of a text string and that's exactly what I need.
Thanks!
I have the following function that loads a csv into a data frame then does some calculations. It takes about 4-5 minutes to do calculation on the csv with a little over 100,000 lines. I was hoping there is a faster way.
def calculate_adeck_errors(in_file):
print(f'Starting Data Calculations: {datetime.datetime.now().strftime("%I:%M%p on %B %d, %Y")}')
pd.set_option('display.max_columns', 12)
# read in the raw csv
adeck_df = pd.read_csv(in_file)
#print(adeck_df)
#extract only the carq items and remove duplicates
carq_data = adeck_df[(adeck_df.MODEL == 'CARQ') & (adeck_df.TAU == 0)].drop_duplicates(keep='last')
#print(carq_data)
#remove carq items from original
final_df = adeck_df[adeck_df.MODEL != 'CARQ']
#print(final_df)
row_list = []
for index, row in carq_data.iterrows():
position_time = row['POSDATETIME']
for index, arow in final_df.iterrows():
if arow['POSDATETIME'] == position_time:
# match, so do calculations
storm_id = arow['STORMID']
model_base_time = arow['MODELDATETIME']
the_hour = arow['TAU']
the_model = arow['MODEL']
point1 = float(row['LAT']), float(row['LON'])
point2 = float(arow['LAT']), float(arow['LON'])
if arow['LAT'] == 0.0:
dist_error = None
else:
dist_error = int(round(haversine(point1, point2, miles=True)))
if arow['WIND'] != 0:
wind_error = int(abs(int(row['WIND']) - int(arow['WIND'])))
else: wind_error = None
if arow['PRES'] != 0:
pressure_error = int(abs(int(row['PRES']) - int(arow['PRES'])))
else:
pressure_error = None
lat_carq = row['LAT']
lon_carq = row['LON']
lat_model = arow['LAT']
lon_model = arow['LON']
wind_carq = row['WIND']
wind_model = arow['WIND']
pres_carq = row['PRES']
pres_model = arow['PRES']
row_list.append([storm_id, model_base_time, the_model, the_hour, lat_carq, lon_carq, lat_model, lon_model, dist_error,
wind_carq, wind_model, wind_error, pres_carq, pres_model, pressure_error])
result_df = pd.DataFrame(row_list)
result_df = result_df.where((pd.notnull(result_df)), None)
result_cols = ['StormID', 'ModelBasetime', 'Model' , 'Tau',
'LatCARQ', 'LonCARQ', 'LatModel', 'LonModel', 'DistError',
'WindCARQ', 'WindModel','WindError',
'PresCARQ', 'PresModel','PresError']
result_df.columns = result_cols
calculate_adeck_errors(infile)
To clarify what I'm doing:
1. The CARQ entries are the control (actual).
2. The other models are the guesses.
3. I'm comparing the control (CARQ) to the guesses to see what their errors are.
4. The basis of the comparison is the MODELBASETIME = POSBASETIME
4. A sample file I'm processing is here: http://vortexweather.com/downloads/adeck/aal062018.csv
I was hoping there is a faster way than i'm doing it, or another pandas method besides iterrows
Many thanks for suggestion.
Bryan
This code takes about 10 seconds to run your entire dataset!
The code looks very similar to what you have written, with the exception that all of the operations within the main_function have been vectorized. See Fast, Flexible, Easy and Intuitive: How to Speed Up Your Pandas Projects
2018-09-13_adeck_error_calculations.ipynb
import pandas as pd
import numpy as np
import datetime
from haversine import haversine
def main_function(df, row):
"""
The main difference here is that everything is vectorized
Returns: DataFrame
"""
df_new = pd.DataFrame()
df_storage = pd.DataFrame()
pos_datetime = df.POSDATETIME.isin([row['POSDATETIME']]) # creates a Boolean map
array_len = len(pos_datetime)
new_index = pos_datetime.index
df_new['StormID'] = df.loc[pos_datetime, 'STORMID']
df_new['ModelBaseTime'] = df.loc[pos_datetime, 'MODELDATETIME']
df_new['Model'] = df.loc[pos_datetime, 'MODEL']
df_new['Tau'] = df.loc[pos_datetime, 'TAU']
# Distance
df_new['LatCARQ'] = pd.DataFrame(np.full((array_len, 1), row['LAT']), index=new_index).loc[pos_datetime, 0]
df_new['LonCARQ'] = pd.DataFrame(np.full((array_len, 1), row['LON']), index=new_index).loc[pos_datetime, 0]
df_new['LatModel'] = df.loc[pos_datetime, 'LAT']
df_new['LonModel'] = df.loc[pos_datetime, 'LON']
def calc_dist_error(row):
return round(haversine((row['LatCARQ'], row['LonCARQ']), (row['LatModel'], row['LonModel']), miles=True)) if row['LatModel'] != 0.0 else None
df_new['DistError'] = df_new.apply(calc_dist_error, axis=1)
# Wind
df_new['WindCARQ'] = pd.DataFrame(np.full((array_len, 1), row['WIND']), index=new_index).loc[pos_datetime, 0]
df_new['WindModel'] = df.loc[pos_datetime, 'WIND']
df_storage['row_WIND'] = pd.DataFrame(np.full((array_len, 1), row['WIND']), index=new_index).loc[pos_datetime, 0]
df_storage['df_WIND'] = df.loc[pos_datetime, 'WIND']
def wind_error_calc(row):
return (row['row_WIND'] - row['df_WIND']) if row['df_WIND'] != 0 else None
df_new['WindError'] = df_storage.apply(wind_error_calc, axis=1)
# Air Pressure
df_new['PresCARQ'] = pd.DataFrame(np.full((array_len, 1), row['PRES']), index=new_index).loc[pos_datetime, 0]
df_new['PresModel'] = df.loc[pos_datetime, 'PRES']
df_storage['row_PRES'] = pd.DataFrame(np.full((array_len, 1), row['PRES']), index=new_index).loc[pos_datetime, 0]
df_storage['df_PRES'] = df.loc[pos_datetime, 'PRES']
def pres_error_calc(row):
return abs(row['row_PRES'] - row['df_PRES']) if row['df_PRES'] != 0 else None
df_new['PresError'] = df_storage.apply(pres_error_calc, axis=1)
del(df_storage)
return df_new
def calculate_adeck_errors(in_file):
"""
Retruns: DataFrame
"""
print(f'Starting Data Calculations: {datetime.datetime.now().strftime("%I:%M:%S%p on %B %d, %Y")}')
pd.set_option('max_columns', 20)
pd.set_option('max_rows', 300)
# read in the raw csv
adeck_df = pd.read_csv(in_file)
adeck_df['MODELDATETIME'] = pd.to_datetime(adeck_df['MODELDATETIME'], format='%Y-%m-%d %H:%M')
adeck_df['POSDATETIME'] = pd.to_datetime(adeck_df['POSDATETIME'], format='%Y-%m-%d %H:%M')
#extract only the carq items and remove duplicates
carq_data = adeck_df[(adeck_df.MODEL == 'CARQ') & (adeck_df.TAU == 0)].drop_duplicates(keep='last')
print('Len carq_data: ', len(carq_data))
#remove carq items from original
final_df = adeck_df[adeck_df.MODEL != 'CARQ']
print('Len final_df: ', len(final_df))
df_out_new = pd.DataFrame()
for index, row in carq_data.iterrows():
test_df = main_function(final_df, row) # function call
df_out_new = df_out_new.append(test_df, sort=False)
df_out_new = df_out_new.reset_index(drop=True)
df_out_new = df_out_new.where((pd.notnull(df_out_new)), None)
print(f'Finishing Data Calculations: {datetime.datetime.now().strftime("%I:%M:%S%p on %B %d, %Y")}')
return df_out_new
in_file = 'aal062018.csv'
df = calculate_adeck_errors(in_file)
>>>Starting Data Calculations: 02:18:30AM on September 13, 2018
>>>Len carq_data: 56
>>>Len final_df: 137999
>>>Finishing Data Calculations: 02:18:39AM on September 13, 2018
print(len(df))
>>>95630
print(df.head(20))
Please don't forget to check the accepted solution. Enjoy!
Looks like you are creating two dataframes out of the same dataframe, and then processing them. Two things that may cut your time.
First, you are iterating over both dataframes and checking for a condition:
for _, row in carq_data.iterrows():
for _, arow in final_df.iterrows():
if arow['POSDATETIME'] == row['POSDATETIME']:
# do something by using both tables
This is essentially an implementation of a join. You are joining carq_data with final_df on 'POSDATETIME'.
As a first step, you should merge the tables:
merged = carq_data.merge(final_df, on=['POSDATETIME'])
At this point you will get multiple rows for each similar 'POSDATETIME'. In the below, let's assume column b is POSDATETIME:
>>> a
a b
0 1 11
1 1 33
>>> b
a b
0 1 2
1 1 3
2 1 4
>>> merged = a.merge(b, on=['a'])
>>> merged
a b_x b_y
0 1 11 2
1 1 11 3
2 1 11 4
3 1 33 2
4 1 33 3
5 1 33 4
Now, to do your conditional calculations, you can use the apply() function.
First, define a function:
def calc_dist_error(row):
return int(round(haversine(row['b_x'], row['b_y'], miles=True))) if row['a'] != 0.0 else None
Then apply it to every row:
merged['dist_error'] = merged.apply(calc_dist_error, axis=1)
Continuing my small example:
>>> merged['c'] = [1, 0, 0, 0, 2, 3]
>>> merged
a b_x b_y c
0 1 11 2 1
1 1 11 3 0
2 1 11 4 0
3 1 33 2 0
4 1 33 3 2
5 1 33 4 3
>>> def foo(row):
... return row['b_x'] - row['b_y'] if row['c'] != 0 else None
...
>>> merged['dist_error'] = merged.apply(foo, axis=1)
>>> merged
a b_x b_y c dist_error
0 1 11 2 1 9.0
1 1 11 3 0 NaN
2 1 11 4 0 NaN
3 1 33 2 0 NaN
4 1 33 3 2 30.0
5 1 33 4 3 29.0
This should help you reduce run time (see also this for how to check using %timeit). Hope this helps!
I am trying to create a flag for date from datetime column. but getting an error after applying the below function.
def f(r):
if r['balance_dt'] <= '2016-11-30':
return 0
else:
return 1
df_obctohdfc['balance_dt_flag'] = df_obctohdfc.apply(f,axis=1)
The error your are getting is because you are comparing string object to datetime object. You can convert the string to datetime.
Ex:
import datetime
def f(r):
if r['balance_dt'] <= datetime.datetime.strptime('2016-11-30', '%Y-%m-%d'):
return 0
else:
return 1
df_obctohdfc['balance_dt_flag'] = df_obctohdfc.apply(f,axis=1)
Note: It is better to do the way jezrael has mention. That is the right way to do it
In pandas is best avoid loops, how working apply under the hood.
I think need convert string to datetime and then cast mask to integer - True to 1 and False to 0 and change <= to >:
timestamp = pd.to_datetime('2016-11-30')
df_obctohdfc['balance_dt_flag'] = (df_obctohdfc['balance_dt'] > timestamp).astype(int)
Sample:
rng = pd.date_range('2016-11-27', periods=10)
df_obctohdfc = pd.DataFrame({'balance_dt': rng})
#print (df_obctohdfc)
timestamp = pd.to_datetime('2016-11-30')
df_obctohdfc['balance_dt_flag'] = (df_obctohdfc['balance_dt'] > timestamp).astype(int)
print (df_obctohdfc)
balance_dt balance_dt_flag
0 2016-11-27 0
1 2016-11-28 0
2 2016-11-29 0
3 2016-11-30 0
4 2016-12-01 1
5 2016-12-02 1
6 2016-12-03 1
7 2016-12-04 1
8 2016-12-05 1
9 2016-12-06 1
Comparing in 1000 rows DataFrame:
In [140]: %timeit df_obctohdfc['balance_dt_flag1'] = (df_obctohdfc['balance_dt'] > timestamp).astype(int)
1000 loops, best of 3: 368 µs per loop
In [141]: %timeit df_obctohdfc['balance_dt_flag2'] = df_obctohdfc.apply(f,axis=1)
10 loops, best of 3: 91.2 ms per loop
Setup:
rng = pd.date_range('2015-11-01', periods=1000)
df_obctohdfc = pd.DataFrame({'balance_dt': rng})
#print (df_obctohdfc)
timestamp = pd.to_datetime('2016-11-30')
import datetime
def f(r):
if r['balance_dt'] <= datetime.datetime.strptime('2016-11-30', '%Y-%m-%d'):
return 0
else:
return 1
I have a numpy array containing a sequence of 1 and -1 of shape (1000,). I need to add some random noise to 10% of this sample. The noise will be a simple switch of the the sign of a given value (e.g. -1 if originally 1). My current solution is as follows:
# create an array of the indices of the values we are going to change
noisy_index = np.random.choice(1000, size=100, replace=False)
# multiply by -1 to change sign of these randomly selected values
mult = np.multiply(correct_labels[noisy_index], -1)
# delete these values from the original dataset
labels_subset = np.delete(correct_labels, noisy_index)
# concatenate the modified values to get back complete (and now noisy) dataset
correct_labels_noisy = np.concatenate((labels_subset, mult), axis=0)
Two questions: 1) is this doing what I described it should be doing? 2) is there a more straightforward and less verbose approach?
The easiest would be multiply the selected values by -1 and assign back to the corresponding indices with:
correct_labels[noisy_index] *= -1
Example:
correct_labels = np.random.choice([1,-1], size=10)
print('original labels: ', correct_labels)
noisy_index = np.random.choice(10, size=3, replace=False)
correct_labels[noisy_index] *= -1
print('modified labels: ', correct_labels)
#original labels: [ 1 -1 -1 1 -1 -1 -1 -1 -1 1]
#modified labels: [ 1 -1 -1 1 1 -1 1 -1 -1 -1]
# ^ ^ ^
print('noisy index: ', noisy_index)
# noisy index: [6 4 9]
I have created a 5 day moving average for 5 years worth of data. How do I iterate through this to show if the moving average is rising or falling for every single day. My code is simply giving me 1 integer answer rather than a rising (+1) or falling(-1) answer for every day. Thank you!
import pandas as pd
df = pd.read_csv('file:///C:/Users/James Brodie/Desktop/USDJPY.csv', header=1, index_col=0)
ma5 = df['PX_LAST'].rolling(window=5).mean()
ma8 = df['PX_LAST'].rolling(window=8).mean()
ma21 = df['PX_LAST'].rolling(window=21).mean()
ma5x = []
for i in ma5:
if i > i-1:
ma5x = 1
elif i < i-1:
ma5x = -1
else:
ma5x = 0
print(ma5x)
Thank you!
ma5 = [5,2,2,3,3,2,5]
ma5x = []
lastItem = ma5[0]
for currItem in ma5[1:]:
if currItem > lastItem:
ma5x.append(1)
elif currItem < lastItem:
ma5x.append(-1)
else:
ma5x.append(0)
lastItem = currItem
print(ma5x)
gives:
[-1, 0, 1, 0, -1, 1]
The elements of a list are best represented in pandas as a Series object (put own index of choice for rows replacing list(range(len(ma5x))) with anything you need):
print('-----------')
import pandas as pd
pd_ma5x = pd.Series(ma5x, index=list(range(len(ma5x))))
print(pd_ma5x)
gives:
-----------
0 -1
1 0
2 1
3 0
4 -1
5 1
dtype: int64
Using own index be aware that pd_ma5x size is one less than that of ma5