I am working on a project that reads a couple of thousand text documents, creates a dataframe from them, and then trains a model on the dataframe. The most time-consuming aspect of the code is the creation of the dataframe.
Here is how I create the dataframe:
I first create 4-5 lists, create a dictionary with 'Column-name' as the key and the previous lists as the values. Then use pd.DataFrame to give the dictionary. I have added print updates after each step and the dataframe creation step takes the most time.
Method I am using:
line_of_interest = []
line_no = []
file_name = []
for file in file_names:
with open(file) as txt:
for i, line in enumerate(txt):
if 'word of interest' in line:
line_of_interest.append(line)
line_no.append(i)
file_name.append()
rows = {'Line_no':line_no,'Line':line_of_interest,'File':file_name}
df = pd.DataFrame(data = rows)
I was wondering if there is a more efficient and less time-consuming way to create the dataframe. I tried looking for similar questions and the only thing I could find was "Most Efficient Way to Create Pandas DataFrame from Web Scraped Data".
Let me know if there is a similar question with a good answer. The only other method of creating a dataframe I know is appending row by row all the values as I discover them, and I don't know a way to check if that is quicker. Thanks!
I have a big table in Hive (dozens to hundreds of millions of rows) from which I want to only choose those that match a regex.
Currently I have a small example to try my code first:
columns = ['id', 'column']
vals = [
(1, "VAL_ID1 BD store"),
(2, "VAL_ID2 BD store"),
(3, "VAL_ID3 BD model"),
(4, "BAD WRONG")
]
df = spark.createDataFrame(vals, columns)
And then I have a regex tested that goes like:
df_regex = df.withColumn('newColumn',F.regexp_extract(df['id'], '^(([a-zA-Z]{2}[a-zA-Z0-9]{1})+(_[a-zA-Z]{2}[a-zA-Z0-9]{1})*)(\s|$)',1))
As I said, this is a test dataframe. In the future I will make it "look" at a very large table. Is there any way to only add rows that match the regex, and thus create a much smaller dataframe?
As it is right now, I am reading every single row, then adding a column withColumn that has an empty field for the rows that do not match the regex. Which makes sense, but I feel like there is benefit in not reading this dataframe two times if I can avoid it.
You want to use the where probably.
df.where(
F.regexp_extract(df['id'], '^(([a-zA-Z]{2}[a-zA-Z0-9]{1})+(_[a-zA-Z]{2}[a-zA-Z0-9]{1})*)(\s|$)',1) != F.lit('')
)
Actually, I tried your regex and it gives no results. But as long as you understand the principle, I think you can use that solution.
EDIT:
I feel like there is benefit in not reading this dataframe two times if I can avoid it.
Spark will read your data only if you perform "action". Transformations are lazy and therefore evaluated only at the end ... so no need to worry about Spark reading your data twice (or more).
I have an empty data frame with about 120 columns, I want to fill it using data I have in a file.
I'm iterating over a file that has about 1.8 million lines.
(The lines are unstructured, I can't load them to a dataframe directly)
For each line in the file I do the following:
Extract the data I need from the current line
Copy the last row in the data frame and append it to the end df = df.append(df.iloc[-1]). The copy is critical, most of the data in the previous row won't be changed.
Change several values in the last row according to the data I've extracted df.iloc[-1, df.columns.get_loc('column_name')] = some_extracted_value
This is very slow, I assume the fault is in the append.
What is the correct approach to speed things up ? preallocate the dataframe ?
EDIT:
After reading the answers I did the following:
I preallocated the dataframe (saved like 10% of the time)
I replaced this : df = df.append(df.iloc[-1]) with this : df.iloc[i] = df.iloc[i-1] (i is the current iteration in the loop).(save like 10% of the time).
Did profiling, even though I removed the append the main issue is copying the previous line, meaning : df.iloc[i] = df.iloc[i-1] takes about 95% of the time.
You may need plenty of memory, whichever option you choose.
However, what you should certainly avoid is using pd.DataFrame.append within a loop. This is expensive versus list.append.
Instead, aggregate to a list of lists, then feed into a dataframe. Since you haven't provided an example, here's some pseudo-code:
# initialize empty list
L = []
for line in my_binary_file:
# extract components required from each line to a list of Python types
line_vars = [line['var1'], line['var2'], line['var3']]
# append to list of results
L.append(line_vars)
# create dataframe from list of lists
df = pd.DataFrame(L, columns=['var1', 'var2', 'var3'])
The Fastest way would be load to dataframe directly via pd.read_csv()
Try separating the logic to clean out unstructured to structured data and then use pd.read_csv to load the dataframe.
You can share the sample unstructured line and logic to take out the structured data, So that might share some insights on the same.
Where you use append you end up copying the dataframe which is inefficient. Try this whole thing again but avoiding this line:
df = df.append(df.iloc[-1])
You could do something like this to copy the last row to a new row (only do this if the last row contains information that you want in the new row):
df.iloc[...calculate the next available index...] = df.iloc[-1]
Then edit the last row accordingly as you have done
df.iloc[-1, df.columns.get_loc('column_name')] = some_extracted_value
You could try some multiprocessing to speed things up
from multiprocessing.dummy import Pool as ThreadPool
def YourCleaningFunction(line):
for each line do the following
blablabla
return(your formated lines with ,) # or use the kind of function jpp just provided
pool = ThreadPool(8) # your number of cores
lines = open('your_big_csv.csv').read().split('\n') # your csv as a list of lines
df = pool.map(YourCleaningFunction, lines)
df = pandas.DataFrame(df)
pool.close()
pool.join()
I have a large pandas dataframe, where I am running groups by operations.
CHROM POS Data01 Data02 ......
1 ....................
1 ...................
2 ..................
2 ............
scaf_9 .............
scaf_9 ............
So, i am doing:
my_data_grouped = my_data.groupby('CHROM')
for chr_, data in my_data_grouped:
do something in chr_
write something from that chr_ data
Everything is fine in small data and in the data where there is no string type CHROM i.e scaff_9. But, with very large data and with scaff_9, I am getting two groups of 2. It really isn't an error message and it is not affecting the computation. The issue is when I write the data by group in the file; I am getting two groups of 2 (splitted unequally).
It is becoming very hard for me to traceback the origin of this problem, since there is no error message and with small data it works well. My only assumption are:
Is there certain limit on the the number of lines in total dataframe vs. grouped dataframe the pandas module can handle. What is the fix to this problem ?
Among all the 2 most of them are treated as integer object and some (later part) as string object being close to scaff_9. Is this possible ?
Sorry, I am only making my assumption here, and it is becoming impossible for me to know the origin of the problem.
Post Edit:
I have also tried to run sort_by(['CHROM']) before doing to groupby, but the problem still persists.
Any possible fix to the issue.
Thanks,
In my opinion there is data problem, obviously some whitespaces, so pandas processes each group separately.
Solution should be remove traling whitespaces first:
df.index = df.index.astype(str).str.strip()
You can also check unique strings values of index:
a = df.index[df.index.map(type) == str].unique().tolist()
If first column is not index:
df['CHROM'] = df['CHROM'].astype(str).str.strip()
a = df.loc[df['CHROM'].map(type) == str, 'CHROM'].unique().tolist()
EDIT:
Last final solution was simplier - casting to str like:
df['CHROM'] = df['CHROM'].astype(str)
I'm having trouble with processing some csv data files for a project. Someone suggested using python/csv reader to help break down the files, which I've had some success with, but not in a way I can use.
This code is a little different from what I was trying before. I am essentially attempting to create an array. In the raw data format, the first 7 rows contain no data, and then each column contains 50 experiments, each with 4000 rows, for 200000 some rows total. What I want to do is take each column, and make it an individual csv file, with each experiment in its own column. So it would be an array of 50 columns and 4000 rows for each data type. The code here does break down the correct values, I think the logic is okay, but it is breaking down the opposite of how I want it. I want the separators without quotes (the commas and spaces) and I want the element values in quotes. Right now it is doing just the opposite for both, element values with no quotes, and the separators in quotes. I've spent several hours trying to figure out how to do this to no avail,
import csv
ifile = open('00_follow_maverick.csv')
epistemicfile = open('00_follower_maverick_EP.csv', 'w')
reader = csv.reader(ifile)
colnum = 0
rownum = 0
y = 0
z = 8
for column in reader:
rownum = 4000 * y + z
for element in column:
writer = csv.writer(epistemicfile)
if y <= 50:
y = y + 1
writer.writerow([element])
writer.writerow(',')
rownum = x * y + z
if y > 50:
y = 0
z = z + 1
writer.writerow(' ')
rownum = x * y + z
if z >= 4008:
break
What is going on: I am taking each row in the raw data file in iterations of 4000, so that I can separate them with commas for the 50 experiments. When y, the experiment indicator here, reaches 50, it resets back to experiment 0, and adds 1 to z, which tells it which row to look at, by the formula of 4000 * y + z. When it completes the rows for all 50 experiments, it is finished. The problem here is that I don't know how to get python to write the actual values in quotes, and my separators outside of quotes.
Any help will be most appreciated. Apologies if this seems a stupid question, I have no programming experience, this is my first attempt ever. Thank you.
Sorry, I'll try to make this more clear. The original csv file has several columns, each of which are different sets of data.
A miniature example of the raw file looks like:
column1 column2 column3
exp1data1time1 exp1data2time1 exp1data3time1
exp1data1time2 exp1data2time2 exp1data3time2
exp2data1time1 exp2data2time1 exp2data3time1
exp2data1time2 exp2data2time2 exp2data3time2
exp3data1time1 exp3data2time1 exp3data3time1
exp3data1time2 exp3data2time2 exp3data3time2
So, the actual version has 4000 rows instead of 2 for each new experiment. There are 40 columns in the actual version, but basically, the data type in the raw file matches the column number. I want to separate each data type or column into an individual csv file.
This would look like:
csv file1
exp1data1time1 exp2data1time1 exp3data1time1
exp1data1time2 exp2data1time2 exp3data1time2
csv file2
exp1data2time1 exp2data2time1 exp3data2time1
exp1data2time2 exp2data2time2 exp3data2time2
csv file3
exp1data3time1 exp2data3time1 exp3data3time1
exp1data3time2 exp2data3time2 exp3data3time2
So, I'd move the raw data in the file to a new column, and each data type to its own file. Right now I'm only going to do one file, until I can move the separate experiments to separate columns in the new file. So, in the code, the above would make the 4000 into 2. I hope this makes more sense, but if not, I will try again.
If I had a cat for each time I saw a bio or psych or chem database in this state:
"each column contains 50 experiments,
each with 4000 rows, for 200000 some
rows total. What I want to do is take
each column, and make it an individual
csv file, with each experiment in its
own column. So it would be an array of
50 columns and 4000 rows for each data
type"
I'd have way too farking many cats.
I didn't even look at your code because the re-mangling you are proposing is just another problem that will have to be solved. I don't fault you, you claim to be a novice and all your peers make the same sort of error. Beginning programmers who have yet to understand how to use arrays often wind up with variable declarations like:
integer response01, response02, response03, response04, ...
and then very, very redundant code when they try to see if every response is - say - 1. I think this is such a seductive error in bio-informatics because it actually models the paper notations they come from rather well. Unfortunately, the sheet-of-paper model isn't the best way to model data.
You should read and understand why database normalization was developed, codified and has come to dominate how people think about structured data. One Wikipedia article may not be sufficient. Using the example I excerpted let me try to explain how I think of it. Your data consists of observations; put the other way the primary datum is a singular observation. That observation has a context though: it is one of a set of 4000 observations, where each set belongs to one of 50 experiments. If you had to attach a context to each observation you'd wind up with an addressing scheme that looks like:
<experiment_number, observation_number, value>
In database jargon, that's a tuple, and it is capable of representing, with no ambiguity and perfect symmetry the entirety of your data. I'm not certain that I've understood the exact structure of your data, so perhaps it is something more like:
<experiment_number, protocol_number, observation_number, value>
where the protocol may be some form of variable treatment type - let's say pH. But note that I didn't call the protocol a pH and I don't record it as such in the database. What I would then need is an ancillary table showing the relevant parameters of the protocol, e.g.:
<protocol_number, acidity, temperature, pressure>
Now we've just built a "relation" that those database people like to talk about; we've also begun normalizing the data. If you need to know the pH for a given protocol, there is one and only one place to find it, in the proper row of the protocol table. Note that I've divorced the data that fit so nicely together on a data-sheet and from the observation table I can't see the pH for a particular dataum. But that's okay, because I can just look it up in my protocol table if needed. This is a "relational join" and if I needed to, I could coalesce all the various parameters from all the various tables and reconstitute the original datasheet in its original, unstructured glory.
I hope this answer is of some use to you. I'm certain that I don't even know what field of study your data is from, but these principles apply across domains from drug trials to purchase requisition processing. Please understand that I'm trying to inform, per your request, and there is zero condescension intended. I welcome further questions on the matter.
Normalization of the dataset
Thanks for giving the example. You have the context I described already, perhaps I can make it more clear.
column1 column2 column3
exp1data1time1 exp1data2time1 exp1data3time1
exp1data1time2 exp1data2time2 exp1data3time2
The columns are an artifice made by the last guy; that is, they carry no relevant information. When parsed into a normal form, your data looks just like my first proposed tuple:
<experiment_number, time, response_number, response>
where I suspect time may actually mean "subject_id" or "trial_number". It may very well look incongruous to you to conjoin all the different response values into the same dataset; indeed based on your desired output, I suspect that it does. At first blush, the objection "but the subject's response to a question about epistemic properties of chairs has no connection to their meta-epistemic beliefs regarding color", but this would be mistaken. The data are related because they have a common experimental subject, and self-correlation is an important concept in sociological analytics.
For example, you may find that respondent A gives the same responses as respondent B, except all of A's responses are biased one higher because of how the subject understood the criteria. This would make a very real difference in the absolute values of the data, but I hope you can see that the question "do A and B actually have different epistemic models?" is salient and valid. One method of data modeling allows this question to be answered easily, your desired method does not.
Working parsing code to follow shortly.
The normalizing code
#!/usr/bin/python
"""parses a csv file containing a particular data layout and normalizes
The raw data set is a csv file of the form::
column1 column2 column3
exp01data01time01 exp01data02time01 exp01data03time01
exp01data01time02 exp01data02time02 exp01data03time02
where there are 40 such columns and the literal column title
is added as context to the output row
it is assumed that the columns are comma separated but
the lexical form of the subcolumns is unspecified.
Output will consist of a single CSV output stream
on stdout of the form::
exp01, time01, data01, column1
for varying actual values of each field.
"""
import csv
import sys
def split_subfields(s):
"""returns a list of subfields of s
this function is expected to be re-written to match the actual,
unspecified lexical structure of s."""
return [s[0:5], s[5:11], s[11:17]]
def normalise_data(reader, writer):
"""returns a list of the column headings from the reader"""
# obtain the headings for use in normalization
names = reader.next()
# get the data rows, split them out by column, add the column name
for row in reader:
for column, datum in enumerate(row):
fields = split_subfields(datum)
fields.append(names[column])
writer.writerow(fields)
def main():
if len(sys.argv) != 2:
print >> sys.stderr, ('usage: %s input.csv' % sys.argv[0])
sys.exit(1)
in_file = sys.argv[1]
reader = csv.reader(open(in_file))
writer = csv.writer(sys.stdout)
normalise_data(reader, writer)
if __name__ == '__main__': main()
Such that the command python epistem.py raw_data.csv > cooked_data.csv yields excerpted output looking like:
exp01,data01,time01,column1
...
exp01,data40,time01,column40
exp01,data01,time02,column1
exp01,data01,time03,column1
...
exp02,data40,time15,column40