lzhw 1.1.15

Compression suite for data frames and tabular data files, csv, excel etc.

lzhw

Compression library for data frames and tabular data files, csv, excel etc.

Compression library to compress big lists and/or pandas dataframes using an optimized algorithm (lzhw) developed from Lempel-Ziv, Huffman and LZ-Welch techniques.

lzhw has a command line tool that can be downloaded from here and can work from command line with no prior python installation.

Manual on how to use it available here.

It works on Windows and soon a Mac version will be available.

Full documentation can be found here

Data Frames compression and decompression works in parallel.

Quick Start

pip install lzhw

import lzhw

sample_data = ["Sunny", "Sunny", "Overcast", "Rain", "Rain", "Rain", "Overcast", 
               "Sunny", "Sunny", "Rain", "Sunny", "Overcast", "Overcast", "Rain", 
               "Rain", "Rain", "Sunny", "Sunny", "Overcaste"]

compressed = lzhw.LZHW(sample_data)
## let's see how the compressed object looks like:
print(compressed.compressed)
# (506460, 128794, 112504)

## its size
print(compressed.size())
# 72

## size of original
from sys import getsizeof
print(getsizeof(sample_data))
# 216

print(compressed.space_saving())
# space saving from original to compressed is 67%

## Let's decompress and check whether there is any information loss
decomp = compressed.decompress()
print(decomp == sample_data)
# True

As we saw, the LZHW class has saved 67% of the space used to store the original list without any loss. This percentage can get better with bigger data that may have repeated sequences. The class has also some useful helper methods as space_saving, size, and decompress() to revert back to original.

Another example with numeric data.

from random import sample, choices

numbers = choices(sample(range(0, 5), 5), k = 20)
comp_num = lzhw.LZHW(numbers)

print(getsizeof(numbers) > comp_num.size())
# True

print(numbers == list(map(int, comp_num.decompress()))) ## make it int again
# True

print(comp_num.space_saving())
# space saving from original to compressed is 73%

Let's look at how the compressed object is stored and how it looks like when printed: LZHW class has an attribute called compressed which is a tuple of integers representing the encoded triplets.

print(comp_num.compressed) # how the compressed is saved (as tuple of 3 integers)
# (8198555, 620206, 3059308)

We can also write the compressed data to files using save_to_file method, and read it back and decompress it using decompress_from_file function.

status = ["Good", "Bad", "Bad", "Bad", "Good", "Good", "Average", "Average", "Good",
          "Average", "Average", "Bad", "Average", "Good", "Bad", "Bad", "Good"]
comp_status = lzhw.LZHW(status)
comp_status.save_to_file("status.txt")
decomp_status = lzhw.decompress_from_file("status.txt")
print(status == decomp_status)
# True

Compressing DataFrames in Parallel

lzhw doesn't work only on lists, it also compress pandas dataframes and save it into compressed files to decompress them later.

import pandas as pd

df = pd.DataFrame({"a": [1, 1, 2, 2, 1, 3, 4, 4],
                   "b": ["A", "A", "B", "B", "A", "C", "D", "D"]})
comp_df = lzhw.CompressedDF(df)
# 100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2003.97it/s]

Let's check space saved by compression

comp_space = 0
for i in range(len(comp_df.compressed)):
	comp_space += comp_df.compressed[i].size()

print(comp_space, getsizeof(df))
# 144 712

## Test information loss
print(comp_df.compressed[0].decompress() == list(map(str, df.a)))
# True

Saving and Loading Compressed DataFrames

With lzhw we can save a data frame into a compressed file and then read it again using save_to_file method and decompress_df_from_file function.

## Save to file
comp_df.save_to_file("comp_df.txt")

## Load the file
original = lzhw.decompress_df_from_file("comp_df.txt")
# 100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2004.93it/s]

print(original)
#   a  b
#0  1  A
#1  1  A
#2  2  B
#3  2  B
#4  1  A
#5  3  C
#6  4  D
#7  4  D

Compressing Bigger DataFrames

Let's try to compress a real-world dataframe german_credit.xlsx file from UCI Machine Learning Repository [1].

Original txt file is 219 KB on desk.

gc_original = pd.read_excel("examples/german_credit.xlsx")
comp_gc = lzhw.CompressedDF(gc_original)
# 100%|█████████████████████████████████████████████████████████████████████████████████| 62/62 [00:00<00:00, 257.95it/s]

## Compare sizes in Python:
comp_space = 0
for i in range(len(comp_gc.compressed)):
	comp_space += comp_gc.compressed[i].size()

print(comp_space, getsizeof(gc_original))
# 4504 548852

print(comp_gc.compressed[0].decompress() == list(map(str, gc_original.iloc[:, 0])))
# True

Huge space saving, 99%, with no information loss!

Let's now write the compressed dataframe into a file and compare the sizes of the files.

comp_gc.save_to_file("gc_compressed.txt")

Checking the size of the compressed file, it is 44 KB. Meaning that in total we saved around 79%. Future versions will be optimized to save more space.

Let's now check when we reload the file, will we lose any information or not.

## Load the file
gc_original2 = lzhw.decompress_df_from_file("gc_compressed.txt")
# 100%|█████████████████████████████████████████████████████████████████████████████████| 62/62 [00:00<00:00, 259.46it/s]

print(list(gc_original2.iloc[:, 13]) == list(map(str, gc_original.iloc[:, 13])))
# True

print(gc_original.shape == gc_original2.shape)
# True

Perfect! There is no information loss at all.

With lzhw also you can choose what columns you are interested in compressing from a data frame. CompressedDF class has an argument selected_cols. And how many rows you want to decompress with n_rows argument.

Please see documentation for deeper look

LZHW Comparison with joblib algorithms

I love joblib. I usually use it for parallelism for its great performance coming with a smooth simplicity.

I once saw this article in its documentation and it is about measuring the performance between different compressors available in it.

Because I am developing a compression library, I wanted to extend the code available in this article adding lzhw to the comparison, just to know where my library stands.

joblib uses three main techniques in this article Zlib, LZMA and LZ4.

I will use 1500000 Sales Records Data.

We will look at Compression and Decompression Duration and The compressed file sizes.

The downloaded compressed file is 53MB on the websites

I will reproduce the code in joblib documentation

data = pd.read_csv("1500000 Sales Records.csv")
print(data.shape)

pickle_file = './pickle_data.joblib'
start = time.time()
with open(pickle_file, 'wb') as f:
    dump(data, f)
raw_dump_duration = time.time() - start
print("Raw dump duration: %0.3fs" % raw_dump_duration)

raw_file_size = os.stat(pickle_file).st_size / 1e6
print("Raw dump file size: %0.3fMB" % raw_file_size)

start = time.time()
with open(pickle_file, 'rb') as f:
    load(f)
raw_load_duration = time.time() - start
print("Raw load duration: %0.3fs" % raw_load_duration)

## ZLIB
start = time.time()
with open(pickle_file, 'wb') as f:
    dump(data, f, compress='zlib')
zlib_dump_duration = time.time() - start
print("Zlib dump duration: %0.3fs" % zlib_dump_duration)

zlib_file_size = os.stat(pickle_file).st_size / 1e6
print("Zlib file size: %0.3fMB" % zlib_file_size)

start = time.time()
with open(pickle_file, 'rb') as f:
    load(f)
zlib_load_duration = time.time() - start
print("Zlib load duration: %0.3fs" % zlib_load_duration)

## LZMA
start = time.time()
with open(pickle_file, 'wb') as f:
    dump(data, f, compress=('lzma', 3))
lzma_dump_duration = time.time() - start
print("LZMA dump duration: %0.3fs" % lzma_dump_duration)

lzma_file_size = os.stat(pickle_file).st_size / 1e6
print("LZMA file size: %0.3fMB" % lzma_file_size)

start = time.time()
with open(pickle_file, 'rb') as f:
    load(f)
lzma_load_duration = time.time() - start
print("LZMA load duration: %0.3fs" % lzma_load_duration)

## LZ4
start = time.time()
with open(pickle_file, 'wb') as f:
    dump(data, f, compress='lz4')
lz4_dump_duration = time.time() - start
print("LZ4 dump duration: %0.3fs" % lz4_dump_duration)

lz4_file_size = os.stat(pickle_file).st_size / 1e6
print("LZ4 file size: %0.3fMB" % lz4_file_size)

start = time.time()
with open(pickle_file, 'rb') as f:
    load(f)
lz4_load_duration = time.time() - start
print("LZ4 load duration: %0.3fs" % lz4_load_duration)

## LZHW
start = time.time()
lzhw_data = lzhw.CompressedDF(data)
lzhw_data.save_to_file("lzhw_data.txt")
lzhw_compression_duration = time.time() - start
print("LZHW compression duration: %0.3fs" % lzhw_compression_duration)

lzhw_file_size = os.stat("lzhw_data.txt").st_size / 1e6
print("LZHW file size: %0.3fMB" % lzhw_file_size)

start = time.time()
lzhw_d = lzhw.decompress_df_from_file("lzhw_data.txt")
lzhw_d_duration = time.time() - start
print("LZHW decompression duration: %0.3fs" % lzhw_d_duration)

# (1500000, 14)
# Raw dump duration: 1.294s
# Raw dump file size: 267.591MB
# Raw load duration: 1.413s
# Zlib dump duration: 6.583s
# Zlib file size: 96.229MB
# Zlib load duration: 2.430s
# LZMA dump duration: 76.526s
# LZMA file size: 72.476MB
# LZMA load duration: 9.240s
# LZ4 dump duration: 1.984s
# LZ4 file size: 152.374MB
# LZ4 load duration: 2.135s
# LZHW compression duration: 81.522s
# LZHW file size: 45.755MB
# LZHW decompression duration: 48.904s

Let's visualize the comparison:

import numpy as np
import matplotlib.pyplot as plt

N = 5
load_durations = (raw_load_duration, zlib_load_duration,
                  lzma_load_duration, lz4_load_duration, lzhw_d_duration)
dump_durations = (raw_dump_duration, zlib_dump_duration,
                  lzma_dump_duration, lz4_dump_duration, lzhw_compression_duration)
file_sizes = (raw_file_size, zlib_file_size, lzma_file_size, lz4_file_size, lzhw_file_size)
ind = np.arange(N)
width = 0.5

plt.figure(1, figsize=(5, 4))
p1 = plt.bar(ind, dump_durations, width)
p2 = plt.bar(ind, load_durations, width, bottom=dump_durations)
plt.ylabel('Time in seconds')
plt.title('Compression & Decompression durations\nof different algorithms')
plt.xticks(ind, ('Raw', 'Zlib', 'LZMA', "LZ4", "LZHW"))
plt.legend((p1[0], p2[0]), ('Compression duration', 'Decompression duration'))

plt.figure(2, figsize=(5, 4))
plt.bar(ind, file_sizes, width, log=True)
plt.ylabel('File size in MB')
plt.xticks(ind, ('Raw', 'Zlib', 'LZMA', "LZ4", "LZHW"))
plt.title('Compressed data size\nof different algorithms')
for index, value in enumerate(file_sizes):
    plt.text(index, value, str(round(value)) + "MB")

By far LZHW outperforms others with acceptable time difference