Python – Finding duplicate files and removing them


I am writing a Python program to find and remove duplicate files from a folder.

I have multiple copies of mp3 files, and some other files. I am using the sh1 algorithm.

How can I find these duplicate files and remove them?

Best Solution

Fastest algorithm - 100x performance increase compared to the accepted answer (really :))

The approaches in the other solutions are very cool, but they forget about an important property of duplicate files - they have the same file size. Calculating the expensive hash only on files with the same size will save tremendous amount of CPU; performance comparisons at the end, here's the explanation.

Iterating on the solid answers given by @nosklo and borrowing the idea of @Raffi to have a fast hash of just the beginning of each file, and calculating the full one only on collisions in the fast hash, here are the steps:

  1. Buildup a hash table of the files, where the filesize is the key.
  2. For files with the same size, create a hash table with the hash of their first 1024 bytes; non-colliding elements are unique
  3. For files with the same hash on the first 1k bytes, calculate the hash on the full contents - files with matching ones are NOT unique.

The code:

#!/usr/bin/env python
# if running in py3, change the shebang, drop the next import for readability (it does no harm in py3)
from __future__ import print_function   # py2 compatibility
from collections import defaultdict
import hashlib
import os
import sys

def chunk_reader(fobj, chunk_size=1024):
    """Generator that reads a file in chunks of bytes"""
    while True:
        chunk =
        if not chunk:
        yield chunk

def get_hash(filename, first_chunk_only=False, hash=hashlib.sha1):
    hashobj = hash()
    file_object = open(filename, 'rb')

    if first_chunk_only:
        for chunk in chunk_reader(file_object):
    hashed = hashobj.digest()

    return hashed

def check_for_duplicates(paths, hash=hashlib.sha1):
    hashes_by_size = defaultdict(list)  # dict of size_in_bytes: [full_path_to_file1, full_path_to_file2, ]
    hashes_on_1k = defaultdict(list)  # dict of (hash1k, size_in_bytes): [full_path_to_file1, full_path_to_file2, ]
    hashes_full = {}   # dict of full_file_hash: full_path_to_file_string

    for path in paths:
        for dirpath, dirnames, filenames in os.walk(path):
            # get all files that have the same size - they are the collision candidates
            for filename in filenames:
                full_path = os.path.join(dirpath, filename)
                    # if the target is a symlink (soft one), this will 
                    # dereference it - change the value to the actual target file
                    full_path = os.path.realpath(full_path)
                    file_size = os.path.getsize(full_path)
                except (OSError,):
                    # not accessible (permissions, etc) - pass on

    # For all files with the same file size, get their hash on the 1st 1024 bytes only
    for size_in_bytes, files in hashes_by_size.items():
        if len(files) < 2:
            continue    # this file size is unique, no need to spend CPU cycles on it

        for filename in files:
                small_hash = get_hash(filename, first_chunk_only=True)
                # the key is the hash on the first 1024 bytes plus the size - to
                # avoid collisions on equal hashes in the first part of the file
                # credits to @Futal for the optimization
                hashes_on_1k[(small_hash, size_in_bytes)].append(filename)
            except (OSError,):
                # the file access might've changed till the exec point got here 

    # For all files with the hash on the 1st 1024 bytes, get their hash on the full file - collisions will be duplicates
    for __, files_list in hashes_on_1k.items():
        if len(files_list) < 2:
            continue    # this hash of fist 1k file bytes is unique, no need to spend cpy cycles on it

        for filename in files_list:
                full_hash = get_hash(filename, first_chunk_only=False)
                duplicate = hashes_full.get(full_hash)
                if duplicate:
                    print("Duplicate found: {} and {}".format(filename, duplicate))
                    hashes_full[full_hash] = filename
            except (OSError,):
                # the file access might've changed till the exec point got here 

if __name__ == "__main__":
    if sys.argv[1:]:
        print("Please pass the paths to check as parameters to the script")

And, here's the fun part - performance comparisons.

Baseline -

  • a directory with 1047 files, 32 mp4, 1015 - jpg, total size - 5445.998 MiB - i.e. my phone's camera auto upload directory :)
  • small (but fully functional) processor - 1600 BogoMIPS, 1.2 GHz 32L1 + 256L2 Kbs cache, /proc/cpuinfo:

Processor : Feroceon 88FR131 rev 1 (v5l) BogoMIPS : 1599.07

(i.e. my low-end NAS :), running Python 2.7.11.

So, the output of @nosklo's very handy solution:

root@NAS:InstantUpload# time ~/scripts/ 
Duplicate found: ./IMG_20151231_143053 (2).jpg and ./IMG_20151231_143053.jpg
Duplicate found: ./IMG_20151125_233019 (2).jpg and ./IMG_20151125_233019.jpg
Duplicate found: ./IMG_20160204_150311.jpg and ./IMG_20160204_150311 (2).jpg
Duplicate found: ./IMG_20160216_074620 (2).jpg and ./IMG_20160216_074620.jpg

real    5m44.198s
user    4m44.550s
sys     0m33.530s

And, here's the version with filter on size check, then small hashes, and finally full hash if collisions are found:

root@NAS:InstantUpload# time ~/scripts/ . "/i-data/51608399/photo/Todor phone"
Duplicate found: ./IMG_20160216_074620 (2).jpg and ./IMG_20160216_074620.jpg
Duplicate found: ./IMG_20160204_150311.jpg and ./IMG_20160204_150311 (2).jpg
Duplicate found: ./IMG_20151231_143053 (2).jpg and ./IMG_20151231_143053.jpg
Duplicate found: ./IMG_20151125_233019 (2).jpg and ./IMG_20151125_233019.jpg

real    0m1.398s
user    0m1.200s
sys     0m0.080s

Both versions were ran 3 times each, to get the avg of the time needed.

So v1 is (user+sys) 284s, the other - 2s; quite a diff, huh :) With this increase, one could go to SHA512, or even fancier - the perf penalty will be mitigated by the less calculations needed.


  • More disk access than the other versions - every file is accessed once for size stats (that's cheap, but still is disk IO), and every duplicate is opened twice (for the small first 1k bytes hash, and for the full contents hash)
  • Will consume more memory due to storing the hash tables runtime