Package 'leakaudit'

Produce a leak-free split assignment

Description

Resolves duplicate groups that span multiple splits by assigning every member of a leaked group to a single split, chosen by priority. This mirrors the common convention of keeping duplicates in train (so no information is discarded, only potential evaluation leakage) but any priority order can be supplied.

Usage

clean_splits(grouped_df, priority = c("train", "val", "test"))

Arguments

grouped_df	A data frame with columns path, group_id, and split, as produced by find_duplicate_groups().
priority	Character vector giving split precedence, highest priority first. Defaults to c("train", "val", "test"): a leaked group touching train and test is fully reassigned to train.

Value

grouped_df with split replaced by the corrected, leak-free assignment, plus a logical column reassigned flagging which rows were changed.

Examples

groups <- data.frame(
  path = c("a.jpg", "b.jpg", "c.jpg"),
  split = c("train", "test", "val"),
  group_id = c(1, 1, 2)
)
clean_splits(groups)

---

Compute perceptual hashes for a set of images

Description

Computes a 64-bit difference hash (dHash) for each image path. dHash is robust to resizing, mild compression, and small crops, which makes it well suited to catching near-duplicate images that differ only in resolution, format, or minor edits – the kind of duplication that commonly leaks across train/validation/test splits.

Usage

compute_hashes(paths, split = NULL)

Arguments

paths	Character vector of file paths to image files. Any format readable by magick::image_read() is supported (jpg, png, tif, ...).
split	A character vector the same length as paths giving the split each image currently belongs to (e.g. "train", "val", "test"). Optional; if omitted, all images are treated as unsplit.

Value

A data frame with columns path, split, and hash (a 16 character hex string encoding the 64-bit dHash). Any path that fails to load produces NA in hash with a warning, rather than stopping the whole run.

Examples

## Not run: 
hashes <- compute_hashes(
  paths = list.files("images/", full.names = TRUE),
  split = rep(c("train", "test"), length.out = 10)
)

## End(Not run)

---

Audit train/val/test leakage from duplicate groups

Description

Given duplicate groups (from find_duplicate_groups()) and each image's split assignment, reports how many groups span more than one split – i.e. how many near-duplicate images "leak" across the train/validation/test boundary – and what fraction of each split is affected.

Usage

dhash_audit(grouped_df)

Arguments

grouped_df

A data frame with columns path, group_id, and split, as produced by find_duplicate_groups() (with split supplied to compute_hashes()).

Value

An object of class leakage_report (a list) with:

leaked_groups

data frame of group_ids that span >1 split, with the splits involved and member count

n_leaked_images

total images sitting in a leaked group

pct_leaked_by_split

named numeric vector, % of each split's images that belong to a leaked group

total_images

total images audited

Examples

groups <- data.frame(
  path = c("a.jpg", "b.jpg", "c.jpg", "d.jpg"),
  split = c("train", "test", "train", "val"),
  group_id = c(1, 1, 2, 3)
)
dhash_audit(groups)

---

Group near-duplicate images by hash distance

Description

Given a table of image hashes (as produced by compute_hashes()), clusters images into duplicate groups: any two images within threshold Hamming distance of each other are placed in the same group, and group membership propagates transitively (A~B~C all end up in one group even if A and C aren't directly close).

Usage

find_duplicate_groups(hash_df, threshold = 5)

Arguments

hash_df	A data frame with at least columns path and hash, as returned by compute_hashes().
threshold	Maximum Hamming distance (in bits, out of 64) for two images to be considered near-duplicates. Defaults to 5, a reasonably conservative value for dHash; lower it for stricter (exact-duplicate-only) matching, raise it to catch more aggressive near-duplicates at the cost of more false positives.

Value

hash_df with an added integer column group_id. Images with no near-duplicates get their own unique group_id. Row order is preserved.

Examples

hashes <- data.frame(
  path = c("a.jpg", "b.jpg", "c.jpg", "d.jpg"),
  hash = c("0000000000000000", "0000000000000001",
           "ffffffffffffffff", "1234567890abcdef")
)
find_duplicate_groups(hashes, threshold = 5)

---

Hamming distance between two hex-encoded perceptual hashes

Description

Computes the number of differing bits between two perceptual hashes encoded as hexadecimal strings (as produced by compute_hashes()).

Usage

hamming_distance(hash1, hash2)

Arguments

hash1	A character scalar or vector of hex-encoded hashes.
hash2	A character scalar or vector of hex-encoded hashes, the same length as hash1 (or length 1, recycled).

Value

An integer vector of Hamming distances (number of differing bits).

Examples

hamming_distance("ff00", "ff01")
hamming_distance(c("ffff", "0000"), "0000")

---