split#

Dataset.split(input_seed: int = 0, split_names: Sequence[str] = ('train', 'valid'), target_split_shares: Sequence[float] = (0.8, 0.2), keep_separate_groups: str | ContinuousGroup | Sequence[str | ContinuousGroup] = ('image_id',), keep_balanced_groups: str | ContinuousGroup | Sequence[str | ContinuousGroup] = ('category_id',), keep_balanced_groups_weights: Sequence[float] | None = None, inplace: bool = False, hist_cost_weight: float = 1, share_cost_weight: float = 1, earth_mover_regularization: float = 0) → Self[source]#

Perform the split operation on annotations and images.

This algorithm works in 2 steps:

  1. divide the dataframe into atomic sub frames. Given the image and annotation

    attributes that need to be kept separate, we can construct sub frame of elements that cannot be in different splits.

  2. Construct the split dataframes iteratively by trying to keep given column

    values with a balanced repartition between splits, along with keeping split sizes as close to target share as possible. Each atomic sub-dataframe is routed to the split that minimize a cost function which try to optimize repartition targets.

Warning

if self.images and self.annotations each have a column with the same name, the column in self.images will be ignored. Make sure column names are mutually exclusive to avoid problems.

See pandas.split_dataframe()

Parameters:

  • input_seed – Seed used for shuffling sub dataframes before beginning step 2 of splitting algorithm. Defaults to 0.

  • split_names – Names of splits. Must be the same length as split_shares. Defaults to (“train”, “valid”).

  • target_split_shares – List of target relative size of each split. Must be the same length as split_names, and will be normalized so that its sum is 1. Defaults to (0.8, 0.2).

  • keep_separate_groups – columns or groups (see group) in annotations or images DataFrame to keep separate. That is for a particular column or group, two rows with the same value cannot be in different splits. Note that image_id will be added to that list, because split happen at the image level. Defaults to (“image_id”,).

  • keep_balanced_groups – columns or groups (see group) in annotations or images DataFrame to keep balanced. That is for a particular group, the distribution of values is the same between original DataFrame and its split, as much as possible. Defaults to (“category_id”,).

  • keep_balanced_groups_weights – Importance of each group to keep balanced when computing histogram cost. If not None, must be a single float or the same size as keep_separate_groups. Defaults to None.

  • inplace – If set, will modify dataframes inplace. This can silently modify some objects (like Datasets) that use them but has a lower memory footprint. Defaults to False.

  • hist_cost_weight – importance of histogram cost for balanced groups. The higher, the more important the histogram cost will be for the decision of where to put each split. Defaults to 1.

  • share_cost_weight – importance of share cost for balanced groups. The higher, the more important the share cost will be for the decision of where to put each split. Defaults to 1.

  • earth_mover_regularization – Regularization parameter applied to sinkhorn’s algorithm during earth mover distance computation. See lours.dataset.split.balanced_group.earth_mover_distance(). Defaults to 0.

Returns:

new Dataset with the split column populated with the corresponding split names.

See also

Example

>>> from lours.utils.doc_utils import dummy_dataset
>>> example = dummy_dataset(
...     200,
...     n_attribute_columns_images={"balanced": 10, "separate": 10},
...     split_names=None,
...     seed=1,
... )
>>> example
Dataset object containing 200 images and 2 objects
Name :
    shake_effort_many
Images root :
    care/suggest
Images :
    width  height        relative_path   type  balanced separate
id
0      955     488  determine/story.jpg   .jpg      send   system
1      131     895       air/method.bmp   .bmp      note   system
2      229     880   political/lead.jpg   .jpg  anything      law
3      840     384        like/safe.bmp   .bmp  anything   likely
4      953     668      suffer/set.jpeg  .jpeg  training   attack
..     ...     ...                  ...    ...       ...      ...
195    122     437    state/almost.tiff  .tiff  anything     star
196    752     300     weight/tend.jpeg  .jpeg     could     rest
197    554     228  remember/summer.png   .png  anything   system
198    688     605       yet/though.png   .png      note   number
199    243     227   describe/road.tiff  .tiff       end   number

[200 rows x 6 columns]
Annotations :
    image_id category_str  category_id  ...   box_y_min   box_width  box_height
id                                      ...
0         77        reach           22  ...   45.427512   40.116677  318.073851
1        137     marriage           15  ...  202.481384  435.389400  475.375279

[2 rows x 7 columns]
Label map :
{14: 'listen', 15: 'marriage', 22: 'reach'}
>>> example.images["separate"].value_counts()
separate
star      27
likely    27
number    27
attack    22
rest      20
law       18
entire    17
enough    16
system    15
often     11
Name: count, dtype: int64
>>> splitted = example.split(
...     keep_balanced_groups=["balanced"], keep_separate_groups=["separate"]
... )
Splitting annotations ...
Separating input data into atomic chunks
1 chunks to distribute across 2 splits
Splitting images ...
Separating input data into atomic chunks
9 chunks to distribute across 2 splits
>>> splitted
Dataset object containing 200 images and 2 objects
Name :
    shake_effort_many
Images root :
    care/suggest
Images :
    width  height        relative_path   type  split  balanced separate
id
0      955     488  determine/story.jpg   .jpg  train      send   system
1      131     895       air/method.bmp   .bmp  train      note   system
2      229     880   political/lead.jpg   .jpg  valid  anything      law
3      840     384        like/safe.bmp   .bmp  train  anything   likely
4      953     668      suffer/set.jpeg  .jpeg  train  training   attack
..     ...     ...                  ...    ...    ...       ...      ...
195    122     437    state/almost.tiff  .tiff  train  anything     star
196    752     300     weight/tend.jpeg  .jpeg  valid     could     rest
197    554     228  remember/summer.png   .png  train  anything   system
198    688     605       yet/though.png   .png  train      note   number
199    243     227   describe/road.tiff  .tiff  train       end   number

[200 rows x 7 columns]
Annotations :
    image_id category_str  category_id  ...   box_y_min   box_width  box_height
id                                      ...
0         77        reach           22  ...   45.427512   40.116677  318.073851
1        137     marriage           15  ...  202.481384  435.389400  475.375279

[2 rows x 8 columns]
Label map :
{14: 'listen', 15: 'marriage', 22: 'reach'}
>>> splitted.images.groupby("split")["separate"].value_counts()
split  separate
train  star        27
       likely      27
       number      27
       attack      22
       entire      17
       enough      16
       system      15
       rest         0
       law          0
       often        0
valid  rest        20
       law         18
       often       11
       entire       0
       star         0
       attack       0
       likely       0
       system       0
       enough       0
       number       0
Name: count, dtype: int64
>>> splitted.images.groupby("split")["balanced"].value_counts()
split  balanced
train  could       21
       coach       20
       end         20
       firm        17
       send        16
       anything    14
       training    13
       lead        10
       note        10
       region      10
valid  could        8
       send         8
       note         6
       firm         5
       anything     5
       training     4
       region       4
       end          4
       coach        3
       lead         2
Name: count, dtype: int64
On this page

This Page