Note
Click here to download the full example code
Custom Dataset - Basics#
Datasets represent a set of recordings that should all be processed in the same way. For example the data of multiple participants in a study, multiple days of recording, or multiple tests. The goal of datasets is to provide a consistent interface to access the raw data, metadata, and potential reference information in an object-oriented way. It is up to you to define, what is considered a single “data-point” for your dataset. Note, that datasets can be arbitrarily nested (e.g. multiple participants with multiple recordings).
Datasets work best in combination with Pipelines
and are further compatible with concepts like GridSearch
and
cross_validation
.
Defining your own dataset#
Fundamentally you only need to create a subclass of Dataset
and define the
create_index
method.
This method should return a dataframe describing all the data-points that should be available in the dataset.
In the following we will create an example dataset, without any real world data, but it can be used to demonstrate most functionality. At the end we will discuss, how gait specific data should be integrated.
We will define an index that contains 5 participants, with 3 recordings each. Recording 3 has 2 trials, while the others have only one. Note, that we implement this as a static index here, but most of the time, you would create the index by e.g. scanning and listing the files in your data directory. It is important that you don’t want to load the entire actual data (e.g. the imu samples) in memory, but just list the available data-points in the index. Then you can filter the dataset first and load the data once you know which data-points you want to access. We will discuss this later in the example.
from itertools import product
from typing import List, Optional, Union
import pandas as pd
trials = list(product(("rec_1", "rec_2", "rec_3"), ("trial_1",)))
trials.append(("rec_3", "trial_2"))
index = [(p, *t) for p, t in product(("p{}".format(i) for i in range(1, 6)), trials)]
index = pd.DataFrame(index, columns=["participant", "recording", "trial"])
index
Now we use this index as the index of our new dataset. To see the dataset in action, we need to create an instance of it. Its string representation will show us the most important information.
from tpcp._dataset import Dataset
class CustomDataset(Dataset):
def create_index(self):
return index
dataset = CustomDataset()
dataset
Subsets#
When working with a dataset, the first thing is usually to select the data you want to use.
For this, you can primarily use the method get_subset
.
Here we want to select only recording 2 and 3 from participant 1 to 4.
Note that the returned subset is an instance of your dataset class as well.
subset = dataset.get_subset(participant=["p1", "p2", "p3", "p4"], recording=["rec_2", "rec_3"])
subset
The subset can then be filtered further. For more advanced filter approaches you can also filter the index directly and use a bool-map to index the dataset
example_bool_map = subset.index["participant"].isin(["p1", "p2"])
final_subset = subset.get_subset(bool_map=example_bool_map)
final_subset
Iteration and Groups#
After selecting the part of the data you want to use, you usually want/need to iterate over the data to apply your processing steps.
By default, you can simply iterate over all rows. Note, that each row itself is a dataset again, but just with a single entry.
for row in final_subset:
print(row)
print("This row contains {} data-point".format(len(row)), end="\n\n")
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_2 trial_1
This row contains 1 data-point
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_3 trial_1
This row contains 1 data-point
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_3 trial_2
This row contains 1 data-point
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_2 trial_1
This row contains 1 data-point
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_3 trial_1
This row contains 1 data-point
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_3 trial_2
This row contains 1 data-point
However, in many cases, we don’t want to iterate over all rows, but rather iterate over groups of the datasets ( e.g. all participants or all tests) individually. We can do that in 2 ways (depending on what is needed). For example, if we want to iterate over all recordings, we can do this:
for trial in final_subset.iter_level("recording"):
print(trial, end="\n\n")
CustomDataset [2 groups/rows]
participant recording trial
0 p1 rec_2 trial_1
1 p2 rec_2 trial_1
CustomDataset [4 groups/rows]
participant recording trial
0 p1 rec_3 trial_1
1 p1 rec_3 trial_2
2 p2 rec_3 trial_1
3 p2 rec_3 trial_2
You can see that we get two subsets, one for each recording label. But what, if we want to iterate over the participants and the recordings together? In these cases, we need to group our dataset first. Note that the grouped_subset shows the new groupby columns as the index in the representation and the length of the dataset is reported to be the number of groups.
grouped_subset = final_subset.groupby(["participant", "recording"])
print("The dataset contains {} groups.".format(len(grouped_subset)))
grouped_subset
The dataset contains 4 groups.
If we now iterate the dataset, it will iterate over the unique groups.
Grouping also changes the meaning of a “single datapoint”.
Each group reports a shape of (1,)
independent of the number of rows in each group.
for group in grouped_subset:
print("This group has the shape {}".format(group.shape))
print(group, end="\n\n")
This group has the shape (1,)
CustomDataset [1 groups/rows]
participant recording trial
participant recording
p1 rec_2 p1 rec_2 trial_1
This group has the shape (1,)
CustomDataset [1 groups/rows]
participant recording trial
participant recording
p1 rec_3 p1 rec_3 trial_1
rec_3 p1 rec_3 trial_2
This group has the shape (1,)
CustomDataset [1 groups/rows]
participant recording trial
participant recording
p2 rec_2 p2 rec_2 trial_1
This group has the shape (1,)
CustomDataset [1 groups/rows]
participant recording trial
participant recording
p2 rec_3 p2 rec_3 trial_1
rec_3 p2 rec_3 trial_2
At any point, you can view all unique groups/rows in the dataset using the groups
attribute.
The order shown here, is the same order used when iterating the dataset.
When creating a new subset, the order might change!
grouped_subset.groups
[CustomDataset(participant='p1', recording='rec_2'), CustomDataset(participant='p1', recording='rec_3'), CustomDataset(participant='p2', recording='rec_2'), CustomDataset(participant='p2', recording='rec_3')]
Note that for an “un-grouped” dataset, this corresponds to all rows.
final_subset.groups
[CustomDataset(participant='p1', recording='rec_2', trial='trial_1'), CustomDataset(participant='p1', recording='rec_3', trial='trial_1'), CustomDataset(participant='p1', recording='rec_3', trial='trial_2'), CustomDataset(participant='p2', recording='rec_2', trial='trial_1'), CustomDataset(participant='p2', recording='rec_3', trial='trial_1'), CustomDataset(participant='p2', recording='rec_3', trial='trial_2')]
In both cases, we can use the group labels (or a subset of them) to index our dataset.
This can be in particular helpful, if you want to recreate specific train test splits provided by cross_validate
.
final_subset.get_subset(groups=final_subset.groups[:3])
If you want, you can also ungroup a dataset again. This can be useful for a nested iteration:
for outer, group in enumerate(grouped_subset):
ungrouped = group.groupby(None)
for inner, subgroup in enumerate(ungrouped):
print(outer, inner)
print(subgroup, end="\n\n")
0 0
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_2 trial_1
1 0
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_3 trial_1
1 1
CustomDataset [1 groups/rows]
participant recording trial
0 p1 rec_3 trial_2
2 0
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_2 trial_1
3 0
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_3 trial_1
3 1
CustomDataset [1 groups/rows]
participant recording trial
0 p2 rec_3 trial_2
Splitting#
If you are evaluating algorithms, it is often important to split your data into a train and a test set, or multiple distinct sets for a cross validation.
The Dataset
objects directly support the sklearn
helper functions for this.
For example, to split our subset into training and testing we can do the following:
from sklearn.model_selection import train_test_split
train, test = train_test_split(final_subset, train_size=0.5)
print("Train:\n", train, end="\n\n")
print("Test:\n", test)
Train:
CustomDataset [3 groups/rows]
participant recording trial
0 p1 rec_3 trial_2
1 p1 rec_2 trial_1
2 p2 rec_3 trial_2
Test:
CustomDataset [3 groups/rows]
participant recording trial
0 p2 rec_3 trial_1
1 p2 rec_2 trial_1
2 p1 rec_3 trial_1
Such splitting always occurs on a data-point level and can therefore be influenced by grouping. If we want to split our datasets into training and testing, but only based on the participants, we can do this:
train, test = train_test_split(final_subset.groupby("participant"), train_size=0.5)
print("Train:\n", train, end="\n\n")
print("Test:\n", test)
Train:
CustomDataset [1 groups/rows]
participant recording trial
participant
p2 p2 rec_2 trial_1
p2 p2 rec_3 trial_1
p2 p2 rec_3 trial_2
Test:
CustomDataset [1 groups/rows]
participant recording trial
participant
p1 p1 rec_2 trial_1
p1 p1 rec_3 trial_1
p1 p1 rec_3 trial_2
In the same way you can use the dataset (grouped or not) with the cross validation helper functions (KFold is just an example, all should work):
CustomDataset [1 groups/rows]
participant recording trial
participant
p2 p2 rec_2 trial_1
p2 p2 rec_3 trial_1
p2 p2 rec_3 trial_2
CustomDataset [1 groups/rows]
participant recording trial
participant
p1 p1 rec_2 trial_1
p1 p1 rec_3 trial_1
p1 p1 rec_3 trial_2
While this works well, it is not always what we want.
Sometimes, we still want to consider each row a single datapoint, but want to prevent that data of e.g. a single
participant is partially put into train- and partially into the test-split.
For this, we can use GroupKFold
in combination with dataset.create_group_labels
.
create_group_labels
generates a unique identifier for each row/group:
group_labels = final_subset.create_group_labels("participant")
group_labels
['p1', 'p1', 'p1', 'p2', 'p2', 'p2']
They can then be used as the group
parameter in GroupKFold
(and similar methods).
Now the data of the two participants is never split between train and test set.
from sklearn.model_selection import GroupKFold
cv = GroupKFold(n_splits=2)
for train, test in cv.split(final_subset, groups=group_labels):
# We only print the train set here
print(final_subset[train], end="\n\n")
CustomDataset [3 groups/rows]
participant recording trial
0 p1 rec_2 trial_1
1 p1 rec_3 trial_1
2 p1 rec_3 trial_2
CustomDataset [3 groups/rows]
participant recording trial
0 p2 rec_2 trial_1
1 p2 rec_3 trial_1
2 p2 rec_3 trial_2
Creating labels also works for datasets that are already grouped. But, the columns that should be contained in the label must be a subset of the groupby columns in this case.
The number of group labels is 4 in this case, as there are only 4 groups after grouping the datset.
group_labels = final_subset.groupby(["participant", "recording"]).create_group_labels("participant")
group_labels
['p1', 'p1', 'p2', 'p2']
Adding Data#
So far we only operated on the index of the dataset. But if we want to run algorithms, we need the actual data (i.e. IMU samples, clinical data, …).
Because the data and the structure of the data can vary widely from dataset to dataset, it is up to you to implement data access. It comes down to documentation to ensure that users access the data in the correct way.
In general we try to follow a couple of conventions to give datasets a consistent feel:
Data access should be provided via
@property
decorator on the dataset objects, loading the data on demand.The names of these properties should follow some the naming scheme (e.g.
data
for the core sensor data) and should return values using the established datatypes (e.g.pd.DataFrames
).The names of values that represent gold standard information (i.e. values you would only have in an evaluation dataset and should never use for training), should have a trailing
_
, which marks them as result similar how sklearn handles it.
This should look something like this:
class CustomDataset(Dataset):
@property
def data(self) -> pd.DataFrame:
# Some logic to load data from disc
raise NotImplementedError()
@property
def sampling_rate_hz(self) -> float:
return 204.8
@property
def reference_events_(self) -> pd.DataFrame:
# Some custom logic to load the gold-standard events of this validation dataset.
# Note the trailing `_` in the name.
raise NotImplementedError()
def create_index(self):
return index
For each of the data-values you need to decide, on which “level” you provide data access. Meaning, do you want/can return data, when there are still multiple participants/recordings in the dataset, or can you only return the data, when there is only a single trial of a single participant left.
Usually, we recommend to always return the data on the lowest logical level (e.g. if you recorded separate IMU
sessions per trial, you should provide access only, if there is just a single trail by a single participant left in
the dataset).
Otherwise, you should throw an error.
This pattern can be simplified using the is_single
or assert_is_single
helper method.
These helpers check based on the provided groupby_cols
if there is really just a single group/row left with the
given groupby settings.
Let’s say data
can be accessed on either a recording
or a trail
level, and segmented_stride_list
can only
be accessed on a trail
level.
Then we could do something like this:
class CustomDataset(Dataset):
@property
def data(self) -> str:
# Note that we need to make our checks from the least restrictive to the most restrictive (if there is only a
# single trail, there is only just a single recording).
if self.is_single(["participant", "recording"]):
return "This is the data for participant {} and rec {}".format(*self.group)
# None -> single row
if self.is_single(None):
return "This is the data for participant {}, rec {} and trial {}".format(*self.group)
raise ValueError(
"Data can only be accessed when their is only a single recording of a single participant in the subset"
)
@property
def sampling_rate_hz(self) -> float:
return 204.8
@property
def segmented_stride_list_(self) -> str:
# We use assert here, as we don't have multiple options.
# (We could also used `None` for the `groupby_cols` here)
self.assert_is_single(["participant", "recording", "trial"], "segmented_stride_list_")
return "This is the segmented stride list for participant {}, rec {} and trial {}".format(*self.group)
def create_index(self):
return index
If we select a single trial (row), we can get data and the stride list:
test_dataset = CustomDataset()
single_trial = test_dataset[0]
print(single_trial.data)
print(single_trial.segmented_stride_list_)
This is the data for participant p1 and rec rec_1
This is the segmented stride list for participant p1, rec rec_1 and trial trial_1
If we only select a recording, we get an error for the stride list:
# We select only recording 3 here, as it has 2 trials.
single_recording = test_dataset.get_subset(recording="rec_3").groupby(["participant", "recording"])[0]
print(single_recording.data)
try:
print(single_recording.segmented_stride_list_)
except Exception as e:
print("ValueError: ", e)
This is the data for participant p1 and rec rec_3
ValueError: The attribute `segmented_stride_list_` of dataset CustomDataset can only be accessed if there is only a single combination of the columns ['participant', 'recording', 'trial'] left in a data subset,
Custom parameter#
Often it is required to pass some parameters/configuration to the dataset. This could be for example the place where the data is stored or if a specific part of the dataset should be included, if some preprocessing should be applied to the data, … .
Such additional configuration can be provided via a custom __init__
and is then available for all methods to be
used.
Note that you must assign the configuration values to attributes with the same name and must not forget to
call super().__init__
class CustomDatasetWithConfig(Dataset):
data_folder: str
custom_config_para: bool
def __init__(
self,
data_folder: str,
custom_config_para: bool = False,
*,
groupby_cols: Optional[Union[List[str], str]] = None,
subset_index: Optional[pd.DataFrame] = None,
):
self.data_folder = data_folder
self.custom_config_para = custom_config_para
super().__init__(groupby_cols=groupby_cols, subset_index=subset_index)
def create_index(self):
# Use e.g. `self.data_folder` to load the data.
return index
Total running time of the script: ( 0 minutes 2.750 seconds)
Estimated memory usage: 11 MB