Note
Go to the end to download the full example code
Label Drift#
This notebooks provides an overview for using and understanding label drift check.
Structure:
What Is Label Drift?#
Drift is simply a change in the distribution of data over time, and it is also one of the top reasons why machine learning model’s performance degrades over time.
Label drift is when drift occurs in the label itself.
For more information on drift, please visit our Drift User Guide.
How Deepchecks Detects Label Drift#
This check detects label drift by using univariate measures on the label properties.
Using Label Properties to Detect Label Drift#
In computer vision specifically, our labels may be complex, and measuring their drift is not a straightforward task. Therefore, we calculate drift on different properties of the label, on which we can directly measure drift.
Which Label Properties Are Used?#
Task Type |
Property name |
What is it |
|---|---|---|
Classification |
Samples Per Class |
Number of images per class |
Object Detection |
Samples Per Class |
Number of bounding boxes per class |
Object Detection |
Bounding Box Area |
Area of bounding box (height * width) |
Object Detection |
Number of Bounding Boxes Per Image |
Number of bounding box objects in each image |
Run the check on a Classification task (MNIST)#
Imports#
Note
In this example, we use the pytorch version of the mnist dataset and model. In order to run this example using tensorflow, please change the import statements to:
from deepchecks.vision.datasets.classification.mnist_tensorflow import load_dataset
from deepchecks.vision.checks import LabelDrift
from deepchecks.vision.datasets.classification.mnist_torch import load_dataset
Loading Data#
train_ds = load_dataset(train=True, batch_size=64, object_type='VisionData')
test_ds = load_dataset(train=False, batch_size=1000, object_type='VisionData')
Running LabelDrift on classification#
check = LabelDrift()
result = check.run(train_ds, test_ds)
result.show()
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:04]
Processing Test Batches:
|█████| 1/1 [Time: 00:04]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Understanding the results#
We can see there is almost no drift between the train & test labels. This means the split to train and test was good (as it is balanced and random). Let’s check the performance of a simple model trained on MNIST.
from deepchecks.vision.checks import ClassPerformance
ClassPerformance().run(train_ds, test_ds)
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
To display the results in an IDE like PyCharm, you can use the following code:
# ClassPerformance().run(train_ds, test_ds, mnist_model).show_in_window()
The result will be displayed in a new window.
MNIST with label drift#
Now, let’s try to separate the MNIST dataset in a different manner that will result in a label drift, and see how it affects the performance. We are going to create a custom collate_fn` in the test dataset, that will select samples with class 0 with a probability of 1/10.
Inserting drift to the test set#
import numpy as np
np.random.seed(42)
def generate_collate_fn_with_label_drift(collate_fn):
def collate_fn_with_label_drift(batch):
batch_dict = collate_fn(batch)
images = batch_dict['images']
labels = batch_dict['labels']
for i in range(len(images)):
image, label = images[i], labels[i]
if label == 0:
if np.random.randint(5) != 0:
batch_dict['labels'][i] = 1
return batch_dict
return collate_fn_with_label_drift
mod_test_ds = load_dataset(train=False, batch_size=1000, object_type='VisionData')
mod_test_ds._batch_loader.collate_fn = generate_collate_fn_with_label_drift(mod_test_ds._batch_loader.collate_fn)
Run the check#
check = LabelDrift()
check.run(train_ds, mod_test_ds)
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:04]
Processing Test Batches:
|█████| 1/1 [Time: 00:04]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Add a condition#
We could also add a condition to the check to alert us to changes in the label distribution, such as the one that occurred here.
check = LabelDrift().add_condition_drift_score_less_than()
check.run(train_ds, mod_test_ds)
# As we can see, the condition alerts us to the presence of drift in the label.
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Results#
We can see the check successfully detects the (expected) drift in class 0 distribution between the train and test sets
But how does this affect the performance of the model?#
ClassPerformance().run(train_ds, mod_test_ds)
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Train Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Processing Test Batches:
|█████| 1/1 [Time: 00:01]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Understanding the results#
We can see the drop in the precision of class 0, which was caused by the class imbalance indicated earlier by the label drift check.
Run the check on an Object Detection task (COCO)#
Note
In this example, we use the pytorch version of the coco dataset and model. In order to run this example using tensorflow, please change the import statements to:
from deepchecks.vision.datasets.detection.coco_tensorflow import load_dataset
from deepchecks.vision.datasets.detection.coco_torch import load_dataset
train_ds = load_dataset(train=True, object_type='VisionData')
test_ds = load_dataset(train=False, object_type='VisionData')
check = LabelDrift()
check.run(train_ds, test_ds)
Processing Train Batches:
| | 0/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:00]
Processing Train Batches:
|█████| 1/1 [Time: 00:00]
Processing Test Batches:
| | 0/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:00]
Processing Test Batches:
|█████| 1/1 [Time: 00:00]
Computing Check:
| | 0/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Computing Check:
|█████| 1/1 [Time: 00:00]
Label drift is detected!#
We can see that the COCO128 contains a drift in the out of the box dataset. In addition to the label count per class, the label drift check for object detection tasks include drift calculation on certain measurements, like the bounding box area and the number of bboxes per image.
Total running time of the script: (0 minutes 22.448 seconds)