Note
Go to the end to download the full example code
Prediction Drift#
This notebooks provides an overview for using and understanding the vision prediction drift check.
Structure:
What Is Prediction 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.
Prediction drift is when drift occurs in the prediction itself. Calculating prediction drift is especially useful in cases in which labels are not available for the test dataset, and so a drift in the predictions is a direct indication that a change that happened in the data has affected the model’s predictions. If labels are available, it’s also recommended to run the Label Drift check.
For more information on drift, please visit our Drift Guide
How Deepchecks Detects Prediction Drift#
This check detects prediction drift by using univariate measures on the prediction properties.
Using Prediction Properties to Detect Prediction Drift#
In computer vision specifically, our predictions may be complex, and measuring their drift is not a straightforward task. Therefore, we calculate drift on different properties of the prediction, on which we can directly measure drift.
Which Prediction 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 PredictionDrift
from deepchecks.vision.datasets.classification.mnist_torch import load_dataset
Load Dataset#
train_ds = load_dataset(train=True, batch_size=64, object_type='VisionData')
test_ds = load_dataset(train=False, batch_size=64, object_type='VisionData')
Running PredictionDrift on classification#
check = PredictionDrift()
result = check.run(train_ds, test_ds)
result
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]
To display the results in an IDE like PyCharm, you can use the following code:
# result.show_in_window()
The result will be displayed in a new window.
Understanding the results#
We can see there is almost no drift between the train & test predictions. 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]
MNIST with prediction drift#
Now, let’s try to separate the MNIST dataset in a different manner that will result in a prediction drift, and see how it affects the performance. We are going to create a custom collate_fn` in the test dataset, that will select a few of the samples with class 0 and change their most of their predicted classes to 1.
Inserting drift to the test set#
import numpy as np
import torch
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
# In 9/10 cases, the prediction vector will change to match the label
if np.random.randint(10) != 0:
batch_dict['predictions'][i] = torch.tensor([0, 1, 0, 0, 0, 0, 0, 0, 0, 0])
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 = PredictionDrift()
result = check.run(train_ds, mod_test_ds)
result
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 about changes in the prediction distribution, such as the one that occurred here.
check = PredictionDrift().add_condition_drift_score_less_than()
result = check.run(train_ds, mod_test_ds)
result
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]
As we can see, the condition alerts us to the presence of drift in the predictions.
Results#
We can see the check successfully detects the (expected) drift in class 0 distribution between the train and test sets. It means the the model correctly predicted 0 for those samples and so we’re seeing drift in the predictions as well as the labels. We note that this check enabled us to detect the presence of label drift (in this case) without needing actual labels for the test data.
But how does this affect the performance of the model?#
result = ClassPerformance().run(train_ds, mod_test_ds)
result
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]
Inferring 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 = PredictionDrift()
result = check.run(train_ds, test_ds)
result
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]
Prediction drift is detected!#
We can see that the COCO128 contains a drift in the out of the box dataset. In addition to the prediction count per class, the prediction 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 22.784 seconds)