Add Spurious Correlations notebook (#94)

Co-authored-by: Elías Snorrason <[email protected]>

Author: allincowell

README.md

@@ -32,6 +32,7 @@ To quickly learn how to run cleanlab on your own data, first check out the [quic
 | 22  | [non_iid_detection](non_iid_detection/non_iid_detection.ipynb)  | Use Datalab to detect non-IID sampling (e.g. drift) in datasets based on numeric features or embeddings. |
 | 23  | [object_detection](object_detection/README.md)  | Train Detectron2 object detection model for use with cleanlab. |
 | 24  | [semantic segmentation](segmentation/training_ResNeXt50_for_Semantic_Segmentation_on_SYNTHIA.ipynb)  | Train ResNeXt semantic segmentation model for use with cleanlab. |
+| 24  | [spurious correlations](spurious_correlations_datalab/detecting_spurious_correlations.ipynb)  | Train a CNN model on spurious and non-spurious versions of a subset of [Food-101](https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/) dataset. Use `Datalab` to detect issues in the spuriously correlated datasets. |
 
 
 ## Instructions

spurious_correlations_datalab/README.md

@@ -0,0 +1,11 @@
+## Instructions
+
+```console
+$ pip install -r requirements.txt
+```
+
+Change the version of `torch` and `torchvision` if necessary.
+
+Start Jupyter Lab and run the notebook: `detecting_spurious_correlations.ipynb`
+
+In this tutorial, we demonstrate the impact of training a model on a dataset with spurious correlations, focusing on a scenario where one class consists predominantly of dark images. We then compare the model's performance on a dataset free from such spurious correlations. Finally, the tutorial shows how these spurious correlation issues can be easily detected using `Datalab`.

spurious_correlations_datalab/detecting_spurious_correlations.ipynb

@@ -0,0 +1,7 @@
+git+https://github.com/cleanlab/cleanlab.git#egg=cleanlab[all]
+cleanvision
+torch
+torchvision
+matplotlib
+numpy
+datasets