Data-centric-AI-perspective

This repo provides the data and code for reproducing the the image classification experiments.

Data

• Please download the prepared dataset data-centric-AI-perspective.zip from this google drive url.

• Extract the files. The resulting file structure will look like:

.
├── README.md
├── data/
    ├── train/              
        ├── cat/
            ├── <ID>.jpg
            ├── ...
        ├── dog/
            ├── <ID>.jpg
            ├── ...
    ├── val/
        ├── cat/
            ├── <ID>.jpg
            ├── ...
        ├── dog/
            ├── <ID>.jpg
            ├── ...
    ├── shapley_val/
        ├── cat/
            ├── <ID>.jpg
            ├── ...
        ├── dog/
            ├── <ID>.jpg
            ├── ...
├── code/
    ├── outputs/ 
    ├── step_1_baseline_finetune.ipynb
    ├── step_2a_extract_features.ipynb
    ├── step_2b_compute_data_shapley.py
    ├── step_3_using_cleaned_data.ipynb              

Code

• Execute the scripts/notebooks in the following order, as indicated by their file names:
  • step_1_baseline_finetune.ipynb: performs standard fine-tuning on the given dataset. The notebook is adopted from the official PyTorch tutorial on fine-tuning torchvision models.
  • step_2a_extract_features.ipynb: extracts image features. The features are used when applying the data Shapley method to the noisy training set.
  • step_2b_compute_data_shapley.py: estimates the Shapley value of each training point.
  • step_3_using_cleaned_data.ipynb: drops the training data points with negative Shapley value, and trains the classifier on the remaining training data points.

Dependencies

• Python 3.6.13 (e.g. conda create -n venv python=3.6.13)
• PyTorch Version: 1.4.0
• Torchvision Version: 0.5.0