Preparing ImageNet64 Dataset with Clustered Label

Step 1: Acquire ImageNet64

1. Download ImageNet64 from ImageNet Download Page.
2. Extract the *.zip files to unveil the dataset batches, including train_data_batch_1 to train_data_batch_10 and val_data.

Organize the extracted data into:

• Training data: downloadeddata/train
• Validation data: downloadeddata/val

Step 2: Setup Environment

Install the required dependencies to ensure the scripts run smoothly:

pip install -r requirements.txt

Script Execution

• Extracting Images: Run extractimages.py to decode and store images in PNG format, categorizing them into directories named after ImageNet synset IDs.

• Data Clustering: Execute clustereddata.py to organize images based on clustered labels. For instance, all cat images (n02124075, n02123394, n02123159, n02123597, n02123045, n02127052) are clustered, enhancing dataset manageability for training purposes. The script also balances the dataset by equalizing the number of images across clusters.

Additional Resources

For comprehensive label mappings and insights into data clustering, refer to the following resources:

• ImageNet Full Label Mapping
• Clustered Label Mapping
• Downsampled ImageNet