Pytorch implementation of CascadeXML: Rethinking Transformers for End-to-end Multi-resolution Training in Extreme Multi-label Classification.
CascadeXML is tested on datasets in The Extreme Classification Repository. Amazon-670K, Amazon-3M, Wiki-500K, AmazonCat-13K and Wiki10-31K are supported. Before training, the datasets have to be prepared and label clusters have to be created. We use Eclare style clusters for CascadeXML.
Run the below command, this will create a new conda environment named cascadexml with all the dependencies required to run CascadeXML.
conda create -n cascadexml python==3.10
conda activate cascadexml
bash install.sh
Prepare the dataset and clusters and run the src/main.py
An example command to train CascadeXML on Wiki10-31K dataset on a single GPU is provided below:
python src/main.py --num_epochs 15 --dataset Wiki10-31K --batch_size 64 --max_len 256 --mn Cascade_Wiki10-31K --topk 64 64 --cluster_name Eclusters_54.pkl --rw_loss 1 1 1 --embed_drops 0.3 0.3 0.4 --warmup 2
Note: Large datasets like Amazon-3M may not fit on a single GPU. For such datasets we recommend enabling the --sparse flag. This flag uses a SparseAdam optimizer instead of standard AdamW and helps reduce memory costs. Unfortunately sparse gradients are not supported by nvcc so DDP cannot be used in this mode. Upto 2 GPUs can be used in sparse mode for model parallel.
Amazon-670K
python src/main.py --lr 1e-4 --num_epochs 15 --dataset Amazon-670K --batch_size 64 --max_len 128 --mn CascadeXML_A670K_2111 --topk 128 256 400 --cluster_name Eclusters_1865.pkl --embed_drops 0.2 0.25 0.4 0.5
Wiki-500K
python src/main.py --lr 1e-4 --num_epochs 15 --dataset Wiki-500K --batch_size 64 --max_len 128 --mn CascadeXML_Wiki500K_2111 --topk 128 256 512 --cluster_name Eclusters_1865.pkl --embed_drops 0.2 0.25 0.35 0.5
If you find this repository useful, please consider giving a star and citing our paper:
@article{Kharbanda2022CascadeXMLRT,
title={CascadeXML: Rethinking Transformers for End-to-end Multi-resolution Training in Extreme Multi-label Classification},
author={Siddhant Kharbanda and Atmadeep Banerjee and Erik Schultheis and Rohit Babbar},
journal={ArXiv},
year={2022},
volume={abs/2211.00640}
}