When in Doubt: Improving Classification Performance with Alternating Normalization

Findings of EMNLP 2021

Menglin Jia, Austin Reiter, Ser-Nam Lim, Yoav Artzi and Claire Cardie
Cornell University, Facebook AI

**arXiv**: https://arxiv.org/abs/2109.13449

Environment settings

This project is tested under Python 3.6, pytorch 1.5.0, torchvision 0.6.0

Preparation

1. Download the data and put the data folder under <DATA_PATH>.
   • Simulation data: the randomly constructed arrays are available here
   • DialogRE: project page
   • Ultra-fine entity typing: project page
   • ImageNet: project page
2. For tuning process, you need to check how many cpus you have first. This reporitory assumes the environment has at least 40 cpus.

Simulation experiments

We have provided the randomly generately array produced from step 1 in the data folder.

# step 1. generate random matrices
python experiments_simulation.py --step 1 \
    --data-root <DATA_PATH>
# step 2. get results
python experiments_simulation.py --step 2 \
    --data-root <DATA_PATH>

Empirical experiments

Ultra-fine entity typing

# Denoise
python experiments_text.py \
    --dataset ultrafine_entity_typing \
    --data-root=<DATA_PATH> --model-type denoise

# multitask
python experiments_text.py \
    --dataset ultrafine_entity_typing --data-root=<DATA_PATH>

Relation extraction

python experiments_text.py \
    --dataset dialogue_re --data-root=<DATA_PATH>

ImageNet

# step 1: prepare the imagenet logits and targets for training and val set
python prepare_imagenet.py --out-dir <OUT_DIR> --data-root=<DATA_PATH>

# step 2: get results
python experiments_visual.py --dataset imagenet

License

This repo are released under the CC-BY-NC 4.0 license. See LICENSE for additional details.