zhuhr213 / snn-conv

ICLR 2023, Spiking Convolutional Neural Networks for Text Classification

Spiking Convolutional Neural Networks for Text Classification (ICLR 2023, Poster)

https://openreview.net/forum?id=pgU3k7QXuz0

Install Requirements

conda create -n snn python=3.7
conda activate snn
pip install -r requirements.txt
pip install -r textattack_r.txt

Shift Pre-trained Word Embeddings to [0, 1]

cd data_preprocess
python tensor_encoder.py
python chinese_tensor_encoder.py

Train Tailored Models

python main.py \
--mode train 
--model_mode ann 
--model_type textcnn

Embed (Prepare for Random Initial Embedding)

Conversion + Normalization methods

python main.py \
--mode conversion
--model_mode snn
--model_type textcnn
--conversion_mode normalize
--conversion_normalize_type model_base or data_base

Conversion + Fine-tuning SNNs

python main.py \
--mode conversion
--model_mode snn
--model_type textcnn
--conversion_mode tune

If Random Initial Embedding (without Pretrain)

Prepare Sentence2index

cd data_preprocess
python sent2id.py

Apply Embedding Layer before fed into SNNs

cd data_preprocess
python snn_wopretrain_tensor_encoder.py

Notes

• Please prepare dateset files and pre-trained word embedding in yourself.
• Shell commands above only show how to run the program in different modes. Detailed hyper-parameters can be set as you want.
• Set parameter random_tensor to True when doing conversion + Normalization or fine-tune if you use random initial embedding.