chinese_ulmfit

Chinese ULMFiT

Universal Language Model Fine-tuning for Text Classification

Download the pre-trained model

Create a virtual environment (you can configure Tsinghua conda source)

conda env create -f env.yml

Unzip Chinese Wikipedia

python -m gensim.scripts.segment_wiki -i -f /data/zhwiki-latest-pages-articles.xml.bz2 -o tmp/wiki2018-11-14.json.gz

Participle Wikipedia

python preprocessing.py segment-wiki --input_file=tmp/wiki2018-11-14.json.gz --output_file=tmp/wiki2018-11-14.words.pkl

Word segmentation material

python preprocessing.py segment-csv --input_file=data/ch_auto.csv --output_file=tmp/ch_auto.words.pkl --label_file=tmp/ch_auto.labels.npy

tokenize Wikipedia material

python preprocessing.py tokenize --input_file=tmp/wiki2018-11-14.words.pkl --output_file=tmp/wiki2018-11-14.ids.npy --mapping_file=tmp/wiki2018-11-14.mapping.pkl

tokenize field materials

python preprocessing.py tokenize --input_file=tmp/ch_auto.words.pkl --output_file=tmp/ch_auto.ids.npy --mapping_file=tmp/ch_auto.mapping.pkl

Pre-training

python pretraining.py --input_file=tmp/wiki2018-11-14.ids.npy --mapping_file=tmp/wiki2018-11-14.mapping.pkl --dir_path=tmp

Fine adjustment

python finetuning.py --input_file=tmp/ch_auto.ids.npy --mapping_file=tmp/ch_auto.mapping.pkl --pretrain_model_file=tmp/models/wiki2018-11-14.h5 --pretrain_mapping_file=tmp/wiki2018-11-14.mapping.pkl --dir_path=tmp --model_id=ch_auto

Training classifiers

python3 train_classifier.py  --id_file=tmp/ch_auto.ids.npy --label_file=tmp/ch_auto.labels.npy --mapping_file=tmp/ch_auto.mapping.pkl  --encoder_file=ch_auto_enc

test

python3 predicting.py --mapping_file=tmp/ch_auto.mapping.pkl --classifier_filename=tmp/models/classifier_1.h5 --num_class=2