Project for the class "Introduction of Natural Language Processing (Fall 2020)"
Clone the repository and Install Anaconda, then create a conda environment for this project and retrieve datasets.
conda env create -f environment.yml
conda activate nlp
# if on leonhard cluster
module load cuda/10.0.130 cudnn/7.5
RECOMMENDED: You can directly download the data folder from THIS LINK, and put it under this repository.
Otherwise, data come from various sources, and the following provides how I scrape and coarsely process them:
It will either load or scrape news conference data from the website, then split by democratic and replican. Under each party, the dataset is split into train, validation and test datasets. Then, it will prepare data for GraphWriter using Open Information Extraction. Lastly, it will prepare data for training PoliIE system
python -m spacy download en_core_web_sm
python dataset.py
Now under data/presidency_project/newsconference/, we should have dem_train.csv, dem_val.csv, dem_test.csv, rep_train.csv, rep_val.csv and rep_test.csv. These six files serve the start to further pre-process needed input files for the following models. The baseline model gpt-2 will directly use these files.
The above command also creates input files for vanilla GraphWriter under data/gwnaive/, and input files for PoliIE under data/scierc
python Questions.py
So far, under the data/ directory, we should have the following structure:
TODO
I suggest to use some computing resources with GPU to train each model. I have used Google Colab and Leonhard Cluster provided by ETH Zurich.
# To Finetune and Generate with Fituned model
python gpt2.py --test ./data/presidency_project/newsconference/rep_test.csv --party Republican --outfn ./generation/gpt2/rep/rep_test_predict.txt --only_gen False
python gpt2.py --test ./data/presidency_project/newsconference/dem_test.csv --party Democratic --outfn ./generation/gpt2/dem/dem_test_predict.txt --only_gen False
# Check more information with
python gpt2.py -h
The training will take quite a long time, and only Tesla V100 fits the memory requirements. Therefore, I provided here a mini-guide when using ETHZ Leonhard to train the model.
# To Start
## Democratic
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=2]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" "python train.py -bsz 1 -t1size 1 -t2size 1 -t3size 1 -datadir ../data/gwnaive/Democratic/ -save ../output/gwnaive/Democratic/ -esz 256 -hsz 256"
## Republican
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=2]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" "python train.py -bsz 4 -t1size 4 -t2size 2 -t3size 1 -datadir ../data/gwnaive/Republican/ -save ../output/gwnaive/Republican/ -esz 256 -hsz 256"
# The training usually takes more than one job, then you need to follow up with last checkpoint you have, e.g, last checkpoint is 9th epoch (0-indexing)
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=2]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" "python train.py -bsz 1 -t1size 1 -t2size 1 -t3size 1 -datadir ../data/presidency_project/newsconference/gwnaive/Democratic/ -save ../output/gwnaive/Democratic/ -ckpt ../output/gwnaive/Democratic/8.vloss-3.715168.lr-0.1 -esz 256 -hsz 256"
To set up the system, follow:
cd scierc
scripts/fetch_required_data.sh
scripts/build_custom_kernels.sh
To train and validate on annotated data over Leonhard cluster
bsub -W 02:00 -N -R "rusage[mem=20480, ngpus_excl_p=4]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" < run_scierc.sh
Use the best trained PoliIE model to automatically label political entities and relations over train, validation and test to generate input data for GraphWriter.
# GENERATE ELMO EMBEDDINGS
### Democratic
python scierc/generate_elmo.py -fn ./data/scierc_predict/dem/train.json -outfn ./data/scierc_predict/dem/train.hdf5
python scierc/generate_elmo.py -fn ./data/scierc_predict/dem/val.json -outfn ./data/scierc_predict/dem/dev.hdf5
python scierc/generate_elmo.py -fn ./data/scierc_predict/dem/test.json -outfn ./data/scierc_predict/dem/test.hdf5
### Republican
python scierc/generate_elmo.py -fn ./data/scierc_predict/rep/train.json -outfn ./data/scierc_predict/rep/train.hdf5
python scierc/generate_elmo.py -fn ./data/scierc_predict/rep/val.json -outfn ./data/scierc_predict/rep/dev.hdf5
python scierc/generate_elmo.py -fn ./data/scierc_predict/rep/test.json -outfn ./data/scierc_predict/rep/test.hdf5
# RUN POLIIE
#republican train
bsub -n 8 -N -R "rusage[mem=2560]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/rep/train.hdf5 --eval_path ./data/scierc_predict/rep/train.json --questions ./data/scierc_predict/rep/train_questions.txt --outfn ./data/gw_scierc/rep/preprocessed.train.tsv --batch_size 128 --ckpt ./data/gw_scierc/rep/train_ckpt.txt'
# republican validation
bsub -n 8 -N -R "rusage[mem=2560]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/rep/dev.hdf5 --eval_path ./data/scierc_predict/rep/val.json --questions ./data/scierc_predict/rep/val_questions.txt --outfn ./data/gw_scierc/rep/preprocessed.val.tsv --batch_size 128 --ckpt ./data/gw_scierc/rep/val_ckpt.txt'
# republican test
bsub -n 8 -N -R "rusage[mem=2560]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/rep/test.hdf5 --eval_path ./data/scierc_predict/rep/test.json --questions ./data/scierc_predict/rep/test_questions.txt --outfn ./data/gw_scierc/rep/preprocessed.test.tsv --batch_size 128 --ckpt ./data/gw_scierc/rep/test_ckpt.txt'
# democratic train
bsub -n 8 -N -R "rusage[mem=5000]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/dem/train.hdf5 --eval_path ./data/scierc_predict/dem/train.json --questions ./data/scierc_predict/dem/train_questions.txt --outfn ./data/gw_scierc/dem/preprocessed.train.tsv --batch_size 128 --ckpt ./data/gw_scierc/dem/train_ckpt.txt'
# democratic validation
bsub -n 8 -N -R "rusage[mem=2560]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/dem/dev.hdf5 --eval_path ./data/scierc_predict/dem/val.json --questions ./data/scierc_predict/dem/val_questions.txt --outfn ./data/gw_scierc/dem/preprocessed.val.tsv --batch_size 128 --ckpt ./data/gw_scierc/dem/val_ckpt.txt'
# democratic test
bsub -n 8 -N -R "rusage[mem=2560]" 'python scierc/predict.py --expt partisan --lm_path_dev ./data/scierc_predict/dem/test.hdf5 --eval_path ./data/scierc_predict/dem/test.json --questions ./data/scierc_predict/dem/test_questions.txt --outfn ./data/gw_scierc/dem/preprocessed.test.tsv --batch_size 128 --ckpt ./data/gw_scierc/dem/test_ckpt.txt'
###### LR 0.01 Embedding size 256 Hidden size 256 ######
# republican
bsub -N -R "rusage[mem=20480, ngpus_excl_p=2]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" "python ./GraphWriter/train.py -bsz 4 -t1size 4 -t2size 2 -t3size 1 -datadir './data/gw_scierc/rep/' -save './gw_scierc_models/rep_lr0.01_256/' -ent_type 'Actor Implementation Institution Achievement OtherPolitical' -esz 256 -hsz 256 -lr 0.01 -ckpt './gw_scierc_models/rep_lr0.01_256/15.vloss-3.703604.lr-0.01'"
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=2]" -R "select[gpu_model0==TeslaV100_SXM2_32GB]" "python ./GraphWriter/train.py -bsz 4 -t1size 4 -t2size 2 -t3size 1 -datadir './data/gw_scierc/rep/' -save './gw_scierc_models/rep/' -ent_type 'Actor Implementation Institution Achievement OtherPolitical'"
python ./GraphWriter/train.py -bsz 1 -t1size 1 -t2size 1 -t3size 1 -datadir './data/gw_scierc/dem/' -save './gw_scierc_models/dem_lr0.01_256/' -ent_type 'Actor Implementation Institution Achievement OtherPolitical' -esz 256 -hsz 256 -lr 0.01
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=8]" "python gpt2.py --test ./data/presidency_project/newsconference/dem_test.csv --party Democratic --outfn ./generation/gpt2/dem/dem_test_predict.txt"
bsub -W 04:00 -N -R "rusage[mem=20480, ngpus_excl_p=8]" "python gpt2.py --test ./data/presidency_project/newsconference/rep_test.csv --party Republican --outfn ./generation/gpt2/rep/rep_test_predict.txt"
For people without ETH Leonhard access, part of the work can also be run on Google Colab Service. Therefore, I provided some noteboooks here (TODO)