1.Preparing

1.0 prepare environment

environment.txt

torch version

conda install pytorch==1.10.2 torchvision==0.11.3 torchaudio==0.10.2 cudatoolkit=11.3.1 -c pytorch -c conda-forge

transformers

pip install transformers==3.4.0

1.1 Prepare data

first put InterMWP dataset in 'data/'

change PLM's vocab, the top two lines in vocab.text in 'chinese-bert-wwm' should be modified as:

[PAD]
[NUM]
...

2. Training Procedure

2.0 Train GTS(Bert) Baseline

run_bert2tree.py

CUDA_VISIBLE_DEVICES=0 python run_bert2tree.py --save_path model/finetune/bert2tree --save

Note: We have already proveid the trained retriever's generated promots, you can skip to 2.2

2.1 train retriever

SoftmaxBCELoss:

CUDA_VISIBLE_DEVICES=0 python run_retriever_attn.py --save_path model/retriever/attn_ep[30] --n_epochs 30

2.2 prompt-enhanced learning for MWP

top K prompts learning:

CUDA_VISIBLE_DEVICES=2 python run_bert2tree.py --save_path model/finetune/bert2tree_top3 --save --logic_path data/logic.json --prompt_text_path model/retriever/attn_ep[30] --retriever_postfix 3

Then generate logic generator training data:

• GTS(BERT)

  CUDA_VISIBLE_DEVICES=5 python evaluate_bert2tree.py --save_path model/finetune/bert2tree/ --save_data_path data/bert2tree --save
  

• InterSolver

  CUDA_VISIBLE_DEVICES=5 python evaluate_bert2tree.py --save_path model/finetune/bert2tree_top3 --logic_path data/logic.json --prompt_text_path model/retriever/attn_ep[30] --retriever_postfix 3 --save_data_path data/bert2tree_bce_ep20_top3 --save
  

2.3 train logic generator

GTS(BERT)-[logic generator]

CUDA_VISIBLE_DEVICES=5 python run_logic_generator.py --data_path data/bert2tree/ --save_path model/logic_generator/bert2tree/ --save --test_while_valid

InterSolver-[logic generator]

CUDA_VISIBLE_DEVICES=5 python run_logic_generator.py --data_path data/bert2tree_bce_ep20_top3/ --save_path model/logic_generator/bert2tree_bce_ep20_top3/ --save --test_while_valid