train_gen.yml

environment:
  image: pytorch/pytorch:1.5-cuda10.1-cudnn7-devel

  setup:
    - pip install transformers==2.11.0
    - pip install spacy
    - pip install tqdm
    - pip install tensorboard
    - pip install wandb
    - pip install pytorch-lightning==0.9.0
    - pip install rouge-score

    - sudo apt-get update
    - sudo apt-get -y install git
    - git clone https://github.com/NVIDIA/apex
    - cd apex
    - pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
    - cd ..

jobs:
  # train GAR with 'answer' as the target (see conf.py for other options)
  - name: generator_train_nq_A
    command:
    - cd gar
    - GEN_TARGET='answer' python train_generator.py --remark generator_train_nq_A  --train_batch_size 128 --eval_batch_size 256 --ckpt_metric val-ROUGE-1

  # train the generative reader
  - name: train_multi_inputs_trivia
    command:
    - cd gar
    - python train_generator.py --remark train_multi_inputs_trivia --output_dir $$OUTPUT_DIR --data_dir $$DATA_DIR/trivia-multi-inputs --train_batch_size 6 --eval_batch_size 12 --max_source_length 1024 --max_target_length 10 --learning_rate 5e-6 --ckpt_metric val-EM

  # test the generative reader
  - name: test_multi_inputs
    command:
    - cd gar
    - python test_generator.py --input_path $$DATA_DIR/nq-multi-inputs-merge4-CV/test.source --output_path $$DATA_DIR/nq-multi-inputs-merge4-CV/gen-test-samples.txt --bs 32 --model_ckpt checkpoint.ckpt --max_source_length 1024 --max_target_length 10 --remark test_multi_inputs

  # train a closed-book generative reader
  - name: train_closedBook
    command:
    - cd gar
    - python train_generator.py  --remark train_closedBook --output_dir $$OUTPUT_DIR --data_dir $$DATA_DIR/closedBook --train_batch_size 512 --eval_batch_size 1024 --learning_rate 1e-4 --ckpt_metric val-EM --max_source_length 20 --max_target_length 10 --num_train_epochs 50 --ckpt_metric val-EM