Awesome

Pipelines for Training and Evaluation

Data-to-text (DART)

• Folder: /dart/

• Datasets: dart (Google drive link)

• Teacher model

  • Init: bart-large
  • Fine-tune: scripts/run_finetune_teacher.sh (Google drive link)

• Student init model:

  • Pre-train init student model: scripts/run_pretrain_student_distill.sh (Google drive link)

• Generate pseudo-target with Teacher model (useful for Seqkd, JS, TVD)

  • scripts/run_teacher_label_all.sh (Google drive link), replace train.json

• Run KD methods (/scripts/)

  • Seqkd: run_seqkd.sh
  • ENGINE: run_engine.sh
  • RKL: run_rkl.sh
  • KL: run_kl_sample.sh
  • JS: run_js.sh
  • TVD: run_tvd_symm.sh

• Decode (/scripts/) (Google drive link)

  • run_eval.sh

• Eval

  • cd /evaluation/ (Google drive link)
  • sh ./run_eval_on_dart.sh (need to modify the $OUTPUT_FILE and download bert-base-uncased model)

• Calculate coverage loss (PPL of teacher)

  • run python3 dart/run_calc_ppl.py --reference_path [teacher_output_path] --input_path [input_path] --model_name [student_model_path] --save_path /tmp/

• Calculate likelihood loss (PPL of student)

  • run python3 dart/run_calc_ppl.py --model_name [teacher_model_path] --input_path [input_path] --reference_path [student_output_path] --save_path /tmp/

Summarization

• Folder: /summa/
• Dataset: xsum (wget https://cdn-datasets.huggingface.co/summarization/xsum.tar.gz)
• Teacher model: https://huggingface.co/facebook/bart-large-xsum
• Student init model:
  • Pre-train init student model: run_pretrain_student_distill.sh (Google drive link)
• Generate pseudo-target with Teacher model (useful for Seqkd, JS, TVD)
  • run_teacher_label.sh (Google drive link), replace train.target
• Run KD methods (/scripts/)
  • Seqkd: run_seqkd.sh
  • ENGINE: run_engine.sh
  • RKL: run_rkl.sh
  • KL: run_kl.sh
  • JS: run_js.sh
  • TVD: run_tvd_symm.sh
• Decode and Evaluate
  • run eval.sh
• Calculate coverage loss (PPL of teacher)
  • run python3 summa/run_calc_ppl.py --reference_path [teacher_output_path] --input_path [input_path] --model_name [student_model_path] --save_path /tmp/
• Calculate likelihood loss (PPL of student)
  • run python3 summa/run_calc_ppl.py --model_name [teacher_model_path] --input_path [input_path] --reference_path [student_output_path] --save_path /tmp/

Machine Translation (WMT16 EN-RO, 100k training data)

• Folder: /t5mt/
• Dataset: wmt_en_ro_100k (Google drive link)
• Teacher model
  • Init: t5-base
  • Fine-tune: scripts_sm/run_finetune_teacher.sh
• Student init model:
  • Pre-train init student model: scripts_sm/run_pretrain_student_distill.sh (Google drive link)
• Generate pseudo-target with Teacher model (useful for Seqkd, JS, TVD)
  • scripts_sm/run_teacher_label.sh (Google drive link), replace train.target
• Run KD methods (/scripts/)
  • Seqkd: run_seqkd.sh
  • ENGINE: run_engine.sh
  • RKL: run_rkl.sh
  • KL: run_kl.sh
  • JS: run_js.sh
  • TVD: run_tvd_symm.sh
• Decode and eval
  • sh scripts_sm/run_eval.sh
• Calculate coverage loss (PPL of teacher)
  • run python3 t5mt/run_calc_ppl.py --reference_path [teacher_output_path] --input_path [input_path] --model_name [student_model_path] --save_path /tmp/
• Calculate likelihood loss (PPL of student)
  • run python3 t5mt/run_calc_ppl.py --model_name [teacher_model_path] --input_path [input_path] --reference_path [student_output_path] --save_path /tmp/

Chat

• Folder: /chat/
• Dataset: Commonsense-Dialogues (Google drive link)
• Teacher model
  • Init: microsoft/DialoGPT-medium
  • Fine-tune: scripts/run_finetune_teacher.sh (Google drive link)
• Student init model:
  • Pre-train init student model: scripts/run_pretrain_student_distill.sh (Google drive link)
• Generate pseudo-target with Teacher model (useful for Seqkd, JS, TVD) (Google drive link), replace train.target
  • scripts/run_teacher_label.sh
• Run KD methods (/scripts/)
  • Seqkd: run_seqkd.sh
  • ENGINE: run_engine.sh
  • RKL: run_rkl.sh
  • KL: run_kl.sh
  • JS: run_js.sh
  • TVD: run_tvd_symm.sh
• Decode and eval
  • sh scripts/run_eval.sh (need to download bert-base-uncased model)
• Calculate coverage loss (PPL of teacher)
  • run python3 chat/run_calc_ppl.py --reference_path [teacher_output_path] --input_path [input_path] --model_name [student_model_path] --save_path /tmp/
• Calculate likelihood loss (PPL of student)
  • run python3 chat/run_calc_ppl.py --model_name [teacher_model_path] --input_path [input_path] --reference_path [student_output_path] --save_path /tmp/

Acknowledgements

• The methods in our codebase are mainly implemented with PyTorch and Huggingface's Transformers libararies
• The pre-distillation part is based on the method proposed in Shleifer & Rush (2020) and their implementation
• We use Maluuba's nlg-eval to measure the BLEU score for the dialogue task.