This project contains the code for Continual Learning for Seq2seq models for Code related tasks.

This codebase has some code or ideas ported from the following repositories.

1. CodeXGLUE
2. CodeT5
3. Learning to Prompt

Use the file src/environment.yml to create a conda environment. The following command can be used conda env create --file=environment.yml.

src
│   dataloaders	# dataloader for CL tasks.
│   evaluator	# Original from CodeT5 for evaluation.
│	plots	# File to make some basic plots related to similarity and prompt frequency.
|	tokenizer	# Original from CodeT5 for some tokenization. Not used for us.
└───sh
|	|	final_runs.sh	# Contains commands for main experiments. These can be used as examples to run the code. For more info on the arguments please look at the config.py file.
└───models
|	|	T5prompt.py
└───utils
│   │   metrics.py	# Main file which implements the metrics.
│   │   replay.py	# Main file to implement replay buffer.
│   │	configs.py	# argparse arguments, etc
│   cont_gen.py	# Main file for running CL experiments.
|	analyse.ipynb	# Main file for analysing the query-key matching analysis.
|	run_gen.py	# Original file from codeT5 to finetune on a single file.
|	run_multi_gen.py	# Modified file from CodeT5 to run multitask learning. Has some hacks to get it works for us.

1. Basic Prompt Pooling: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --keys_agg=random --pool_freq --name=pool100 --pool_size=100 --prompt_method=pool --num_prompts_per_task=20 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=8 --eval_batch_size=32 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=final_1 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

2. PP + ER: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --replay=ring --buffer_size=500 --buffer_bs=2 --keys_agg=random --pool_freq --name=pool100_ER500 --pool_size=100 --prompt_method=pool --num_prompts_per_task=20 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=8 --eval_batch_size=32 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=final_1 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

3. PP + TF + ER: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --pool_teacher --num_shared_keys_per_pair=2 --replay=ring --buffer_size=500 --buffer_bs=2 --keys_agg=random --pool_freq --name=pool100_teacher_ER500 --pool_size=100 --prompt_method=pool --num_prompts_per_task=20 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=16 --eval_batch_size=64 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=teacher_tune --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

4. PP + Fixed Assignment: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --pool_freq --prompt_lr=100 --name=pool200_fixed_selection_plr100 --pool_size=200 --prompt_method=pool_fixed --num_pool_prompt_tokens=5 --num_prompts_per_task=20 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=8 --eval_batch_size=32 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=final_1 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

5. ShPT: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --prompt_lr=100 --name=shpt --prompt_method=shpt --num_prompts_per_task=100 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=8 --eval_batch_size=32 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=final_1 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

6. TSPT: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --name=TSPT --prompt_method=tspt --prompt_lr=100 --num_prompts_per_task=100 --train_only_prompts --bleu_samples=5000 --warmup_steps=500 --train_batch_size=8 --eval_batch_size=32 --log_steps=10 --data_num=-1 --save_last_checkpoints --always_save_model --project_name=final_1 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

7. NSL + ER: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --replay=ring --buffer_size=500 --add_task_prefix --add_lang_ids --bleu_samples=5000 --log_steps=10 --data_num=-1 --warmup_steps=500 --save_last_checkpoints --always_save_model --project_name=final_1 --name=nsl_ER500 --stream=concode_none,translate_java-cs,summarize_ruby,refine_small

8. Summ + NSL + ER: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python cont_gen.py --name=nsl_ER500 --replay=ring --buffer_size=500 --add_task_prefix --add_lang_ids --bleu_samples=5000 --log_steps=10 --data_num=-1 --warmup_steps=100 --train_batch_size=8 --eval_batch_size=32 --save_last_checkpoints --always_save_model --project_name=summ_pool --stream=summarize