bentrevett / CodeSearchNet

CodeSearchNet

Setup

1. download_data.py to get data for all 6 languages
2. create_6L_data.py to create 6L dataset
3. create_5L_data.py to create 5L dataset
4. python bpe.py --lang java --vocab_max_size 10000 --bpe_pct 0.5 to create bpe dataset for java
5. python bpe.py --lang 6L --vocab_max_size 10000 --bpe_pct 0.5 to create bpe dataset for 6L
6. python bpe.py --lang 5L --vocab_max_size 10000 --bpe_pct 0.5 to create bpe dataset for 5L
7. python create_vocab.py to create vocabularies

Initial Experiments

1. python main_code_retrieval.py --lang java --model transformer --seed 1 model on code retrieval task, in java, w/ random initial weights
2. python main_method_prediction.py --lang java --model transformer --seed 1 model on method prediction task, in java, w/ random initial weights

Experiments w/ Extra Data

1. python main_code_retrieval.py --lang 6L-java --model transformer --seed 1
2. python main_method_prediction.py --lang 6L-java --model transformer --seed 1

Experiments w/ Extra Data (but no Java!)

1. python main_code_retrieval.py --lang 5L-java --model transformer --seed 1
2. pythom main_method_prediction.py --lang 5L-java --model transformer --seed 1

Pre-Training

1. python main_sequence_lm.py --lang java --model --transformer --data code --seed 1
2. python main_sequence_lm.py --lang java --model --transformer --data code --seed 1
3. python main_code_retrieval.py --lang java --model transformer --seed 1 --load model on code retrieval task, in java, fine tuned
4. python main_method_prediction.py --lang java --model transformer --seed 1 --load model on method prediction task, in java, fine tuned