xing-hu / tl-codesum Goto Github PK

Hi, could you please share how can I solve this error. Thanks

usage: main.py [-h] [-v] [--debug] [--reset] [--reset-learning-rate]
[--learning-rate LEARNING_RATE] [--purge] [--crash-test]
[--decode [DECODE [DECODE ...]]]
[--align [ALIGN [ALIGN ...]]] [--eval [EVAL [EVAL ...]]]
[--train] [--save] [--gpu-id GPU_ID] [--no-gpu NO_GPU]
[--beam-size BEAM_SIZE]
[--len-normalization LEN_NORMALIZATION]
[--no-early-stopping] [--ensemble] [--average]
[--checkpoints CHECKPOINTS [CHECKPOINTS ...]]
[--output OUTPUT] [--max-steps MAX_STEPS]
[--max-test-size MAX_TEST_SIZE] [--remove-unk]
[--raw-output] [--pred-edits] [--model-dir MODEL_DIR]
[--batch-size BATCH_SIZE] [--no-fix]
[--max-output-len MAX_OUTPUT_LEN] [--unk-replace]
[--lexicon LEXICON] [--temperature TEMPERATURE]
[--attn-temperature ATTN_TEMPERATURE]
[--align-encoder-id ALIGN_ENCODER_ID] [--tf-seed TF_SEED]
[--seed SEED]
config
main.py: error: unrecognized arguments: ../config/code2nl.yaml ../config/default.yaml ../config/multi_attn.yaml ../config/transfer_fix.yaml ../config/transfer.yaml

Good morning.
I'm Bernardo Lourenço, a MSc student in Computer Science and Engineering @ Instituto Superior Técnico (Lisbon, Portugal).
I was analyzing the TL-CodeSum dataset and I was thinking to use this dataset to train my model. Can you tell me the repositories where you got each sample of the dataset?
I wish you a good week 😄
Kind regards,
Bernardo Lourenço

After training the api2nl.yaml configure file and get 0.26 valid score，I start with the transfer.yaml training but only get 0.13 valid score.
Is there any problems in my progress?
Looking forward to your reply!

How many epochs did you train each model mentioned in the paper?

When training the "API-Only" model, part of the log is:

hypotheses:18940, references: 18940
corpus_bleu: 0.0160 avg_score: 0.1902
11/14 21:36:38 valid score=0.0160 avg_score=0.1902
11/14 21:36:38 saving model to ../data_pre/model/api2nl/checkpoints
11/14 21:36:38 finished saving model
11/14 21:48:50 step 56000 epoch 6 learning rate 0.429 step-time 0.364 loss 53.745
11/14 21:49:52   valid eval: loss 56.68
11/14 21:49:52 starting decoding
hypotheses:18940, references: 18940
corpus_bleu: 0.0145 avg_score: 0.1670
11/14 21:51:54 valid score=0.0145 avg_score=0.1670
11/14 21:51:54 saving model to ../data_pre/model/api2nl/checkpoints
11/14 21:51:54 finished saving model

Is "avg_score" BLEU-4 score? If it is, the perf at epoch 6 seem far from 0.2645 reported in paper.

Therefore I'd like to know how many epochs you trained each model.

I trained the api model first on your dataset based on api2nl.yaml, then run the transfer model based on transfer.yaml. This is my final result.
I'm looking forward to your reply.

Can you also provide the code or tool for pre-processing source code?
(parsing source code, extracting javadoc, and extracting api sequences etc.)
Thanks!

I am producing the dataset for my project these days, could you share the code for extracting api sequences? Great thanks!