AOCG

<pre>### The replication package of AOCG<br><br>The repository is divided into two parts: datasets and the code of our AOCG method.<br><br><br>### Requirements<br>```<br>- python 3.8<br>- Java 1.8.0<br>- transformers 4.5.1<br>- tree-sitter 0.2.2<br>- Pytorch 1.7.1<br>```<br><br>### Data Preprocessing<br>```<br>Experimental datasets contain the API_SUM dataset, the Hearthstone dataset, and the MBPP dataset. We use tree sitter tool to automatically extract the API terms and sketches of programs.<br><br>Take the MBPP dataset as an example:<br><br>To extract API terms, run 'data_process/api_extract.py' and acquire the 'api_terms.jsonl' <br><br>To extract sketches, run 'data_process/sketch_extract.py' and acquire the 'sketches.jsonl' <br><br>Put the API terms, sketches, complete codes, and requirements into the 'final_train.jsonl' and 'final_test.jsonl'.<br><br>```<br><br><br>### Training<br>Given a specific requirement, the APIer predicts API terms, and the Sketcher outputs corresponding the sketch based on the API terms and requirements. And the Coder fills the sketch to a complete program according to the API terms, sketch and requirement.<br><br>```<br>export CUDA_VISIBLE_DEVICES=0<br><br>python AOCG_finetune.py \<br>--stage_1 nl_pp \<br>--stage_2 nl_pp_ss \<br>--stage_3 nl_ss_pp_code \<br>--local_rank -1 <br><br>```<br><br><br>### Inference<br>The AOCG predicts code snippets in a progressive generation manner, and write the predicted codes into 'xx.output'.<br><br>```<br>export CUDA_VISIBLE_DEVICES=0<br><br>python AOCG_inference.py \<br>--stage_1 nl_pp \<br>--stage_2 nl_pp_ss \<br>--stage_3 nl_ss_pp_code \<br>--local_rank -1 <br><br>```<br><br><br>### Evaluation<br>```<br>After acquiring the generated codes, evaluate the programs by running 'evaluator/evaluation.py'.<br><br>```</pre>