import os import csv import time import argparse import jsonlines import pandas as pd from statistics import mean from jinja2 import Environment, FileSystemLoader from utils import pass_at_k_continuous_vals, diff_bleu, post_process_generations, statistical_significance_test, remove_comments, remove_blank_lines,get_files_with_syntax_errors parser = argparse.ArgumentParser() parser.add_argument("--data_subset", type = str, default = "latency", help = "latency/resource_util/runtime_efficiency/maintenance/security") parser.add_argument("--model", type = str, default = "wizardcoder", help = "model name") parser.add_argument("--model_path", type = str, required = True, help = "HF path for OS models") parser.add_argument("--prompt", type = str, default = "base_prompt", help = "base_prompt/coding_concepts/chain_of_thought/one-shot") parser.add_argument("--num_samples", type = int, default = 1, help = "Number of samples") parser.add_argument("--score_k", type = str, default = "1,5,10,20", help = "K value for score@k (should not be greater than num_samples and can be comma-separated)") parser.add_argument("--metric", type = str, default = "runtime", help = "runtime/diffbleu/codeql-diffbleu") args = parser.parse_args() args.model = args.model_path.split("/")[-1] generations_path = os.path.join("generations", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples", "generated_outputs.jsonl") if(args.metric == "classification"): generations_path = os.path.join("generations", "classification", args.data_subset, os.path.split(args.model_path)[1], args.prompt, f"{args.num_samples}_samples", "generated_outputs.jsonl") left_predictions = [] right_predictions = [] left_labels = [] right_labels = [] with jsonlines.open(generations_path) as reader: for generation in reader: left_predictions.append(generation['left_output']) left_labels.append(generation['classification_left_label']) right_predictions.append(generation['right_output']) right_labels.append(generation['classification_right_label']) left_accuracy = sum([1 if prediction == label else 0 for prediction, label in zip(left_predictions, left_labels)]) / len(left_labels) left_consistency = sum([1 if prediction is not None else 0 for prediction in left_predictions]) / len(left_predictions) right_accuracy = sum([1 if prediction == label else 0 for prediction, label in zip(right_predictions, right_labels)]) / len(right_labels) right_consistency = sum([1 if prediction is not None else 0 for prediction in right_predictions]) / len(right_predictions) joint_accuracy = [1 if left_prediction == left_label and right_prediction == right_label else 0 for left_prediction, left_label, right_prediction, right_label in zip(left_predictions, left_labels, right_predictions, right_labels)] joint_accuracy = sum(joint_accuracy) / len(joint_accuracy) result_string = {"Model":args.model, "left_accuracy":round((left_accuracy*100),1), "right_accuracy":round((right_accuracy*100),1), "joint_accuracy":round((joint_accuracy*100),1), "left_consistency":round((left_consistency*100),1), "right_consistency":round((right_consistency*100),1)} output_path = os.path.join("evaluation_results","classification",args.data_subset,os.path.split(args.model_path)[1],args.prompt,f"{args.num_samples}_samples") if not os.path.exists(output_path): os.makedirs(output_path) samples = pd.DataFrame([result_string]) samples.to_json(os.path.join(output_path,"results.jsonl"), orient="records", lines=True) print("{}".format(result_string)) # To calculate runtimes(Applicable for non-func runtime_efficiency) elif args.metric == "runtime": start_time = time.time() with jsonlines.open(generations_path) as reader: samples = [] for generation in reader: parsed_generations = [] for l in range(args.num_samples): generated_answers = post_process_generations(generated_answers=generation['generated_answers'][l], model = args.model, prompt = args.prompt, pl = generation['pl'])[1] parsed_generations.append(generated_answers) samples.append(dict(problem_id = generation['problem_id'], submission_id_v0 = generation['submission_id_v0'], cpu_time_v0 = generation['cpu_time_v0'], cpu_time_v1 = generation['cpu_time_v1'], input=generation['source_code'], target=generation['target_code'], generated_answers=parsed_generations, inference_time=generation['inference_time'])) samples = pd.DataFrame(samples) path = os.path.join("src", "evaluation", "pie-perf", "generated_outputs.jsonl") samples.to_json(path, orient="records", lines = True) env = Environment(loader = FileSystemLoader(os.path.join("src", "evaluation", "pie-perf", "data", "sample"))) template = env.get_template('sample_eval_config_template.yaml') output_path = os.path.join("evaluation_results", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples", "generated_outputs.report") rendered_yaml = template.render(output_path = output_path) config_file_path = os.path.join("src", "evaluation", "pie-perf", "data", "sample", "sample_eval_config.yaml") f=open(config_file_path, "w") f.write(rendered_yaml) f.close() path = os.path.split(output_path)[0] if not os.path.exists(path): os.makedirs(path) run_file = os.path.join("src", "evaluation", "pie-perf", "src", "codenet_eval", "run_eval.py") os.system(f'python3 {run_file} --eval_config {config_file_path}') k_values = list(map(int,args.score_k.split(","))) scores = statistical_significance_test(output_path,args.num_samples,k_values) results = {"model":args.model, "prompt":args.prompt, "num_samples":args.num_samples} for i,j in zip(range(2,len(results),3),range(len(k_values))): results[f"Average Speedups@{k_values[j]},{args.num_samples}"] = scores[i-2] results[f"Correct@{k_values[j]},{args.num_samples}"] = scores[i-1] results[f"Improvements@{k_values[j]},{args.num_samples}"] = scores[i] samples = pd.DataFrame([results]) samples.to_json(os.path.join(path, "results.jsonl"), orient="records", lines=True) print("{}".format(results)) # To calculate diffbleu(Applicable for all splits) elif args.metric == "diffbleu": k_values = list(map(int,args.score_k.split(","))) overall_score={} for k in k_values: overall_score[k] = [] passed = 0 count = 0 with jsonlines.open(generations_path) as reader: for generation in reader: count += 1 scores = [] for l in range(args.num_samples): generated_answers = post_process_generations(generated_answers = generation['generated_answers'][l], model = args.model, prompt = args.prompt, pl = generation['pl']) passed += generated_answers[0] diff_score_bleu = diff_bleu(source_code = generation['source_code'], target = generation['target_code'], generated_answers = generated_answers[1], pl = generation['pl']) scores.append(diff_score_bleu) scores.sort(reverse = True) for k in k_values: overall_score[k].append(pass_at_k_continuous_vals(n = args.num_samples, k = k, vals = scores)) scores = [] scores.append((passed*100)/(count*args.num_samples)) results = {"model":args.model, "prompt":args.prompt, "num_samples":args.num_samples} for k in k_values: results[f"Score@{k},{args.num_samples}"] = round(mean(overall_score[k])*100,1) scores.append(round(mean(overall_score[k])*100,1)) results["Passed"] = (passed*100)/(count*args.num_samples) samples = pd.DataFrame([results]) path = os.path.join("evaluation_results", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples") if not os.path.exists(path): os.makedirs(path) samples.to_json(os.path.join(path,"results_{}.jsonl".format(args.metric)), orient="records", lines=True) print("Pass Rate: {}, DiffBleu Score: {}".format(scores[0], scores[1])) # To run codeql(Applicable for security and maintenance) elif args.metric == "codeql": all_check_paths={} query_lang = {} with jsonlines.open(generations_path) as reader: for generation in reader: query = generation['codeql_check'].split("/")[-1].split(".ql")[0] all_check_paths[query]=generation['codeql_check'] code_path="evaluation_results/edit/{}/{}/{}/{}_samples/generated_code/{}/".format(args.data_subset, args.model, args.prompt, args.num_samples, query) if not os.path.exists(code_path): os.makedirs(code_path) if(generation['pl'] == "python"): ext = ".py" pl = "Python" query_lang[query] = "python" else: ext = ".c" pl = "C" query_lang[query] = "cpp" for index in range(len(generation['generated_answers'])): code_path_indexed = code_path + "{}_{}{}".format(generation['file_path'].split("/")[-2]+"_"+generation['file_path'].split("/")[-1].split(ext)[0], index, ext) f = open(code_path_indexed,"w+") generated_answers = post_process_generations(generated_answers=generation['generated_answers'][index], model=args.model, prompt=args.prompt, pl=generation['pl'])[1] code = remove_comments(generated_answers, generation['pl']) if remove_blank_lines(code).strip() == "": generated_answers = generation['source_code'] f.write(generated_answers) f.close() if(pl == "C"): f = open(code_path+"Makefile", "w+") f.write("SRCS=$(wildcard *.c)\nOBJS=$(SRCS:.c=.o)\n\nall: $(OBJS)\n\n%.o: %.c\n gcc -g -O -c $< -o $@ || (echo \"Deleting $<\" && echo \"$<\" >> rejected_files.log && mv $< $<.reject)\n\nclean:\n\trm -rf *.o") f.close() for query in all_check_paths.keys(): code_path_generations = "evaluation_results/edit/{}/{}/{}/{}_samples/generated_code/".format(args.data_subset, args.model, args.prompt, args.num_samples) code_path_db = "evaluation_results/edit/{}/{}/{}/{}_samples/generated_code_db/".format(args.data_subset, args.model, args.prompt, args.num_samples) if not os.path.exists(code_path_db): os.makedirs(code_path_db) code_path_results="evaluation_results/edit/{}/{}/{}/{}_samples/generated_code_results/".format(args.data_subset, args.model, args.prompt, args.num_samples) if not os.path.exists(code_path_results): os.makedirs(code_path_results) os.system("codeql-home/codeql/codeql database create --quiet --language={} --source-root={}{} {}{}".format(query_lang[query], code_path_generations, query, code_path_db, query)) os.system("codeql-home/codeql/codeql database analyze --rerun {}{} {} --format=csv --output={}{}.csv --threads=0".format(code_path_db, query, all_check_paths[query], code_path_results, query)) k_values = list(map(int,args.score_k.split(","))) overall_score={} for k in k_values: overall_score[k] = [] syntax_errors = {} syn_errors = [] done = [] scores_dump = [] with jsonlines.open(generations_path) as reader: parsed = 0 for generation in reader: query = generation['codeql_check'].split("/")[-1].split(".ql")[0] code_path ="evaluation_results/edit/{}/{}/{}/{}_samples/generated_code/{}/".format(args.data_subset, args.model, args.prompt, args.num_samples, query) scores = [] code_path_results = "evaluation_results/edit/{}/{}/{}/{}_samples/generated_code_results/{}.csv".format(args.data_subset, args.model, args.prompt, args.num_samples, query) code_path_generations = "evaluation_results/edit/{}/{}/{}/{}_samples/generated_code/{}/".format(args.data_subset, args.model, args.prompt, args.num_samples, query) code_path_db = "evaluation_results/edit/{}/{}/{}/{}_samples/generated_code_db/".format(args.data_subset, args.model, args.prompt, args.num_samples) errors=[] with open(code_path_results) as f: csvfile = csv.reader(f) for error in csvfile: errors.append(error[-5].split("/")[1]) errors = list(set(errors)) index = 0 scores = [] ans = [] syn=get_files_with_syntax_errors(generated_code_path = code_path_generations, codeql_db_path = code_path_db, query = query) if(len(syn)>0 and query not in done): syn_errors += syn try: syntax_errors[query] += syn except: syntax_errors[query] = syn done.append(query) for index in range(len(generation['generated_answers'])): if(generation['pl'] == "python"): ext = ".py" pl = "Python" else: ext = ".c" pl = "C" filename = "{}_{}{}".format(generation['file_path'].split("/")[-2]+"_"+generation['file_path'].split("/")[-1].split(ext)[0], index, ext) index += 1 if(filename in errors or filename in syn_errors): scores.append(0) else: scores.append(1) scores.sort(reverse = True) for k in k_values: overall_score[k].append(pass_at_k_continuous_vals(n = args.num_samples, k = k, vals = scores)) scores_dump.append(scores) scores=[] path = os.path.join("evaluation_results", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples") f = open(os.path.join(path,"results.txt"), 'w') f.write(str(scores_dump)) f.close() results = {"model":args.model, "prompt":args.prompt, "num_samples":args.num_samples} for k in k_values: results[f"Score@{k},{args.num_samples}"] = round(mean(overall_score[k])*100,1) results["syntax_errors"] = syntax_errors results["no_of_syntax"] = len(syn_errors) samples = pd.DataFrame([results]) path = os.path.join("evaluation_results", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples") samples.to_json(os.path.join(path, "results_{}.jsonl".format(args.metric)), orient="records", lines=True) print("{}".format(results)) # get codeql*diffbleu numbers(Applicable for security and maintenance) elif args.metric == "codeql-diffbleu": k_values = list(map(int, args.score_k.split(","))) overall_score = {} for k in k_values: overall_score[k] = [] generations_path = os.path.join("generations","edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples", "generated_outputs.jsonl") passed = 0 count = 0 with jsonlines.open(generations_path) as reader: res_path = os.path.split(generations_path)[0].split('/') res_path.insert(1,"evaluation_results") res_path = os.path.join("/".join(res_path[1:]), "results.txt") codeql_results = eval(open(res_path).read()) for generation,res in zip(reader, codeql_results): scores=[] count += 1 for l in range(len(generation['generated_answers'])): generated_answers=post_process_generations(generated_answers = generation['generated_answers'][l], model = args.model, prompt = args.prompt, pl = generation['pl']) passed += generated_answers[0] diff_score_bleu=res[l]*diff_bleu(source_code = generation['source_code'], target = generation['target_code'], generated_answers = generated_answers[1], pl = generation['pl']) scores.append(diff_score_bleu) scores.sort(reverse = True) for k in k_values: overall_score[k].append(pass_at_k_continuous_vals(n = args.num_samples, k = k, vals = scores)) scores = [] scores.append((passed*100)/(count*args.num_samples)) results = {"model":args.model, "prompt":args.prompt, "num_samples":args.num_samples} for k in k_values: results[f"Score@{k},{args.num_samples}"] = round(mean(overall_score[k])*100,1) scores.append(round(mean(overall_score[k])*100,1)) results["Passed"] = (passed*100)/(count*args.num_samples) scores.append((passed*100)/(count*args.num_samples)) samples = pd.DataFrame([results]) path = os.path.join("evaluation_results", "edit", args.data_subset, args.model, args.prompt, f"{args.num_samples}_samples") samples.to_json(os.path.join(path,"results_{}.jsonl".format(args.metric)), orient="records", lines=True) print("{}".format(results))