import itertools
import json
import multiprocessing
import numpy as np
from typing import Dict
from datasets import load_dataset
from .testing_util import run_test
DATASET = "codeparrot/apps"
TIMEOUT = 10
def check_correctness(sample, generation, timeout, debug=True):
"""Check correctness of code generation with a global timeout.
The global timeout is to catch some extreme/rare cases not handled by the timeouts
inside `run_test`"""
def _temp_run(sample, generation, debug, result):
result.append(run_test(sample, test=generation, debug=debug))
manager = multiprocessing.Manager()
result = manager.list()
p = multiprocessing.Process(target=_temp_run, args=(sample, generation, debug, result))
p.start()
p.join(timeout=timeout + 1)
if p.is_alive():
p.kill()
if not result:
in_outs = json.loads(sample["input_output"])
# consider that all tests failed
result = [[-1 for i in range(len(in_outs["inputs"]))]]
if debug:
print(f"global timeout")
return result[0]
def evaluate_generations(generations: list, level: str = "all", debug: bool = False):
"""We take the list of code generations and try to compile them
and the run their corresponding unit tests which are retrieved from the APPS dataset.
Args:
generations: list of code generations (same order as samples in APPS dataset)
level: difficulty level used in the generation, can be "all", "introductory", "interview" or "competition"
Returns:
results: dictionary of results, key is the problem index, value is a list of results for each generation
[-2] = compile error, [-1] = runtime error [False] = failed test case [True] = passed test case
"""
# generations are code generations in the same order of the dataset
apps_eval = load_dataset(DATASET, split="test", difficulties=[level])
results = {}
for index in range(len(generations)):
# code generations for problem (index)
problem_generations = generations[index]
# get corresponding samples from APPS dataset
sample = apps_eval[index]
res = []
# loop over the generations
for o_idx, o in enumerate(problem_generations):
curr_res = [-2]
try:
curr_res = check_correctness(sample, o, timeout=TIMEOUT, debug=debug)
if debug:
print(f"\nSuccessful compilation of task {index}!")
fixed = []
for e in curr_res:
if isinstance(e, np.ndarray):
e = e.item(0)
if isinstance(e, np.bool_):
e = bool(e)
fixed.append(e)
curr_res = fixed
if not np.all(curr_res):
if debug:
print(f"Results were not True for all test cases")
except Exception as e:
if debug:
print(f"Compilation failed, test framework exception = {repr(e)}{e}\n")
break
finally:
assert isinstance(curr_res, list)
res.append(curr_res)
results[index] = res
return results
def estimate_pass_at_k(num_samples, num_correct, k):
"""Estimates pass@k of each problem and returns them in an array."""
def estimator(n: int, c: int, k: int) -> float:
"""Calculates 1 - comb(n - c, k) / comb(n, k)."""
if n - c < k:
return 1.0
return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1))
if isinstance(num_samples, int):
num_samples_it = itertools.repeat(num_samples, len(num_correct))
else:
assert len(num_samples) == len(num_correct)
num_samples_it = iter(num_samples)
return np.array([estimator(int(n), int(c), k) for n, c in zip(num_samples_it, num_correct)])
def get_results(results: Dict[int, list], count_errors: bool = False, k_list: list = [1, 10, 100]):
"""
Given the results evaluated against the testcases we output some statistics.
For single generations:
>>> example_results = {0: [[-2]], 1: [[False,False]], 2: [[True,True]], 3: [[False,True,False,True]], 4: [[-1,-1]]}
>>> get_results(example_results, count_errors=True)
Computing accuracy metrics...
number of compile errors = 1 avg = 0.2
number of runtime errors = 1 avg = 0.2
number of problems evaluated = 5
Average Accuracy : 0.3
Strict Accuracy : 0.2
{'avg_accuracy': 0.3, 'strict_accuracy': 0.2, 'pass_at_k': None}
For multiple generations:
>>> example_results = {0: [[-2], [True, True, True]], 1: [[-1,-1, -1], [True, False, True]]}
>>> get_results(example_results, k_list=[1, 2])
Computing pass@k metric for multiple generations...
{'pass@1': 0.25, 'pass@2': 0.5}
{'avg_accuracy': None, 'strict_accuracy': None, 'pass_at_k': {'pass@1': 0.25, 'pass@2': 0.5}}
"""
metrics = {"avg_accuracy": None, "strict_accuracy": None, "pass_at_k": None}
if len(results[0]) == 1:
# for single generations we compute average accuracy and stric accuracy: original APPS metrics
print("Computing accuracy metrics...")
res = []
per_prob_res = []
all_correct = []
for index in results:
problem_results = np.asarray(results[index])
res.extend(problem_results)
per_prob_res.append(np.mean(problem_results > 0))
all_correct.append(np.all(problem_results > 0))
# we count campilation and runtime errors once per pronlem
compile_errors = len([e for e in res if -2 in e])
runtime_errors = len([e for e in res if -1 in e])
total_testcases = len(res)
if count_errors:
print(f"number of compile errors = {compile_errors} avg = {compile_errors / total_testcases}")
print(f"number of runtime errors = {runtime_errors} avg = {runtime_errors / total_testcases}")
print(f"number of problems evaluated = {total_testcases}")
print(f"Average Accuracy : {np.mean(per_prob_res)}")
print(f"Strict Accuracy : {np.mean(all_correct)}")
metrics["avg_accuracy"] = np.mean(per_prob_res)
metrics["strict_accuracy"] = np.mean(all_correct)
else:
# for multiple generations we use pass@k metric used in the HumanEval benchmark
# we use strict accuracy, a generation is valid if it has to pass all the tests
print("Computing pass@k metric for multiple generations...")
# total is list with nb generations per task (task=index)
# correct is number of generations that passed all tests per task
total = []
correct = []
for index in results:
all_correct = []
for generation in results[index]:
gen = np.array(generation)
all_correct.append(np.all(gen>0))
total.append(len(all_correct))
correct.append(sum(all_correct))
total = np.array(total)
correct = np.array(correct)
ks = k_list
pass_at_k = {f"pass@{k}": estimate_pass_at_k(total, correct, k).mean() for k in ks if (total >= k).all()}
print(pass_at_k)
metrics["pass_at_k"] = pass_at_k
return metrics
def compute_metrics(generations, level="all", k_list=[1, 10, 100], count_errors=True, debug=False):
"""Return metrics for the given generations.
Args:
generations: list of code generations for each problem (each generation is a list of generations)
k_list: list of k values to compute pass@k when using multiple generations
count_errors: whether to count compilation and runtime errors when using single generations
level: difficulty level in APPS dataset that was used for the given generations (from: "all", "introductory", "interview", "competition")
Returns:
metrics: dict of metrics
Examples:
>>> import json
>>> # lists of solutions to the two first APPS problems (note not all solutions pass all tests)
>>> solution_sample1 = json.load(open("test_examples/solutions_problem_1.json", "r"))
>>> solution_sample2 = json.load(open("test_examples/solutions_problem_2.json", "r"))
>>> single_solutions = [solution_sample1[:1], solution_sample2[:1]]
>>> compute_metrics(single_solutions, level="all")
Computing accuracy metrics...
number of compile errors = 0 avg = 0.0
number of runtime errors = 0 avg = 0.0
number of problems evaluated = 2
Average Accuracy : 1.0
Strict Accuracy : 1.0
{'avg_accuracy': 1.0, 'strict_accuracy': 1.0, 'pass_at_k': None}
>>> multiple_solutions = [solution_sample1[:3], solution_sample2[:3]]
>>> compute_metrics(multiple_solutions, level="all", k_list=[1, 2, 3])
Computing pass@k metric for multiple generations...
{'pass@1': 1.0, 'pass@2': 1.0, 'pass@3': 1.0}
{'avg_accuracy': None, 'strict_accuracy': None, 'pass_at_k': {'pass@1': 1.0, 'pass@2': 1.0, 'pass@3': 1.0}}
"""
results = evaluate_generations(generations, level=level, debug=debug)
metrics = get_results(results, count_errors=count_errors, k_list=k_list)
return metrics
# import doctest
# doctest.testmod()