apps_metric / example_script.py
loubnabnl's picture
loubnabnl HF staff
use evaluate metric for evaluation
73ef22f
raw
history blame
6.3 kB
"""This is an example script to evaluate a code generation model on APPS, you can also use the APPS solutions as code generations
> python example_script.py --model_ckpt MODEL_NAME --num_tasks 10 --difficulty introductory --n_samples 1
> python example_script.py --use_solutions True --num_tasks 10 --difficulty introductory --n_samples 1"""
import json
import pprint
from tqdm import tqdm
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
from evaluate import load
def generate_prompt(sample):
starter_code = None if len(sample["starter_code"]) == 0 else sample["starter_code"]
try:
input_outpout = json.loads(sample["input_output"])
fn_name = None if not input_outpout.get("fn_name") else input_outpout["fn_name"]
except ValueError:
fn_name = None
_input = "\nQUESTION:\n"
_input += sample["question"]
if starter_code:
_input += starter_code
if fn_name:
_input += "\nUse Standard Input format"
else:
_input += "\nUse Call-Based format"
_input += "\nANSWER:\n"
return _input
def complete_code(pipe, prompt, num_completions=1, max_length=256, **gen_kwargs):
"""Complete prompt with text generation pipeline and return num_completions."""
prompt = pipe.tokenizer.eos_token + prompt
try:
code_gens = pipe(prompt, num_return_sequences=num_completions, max_length=max_length, **gen_kwargs)
return [code_gen["generated_text"][len(prompt):] for code_gen in code_gens]
except IndexError:
print("prompt is longer than the context size of the model, generation skipped")
code_gens = ""
return [""]
def make_generations(dataset, args, model, tokenizer):
set_seed(args.seed)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=args.device_int)
# Generation settings
gen_kwargs = {
"do_sample": args.do_sample,
"temperature": args.temperature,
"top_p": args.top_p,
"top_k": args.top_k
}
# Generate completions for evaluation set
n_tasks = args.num_tasks if args.num_tasks is not None else len(dataset)
print(f"ntasks is {n_tasks}")
generations = []
for task in tqdm(range(n_tasks)):
task_generations = []
prompt = generate_prompt(dataset[task]).strip()
task_generations.extend(complete_code(pipe, prompt, num_completions=args.n_samples, max_length=args.max_length, **gen_kwargs))
generations.append([gen.replace(args.eos, "") for gen in task_generations])
return generations
def main(args):
DATA_PATH = "codeparrot/apps"
argsdict = vars(args)
print(pprint.pformat(argsdict))
# setup
print("Loading evaluation dataset...")
dataset = load_dataset(DATA_PATH, split="test", difficulties=[args.difficulty])
if args.use_solutions:
print("Using data solutions as code generations")
model = None
tokenizer = None
generations = []
for index in range(args.num_tasks+1):
try:
sol = json.loads(dataset[index]["solutions"])
generations.append(sol[:args.n_solutions])
except ValueError:
print(f"No solutions for task {index} or not enough to have {args.n_solutions} solutions")
break
else:
print("Loading tokenizer and model...")
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer)
model = AutoModelForCausalLM.from_pretrained(args.model_ckpt)
generations = make_generations(dataset, args, model, tokenizer)
metric = load("loubnabnl/apps_metric")
results = metric.compute(predictions=generations, level=args.difficulty, k_list=args.k_list, count_errors=args.count_errors, debug=args.debug)
print(results)
with open(args.output_file, "w") as fp:
json.dump(results, fp)
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Testing a Language Model on APPS Python Code dataset")
#model and tokenizer arguments
parser.add_argument("--model_ckpt", default="loubnabnl/apps-1.5B-model", type=str, help="path to model checkpoint.")
parser.add_argument("--tokenizer", default="gpt2", type=str, help="tokenizer to use.")
parser.add_argument("--eos", default="<|endoftext|>", type=str, help="end of sentence token.")
# generation arguments
parser.add_argument("--do_sample", default=True, type=bool, help="do sampling in generation")
parser.add_argument("--temperature", default=0.2, type=float, help="temperature for sampling")
parser.add_argument("--top_p", default=0.95, type=float, help="top p for sampling")
parser.add_argument("--top_k", default=0, type=float, help="top k for sampling")
parser.add_argument("--max_length", default=1024, type=int, help="max length of generated code")
# evaluation arguments
parser.add_argument("--difficulty", default="all", type=str, help="difficulty level to select in the dataset from:\
'all', 'introductory', 'interview' and 'competition' ")
parser.add_argument("--num_tasks", default=6, type=int, help="number of tasks to evaluate")
parser.add_argument("--use_solutions", default=False, type=bool, help="use solutions instead of generating new code")
parser.add_argument("--n_samples", default=1, type=int, help="number of samples to generate")
parser.add_argument("--n_solutions", default=1, type=int, help="number of solutions to use")
parser.add_argument("--k_list", default=[1, 2, 3], type=list, help="list of k values to evaluate pass@k")
parser.add_argument("--count_errors", default=False, type=bool, help="count compilation and runtime errors for single generations")
# configuration
parser.add_argument("--seed", default=0, type=int, help="generation seed")
parser.add_argument("--device_int", default=-1, type=int, help="device on which code generation is run, if positive use GPU")
parser.add_argument("--debug", default=False, type=bool, help="debug mode")
# save
parser.add_argument("--output_file", default="apps_metrics.json", type=str, help="output file to save the results")
args = parser.parse_args()
main(args)