File size: 9,389 Bytes
2a26d3b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 |
import datetime
import json
import multiprocessing
import os
import re
import string
import subprocess
import time
import numpy as np
import torch
import torch.distributed as dist
from attrdict import AttrDict
from human_eval.evaluation import evaluate_functional_correctness
from tqdm import tqdm
from transformers import (AutoModelForCausalLM, AutoTokenizer,
StoppingCriteria, StoppingCriteriaList)
from utils.dataset import MBPPDataset
from utils.utils import cleanup_code
class KeywordsStoppingCriteria(StoppingCriteria):
def __init__(self, keywords_str, tokenizer):
StoppingCriteria.__init__(self)
self.current_context = []
self.tokenizer = tokenizer
self.keywords_str = keywords_str
def __call__(
self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
) -> bool:
self.current_context.append(input_ids[0][-1].item())
current_context = self.tokenizer.decode(self.current_context)
for word in self.keywords_str:
if word in current_context:
return True
return False
class MBPP:
"""
MBPP evaluation class.
"""
def __init__(
self,
data_root,
max_seq_len=2048,
language="python",
max_gen_len=200,
batch_size=512,
log_dir=None,
temperature=0,
issft=False,
top_p=0.95,
model_name="",
inference_increment=True,
tokenizer_cfg=None,
n_sample=40,
k_sample=1,
):
self.data_root = data_root
self.max_seq_len = max_seq_len
self.max_gen_len = max_gen_len
self.batch_size = batch_size
self.k = k_sample
self.n_sample = n_sample
self.language = language
self.log_dir = log_dir
self.sft = issft
self.temperature = temperature
self.top_p = top_p
self.model_name = tokenizer_cfg["model_path"].replace("/", "_")
self.inference_increment = inference_increment
os.makedirs(self.log_dir, exist_ok=True)
tokenizer_cls = tokenizer_cfg.pop("cls")
try:
self.tokenizer = AutoTokenizer.from_pretrained(
tokenizer_cfg.pop("model_path"), trust_remote_code=True
)
except Exception as e:
print(e)
assert False
@torch.no_grad()
def eval_model(self, gpt, accelerator):
"""
Evaluate the model.
"""
assert (
self.log_dir is not None
), "log_dir should not be None when evaluating MBPP"
dataset = MBPPDataset(self.data_root, samplenum=self.n_sample)
nprompt = len(dataset) // self.n_sample
dp_rank = accelerator.process_index
dp_size = accelerator.num_processes
if self.k > 1:
assert self.n_sample >= 80, "MBPP PASS@80 needs n_sample >= 80"
gpt.eval()
prompt_indices_split = np.array_split(range(nprompt), dp_size)
prompt_indices = prompt_indices_split[dp_rank]
indices = []
for x in prompt_indices:
for j in range(self.n_sample):
indices.append(x * self.n_sample + j)
all_num = len(indices)
processed_num = 0
log_file = os.path.join(
self.log_dir,
f"{self.model_name}_rank{dp_rank}_bs{self.batch_size}_shot_log_{self.language}.json",
)
tmpfile = open(log_file, "w")
totoalnum = 0
start_time = time.time()
for idx in tqdm(range(0, len(indices), self.batch_size)):
prompt_list = []
prompt_lens = []
answers_list = []
test_list = []
taskid = []
tokenized_prompt_lens = []
for j in indices[idx : idx + self.batch_size]:
data = dataset[j]
prompt = dataset.prompt
prompt1 = data["prompt"]
tests = "\n".join(data["test"])
test_list.append(data["test"])
prompt_curr = f"You are an expert Python programmer, and here is your task: {prompt1} Your code should pass these tests:\n\n{tests}\n[BEGIN]"
fprompt = ""
for i in range(len(prompt) - 1, -1, -1):
finalprompt = prompt[i] + prompt_curr
curr_seq_len = len(self.tokenizer.encode(finalprompt))
if curr_seq_len >= self.max_seq_len - self.max_gen_len:
continue
else:
fprompt = finalprompt
break
if fprompt == "":
fprompt = prompt_curr
encodelist = self.tokenizer.encode(fprompt)
while True:
try:
fprompt = self.tokenizer.decode(
encodelist[: self.max_seq_len - self.max_gen_len]
)
break
except:
encodelist.pop(-1)
prompt_list.append(fprompt)
answers_list.append(data["code"])
prompt_lens.append(len(fprompt))
taskid.append(data["task_id"])
tokenized_prompt = self.tokenizer(
prompt_list, padding=True, return_tensors="pt"
)
inputids = tokenized_prompt["input_ids"].to(gpt.device)[
:, -self.max_seq_len :
]
attenion_mask = tokenized_prompt["attention_mask"].to(gpt.device)[
:, -self.max_seq_len :
]
if self.temperature == 0:
stop_criteria = KeywordsStoppingCriteria(["[DONE]"], self.tokenizer)
decoded = gpt.generate(
input_ids=inputids,
attention_mask=attenion_mask,
max_new_tokens=self.max_gen_len,
top_p=self.top_p,
eos_token_id=self.tokenizer.eos_token_id,
do_sample=False,
stopping_criteria=StoppingCriteriaList([stop_criteria]),
pad_token_id=self.tokenizer.eos_token_id,
)
else:
decoded = gpt.generate(
tokenized_prompt_lens,
max_new_tokens=self.max_gen_len,
temperature=self.temperature,
top_p=0.95,
inference_increment=True,
stopping_criteria=StoppingCriteriaList([stop_criteria]),
pad_token_id=self.tokenizer.eos_token_id,
)
for local_idx, text in enumerate(decoded):
prediction = decoded[local_idx]
prediction = self.tokenizer.decode(prediction, skip_special_tokens=True)
# print(prediction)
suffixprediction = prediction[prompt_lens[local_idx] :]
suffixprediction = suffixprediction.split("[DONE]")[0].strip()
res = {"task_id": taskid[local_idx], "generation": suffixprediction}
tmpfile.write(json.dumps(res) + "\n")
tmpfile.flush()
totoalnum += 1
self.log_score(dp_rank, totoalnum, all_num, start_time, self.batch_size)
tmpfile.close()
accelerator.wait_for_everyone()
self._calculate_final_score(accelerator)
def log_score(self, dp_rank, processed_num, all_num, start_time, bs):
"""
Log the score.
"""
mem = torch.cuda.max_memory_allocated() / (1 << 30)
avg_time = (time.time() - start_time) / processed_num * bs
print(
f"DP RANK:{dp_rank} process_num/all_num:{int(processed_num)}/{all_num} "
f"avg_time_per_batch:{avg_time:.2f} s "
f"still_need:{((all_num - processed_num) // bs + 1) * avg_time / 60:.2f} m",
f"mem:{mem:.3f} GiB bs:{bs}",
flush=True,
)
if processed_num == all_num:
print(
f"EVAL DONE! Process time {(time.time() - start_time) / 60:.2f} m",
flush=True,
)
def _calculate_final_score(self, accelerator):
"""
Calculate the final score.
"""
if accelerator.is_local_main_process:
logfilepath = os.path.join(self.log_dir, f"final_{self.model_name}.jsonl")
logfile = open(logfilepath, "w")
for i in range(accelerator.num_processes):
tmplogfile = os.path.join(
self.log_dir,
f"{self.model_name}_rank{i}_bs{self.batch_size}_shot_log_{self.language}.json",
)
logfile.write(open(tmplogfile).read().strip() + "\n")
os.remove(tmplogfile)
logfile.close()
timeout = 10
runlang = self.language
res = evaluate_functional_correctness(
input_file=logfilepath,
problem_file=os.path.join(self.data_root, f"mbpp_test.jsonl"),
tmp_dir=self.log_dir,
timeout=timeout,
language=runlang,
)
print("score is", res["pass@%d" % self.k])
os.remove(logfilepath)
return
|