Spaces:
Runtime error
Runtime error
File size: 16,687 Bytes
f239efc |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 |
from functools import partial
import os
import json
from typing import OrderedDict
import tqdm
import torch
from PIL import Image
import ast
import numpy as np
from multiprocessing import Pool
from decord import VideoReader, cpu
import os
from tasks.eval.eval_utils import (
dump_json,
load_json,
EvalDataset,
)
from dataclasses import dataclass
from openai import OpenAI
client = OpenAI(
# This is the default and can be omitted
api_key=os.environ.get("OPENAI_API_KEY"),
)
task_type2chatgpt_contents = OrderedDict({
"MSVD_QA": {
"system": "You are an intelligent chatbot designed for evaluating the correctness of generative outputs for question-answer pairs. "
"Your task is to compare the predicted answer with the correct answer and determine if they match meaningfully. Here's how you can accomplish the task:"
"------"
"##INSTRUCTIONS: "
"- Focus on the meaningful match between the predicted answer and the correct answer.\n"
"- Consider synonyms or paraphrases as valid matches.\n"
"- Evaluate the correctness of the prediction compared to the answer.",
"user": """Please evaluate the following video-based question-answer pair:\n\n"""
"""Question: {question}\n"""
"""Correct Answer: {answer}\n"""
"""Predicted Answer: {pred}\n\n"""
"""Provide your evaluation only as a yes/no and score where the score is an integer value between 0 and 5, with 5 indicating the highest meaningful match. """
"""Please generate the response in the form of a Python dictionary string with keys 'pred' and 'score', where value of 'pred' is a string of 'yes' or 'no' and value of 'score' is in INTEGER, not STRING."""
"""DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. """
"""For example, your response should look like this: {{'pred': 'yes', 'score': 4.8}}."""
},
"MSRVTT_QA": {
"system": "You are an intelligent chatbot designed for evaluating the correctness of generative outputs for question-answer pairs. "
"Your task is to compare the predicted answer with the correct answer and determine if they match meaningfully. Here's how you can accomplish the task:"
"------"
"##INSTRUCTIONS: "
"- Focus on the meaningful match between the predicted answer and the correct answer.\n"
"- Consider synonyms or paraphrases as valid matches.\n"
"- Evaluate the correctness of the prediction compared to the answer.",
"user": """Please evaluate the following video-based question-answer pair:\n\n"""
"""Question: {question}\n"""
"""Correct Answer: {answer}\n"""
"""Predicted Answer: {pred}\n\n"""
"""Provide your evaluation only as a yes/no and score where the score is an integer value between 0 and 5, with 5 indicating the highest meaningful match. """
"""Please generate the response in the form of a Python dictionary string with keys 'pred' and 'score', where value of 'pred' is a string of 'yes' or 'no' and value of 'score' is in INTEGER, not STRING."""
"""DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. """
"""For example, your response should look like this: {{'pred': 'yes', 'score': 4.8}}."""
# """Make sure you only response with text that Follows Python syntax. For example, your response should look like this: {'pred': 'yes', 'score': 4.8}."""
},
"ActivityNet": {
"system": "You are an intelligent chatbot designed for evaluating the correctness of generative outputs for question-answer pairs. "
"Your task is to compare the predicted answer with the correct answer and determine if they match meaningfully. Here's how you can accomplish the task:"
"------"
"##INSTRUCTIONS: "
"- Focus on the meaningful match between the predicted answer and the correct answer.\n"
"- Consider synonyms or paraphrases as valid matches.\n"
"- Evaluate the correctness of the prediction compared to the answer.",
"user": """Please evaluate the following video-based question-answer pair:\n\n"""
"""Question: {question}\n"""
"""Correct Answer: {answer}\n"""
"""Predicted Answer: {pred}\n\n"""
"""Provide your evaluation only as a yes/no and score where the score is an integer value between 0 and 5, with 5 indicating the highest meaningful match. """
"""Please generate the response in the form of a Python dictionary string with keys 'pred' and 'score', where value of 'pred' is a string of 'yes' or 'no' and value of 'score' is in INTEGER, not STRING."""
"""DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. """
"""For example, your response should look like this: {{'pred': 'yes', 'score': 4.8}}."""
# """Make sure you only response with text that Follows Python syntax. For example, your response should look like this: {'pred': 'yes', 'score': 4.8}."""
},
"TGIF_QA": {
"system": "You are an intelligent chatbot designed for evaluating the correctness of generative outputs for question-answer pairs. "
"Your task is to compare the predicted answer with the correct answer and determine if they match meaningfully. Here's how you can accomplish the task:"
"------"
"##INSTRUCTIONS: "
"- Focus on the meaningful match between the predicted answer and the correct answer.\n"
"- Consider synonyms or paraphrases as valid matches.\n"
"- Evaluate the correctness of the prediction compared to the answer.",
"user": """Please evaluate the following video-based question-answer pair:\n\n"""
"""Question: {question}\n"""
"""Correct Answer: {answer}\n"""
"""Predicted Answer: {pred}\n\n"""
"""Provide your evaluation only as a yes/no and score where the score is an integer value between 0 and 5, with 5 indicating the highest meaningful match. """
"""Please generate the response in the form of a Python dictionary string with keys 'pred' and 'score', where value of 'pred' is a string of 'yes' or 'no' and value of 'score' is in INTEGER, not STRING."""
"""DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. """
"""For example, your response should look like this: {{'pred': 'yes', 'score': 4.8}}."""
# """Make sure you only response with text that Follows Python syntax. For example, your response should look like this: {'pred': 'yes', 'score': 4.8}."""
},
})
# Follow the instructions carefully and be helpful and precise with your answer.
def check_ans_qa(question, pred, gt, task_type, model="gpt-3.5-turbo-0125"):
try:
# Compute the temporal understanding score
user_input = task_type2chatgpt_contents[task_type]['user']
user_input = user_input.format(question=question, answer=gt, pred=pred)
completion = client.chat.completions.create(
model=model,
messages=[
{
"role": "system",
"content": task_type2chatgpt_contents[task_type]['system'],
},
{
"role": "user",
"content": user_input,
}
]
)
# Convert response to a Python dictionary.
# response_message = completion["choices"][0]["message"]["content"]
response_message = completion.choices[0].message.content
response_dict = ast.literal_eval(response_message)
pred = response_dict['pred']
score = response_dict['score']
if not pred in ('yes', 'no') or not isinstance(score, (int, float)):
raise ValueError(f"{model} doesn't follow")
flag = pred == 'yes'
except Exception as e:
import traceback
traceback.print_exc()
flag, score = False, 0
print(
f"GPT cannot deal with:\n"
f"--pred: {pred}\n"
f"--gt: {gt}\n"
f"--gpt responded: {response_message}\n"
"--will assign flag=False and score=0"
)
print(f"Dumb Answer in {task_type}")
return flag, score
def chatgpt_eval(res, model="gpt-3.5-turbo-0125"):
pred = res['pred']
gt = res['gt']
question=res['question']
task_type = res['task_type']
correct, score = check_ans_qa(question=question, pred=pred, gt=gt,task_type=task_type, model=model) # acc is bool, score is given by chatgpt
# update the scores in result_list for this sample
res['score'] = score
res['correct'] = correct
return res
def save_results(result_list, save_path, model="gpt-3.5-turbo-0125"):
dump_json(result_list, save_path, 'inference_results.json')
with Pool(7) as pool:
func = partial(chatgpt_eval, model=model)
result_list = [ res for res in tqdm.tqdm(pool.imap_unordered(func, result_list), total=len(result_list), desc='Language Chat Model Automated Evaluation...')]
# result_list = pool.map(partial(chatgpt_eval, model=model), result_list)
# result_list = [chatgpt_eval(res, model=model) for res in result_list]
final_res, acc_dict = {}, {}
correct, total, total_score = 0, 0, 0
for i, res in enumerate(result_list):
task_type = res['task_type']
if task_type not in acc_dict:
acc_dict[task_type] = {
'correct': 0,
'total': 0,
'score': 0,
} # correct, total
acc_dict[task_type]['total'] += 1
acc_dict[task_type]['correct'] += res['correct']
acc_dict[task_type]['score'] += res['score']
for k, v in acc_dict.items():
final_res[k] = {
'acc': v['correct'] / v['total'] * 100,
'score': v['score'] / v['total']
}
correct += v['correct']
total += v['total']
total_score += v['score']
final_res['Avg_Acc'] = correct / total * 100
final_res['Avg_Score'] = total_score / total
all_results = {
"acc_dict": acc_dict,
"result_list": result_list
}
dump_json(all_results, save_path, 'all_results.json')
dump_json(final_res, save_path, 'upload_leaderboard.json')
def load_results(save_path):
json_data = load_json(save_path, 'inference_results.json')
return json_data
@dataclass
class OpenendQASample():
question: str
answer: str
class VideoQABenchDataset(EvalDataset):
data_dir = "DATAS/VideoQA"
data_list_info = OrderedDict({
"MSVD_QA": OrderedDict(
q_json_relpath="MSVD_Zero_Shot_QA/test_q.json",
a_json_relpath="MSVD_Zero_Shot_QA/test_a.json",
prefix="DATAS/VideoQA/MSVD_Zero_Shot_QA/videos",
data_type="video",
bound=False,
question_key='question',
answer_key='answer',
name_key='video_name',
postfix=('avi',),
),
"MSRVTT_QA": OrderedDict(
q_json_relpath="MSRVTT_Zero_Shot_QA/test_q.json",
a_json_relpath="MSRVTT_Zero_Shot_QA/test_a.json",
prefix="DATAS/VideoQA/MSRVTT_Zero_Shot_QA/videos/all",
data_type="video",
bound=False,
question_key='question',
answer_key='answer',
name_key='video_name',
postfix=('mp4', ),
), # don't has start & end
"ActivityNet": OrderedDict(
q_json_relpath="ActivityNet/test_q.json",
a_json_relpath="ActivityNet/test_a.json",
prefix="DATAS/VideoQA/ActivityNet/all_test",
data_type="video",
bound=False,
question_key='question',
answer_key='answer',
name_key='video_name',
postfix=('mp4', 'mkv', 'webm'),
), # don't has start & end
"TGIF_QA": OrderedDict(
q_json_relpath="TGIF_QA/test_q.json",
a_json_relpath="TGIF_QA/test_a.json",
prefix="DATAS/VideoQA/TGIF_QA/tgif_videos",
data_type="gif",
bound=False,
question_key='question',
answer_key='answer',
name_key='video_name',
postfix=('gif',),
), # don't has start & end
})
def __init__(self, *args, **kwargs):
# test_ratio for videoqa is for each sub dataset
test_ratio = kwargs.pop('test_ratio', None)
kwargs['test_ratio'] = None
test_datasets = kwargs.pop('test_datasets', None)
super().__init__(*args, **kwargs)
test_ratio = 1 if test_ratio is None else test_ratio
self.test_ratio = test_ratio
if test_datasets is not None:
data_list_info = {k:v for k,v in self.data_list_info.items() if k in test_datasets}
else:
data_list_info = self.data_list_info
data_dir = self.data_dir
self.data_list = []
for k, v in data_list_info.items():
with open(os.path.join(data_dir, v['q_json_relpath']), 'r') as f:
quesions_json_data = json.load(f)
with open(os.path.join(data_dir, v['a_json_relpath']), 'r') as f:
answers_json_data = json.load(f)
indexs = list(range(len(quesions_json_data)))
np.random.RandomState(42).shuffle(indexs)
num_samples = int(len(indexs) * self.test_ratio) if 0 < self.test_ratio <= 1 else int(self.test_ratio)
indexs = indexs[:num_samples]
for i in indexs:
question_data = quesions_json_data[i]
answer_data = answers_json_data[i]
data = {}
# why do we have anet's video name not in the original json file???
if k == "ActivityNet":
question_data['video_name'] = 'v_' + question_data['video_name']
data.update(**question_data)
data.update(**answer_data)
self.data_list.append({
'task_type': k,
'data': data,
**v, # all the infos
})
print(len(self.data_list))
def __len__(self):
return len(self.data_list)
def __getitem__(self, idx):
decord_method = self.decord_method[self.data_list[idx]['data_type']]
bound = None
if self.data_list[idx]['bound']:
bound = (
self.data_list[idx]['data']['start'],
self.data_list[idx]['data']['end'],
)
video_name_key = self.data_list[idx]['name_key']
video_name = self.data_list[idx]['data'][video_name_key]
video_postfixs = self.data_list[idx]['postfix']
video_paths = []
for p in video_postfixs:
video_path = os.path.join(self.data_list[idx]['prefix'], video_name + '.' + p)
if os.path.exists(video_path):
video_paths.append(video_path)
assert len(video_paths) > 0, f'no video named {video_name}'
# video_filename = self.data_list[idx]['data'][video_name_key] + video_postfix
video_path = video_paths[0]
images_group = decord_method(video_path, bound)
question_key = self.data_list[idx]['question_key']
answer_key = self.data_list[idx]['answer_key']
sample = OpenendQASample(
question=self.data_list[idx]['data'][question_key],
answer=self.data_list[idx]['data'][answer_key]
)
question, answer = self.qa_template(sample)
return {
'video_pils': images_group, # some might use the original pils and do their own transforms
'question': question,
'video_path': video_path,
'answer': answer,
'task_type': self.data_list[idx]['task_type']
}
def qa_template(self, data: OpenendQASample):
answer = data.answer
question = data.question
# by far, might use some prompting.
return question, answer
|