File size: 4,239 Bytes
63fba53 | 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 | import torch
import os
import json
import numpy as np
from tqdm import tqdm
from datasets import load_dataset, concatenate_datasets
from mathruler.grader import extract_boxed_content
from .data_utils import load_yaml, construct_prompt, save_json, process_single_sample, CAT_SHORT2LONG,DOMAIN_CAT2SUB_CAT
from .eval_utils import evaluate,parse_multi_choice_response, parse_open_response
from ..utils import extract
def run_model(samples, model,):
out_samples = {}
out_response = {}
with torch.no_grad():
messages_list = []
current_messages = []
current_samples = []
for sample in tqdm(samples):
messages = {"prompt":sample["final_input_prompt"],"image":sample["image"]}
current_messages.append(messages)
current_samples.append(sample)
if len(current_messages) >= 2000:
messages_list.append([current_messages,current_samples])
current_messages = []
current_samples = []
if current_messages:
messages_list.append([current_messages,current_samples])
for current_messages,current_samples in tqdm(messages_list):
outputs = model.generate_outputs(current_messages)
for sample,response in zip(current_samples,outputs):
if "<answer>" in response:
response = extract(response,"answer")
if extract_boxed_content(response) != "None":
response = extract_boxed_content(response)
if sample["question_type"] == "multiple-choice":
pred_ans = parse_multi_choice_response(response,sample["all_choices"],sample["index2ans"])
else:
pred_ans = parse_open_response(response)
out_samples[sample['id']] = pred_ans
out_response[sample['id']] = response
return out_samples,out_response
def eval_MMMU_test(model,dataset_path,output_path,subset):
sub_dataset_list = []
for subject in tqdm(DOMAIN_CAT2SUB_CAT[subset]):
sub_dataset = load_dataset(dataset_path, subject, split="test")
sub_dataset_list.append(sub_dataset)
with open("utils/MMMU/answer.json", "r") as f:
gold_answers = json.load(f)
chunk_idx = int(os.environ.get("chunk_idx",0))
num_chunks = int(os.environ.get("num_chunks",1))
samples = []
dataset = concatenate_datasets(sub_dataset_list)
for idx,sample in tqdm(enumerate(dataset)):
if idx % num_chunks == chunk_idx:
sample = process_single_sample(sample)
sample = construct_prompt(sample)
samples.append(sample)
if num_chunks == 1:
results_path = os.path.join(output_path,"results.json")
response_path = os.path.join(output_path,"response.json")
out_samples,out_response = run_model(samples,model)
save_json(results_path,out_samples)
save_json(response_path,out_response)
return evaluate(out_samples,gold_answers)
elif num_chunks > 1:
results_path = os.path.join(output_path,f"results_{chunk_idx}.json")
response_path = os.path.join(output_path,f"response_{chunk_idx}.json")
out_samples,out_response = run_model(samples,model)
save_json(results_path,out_samples)
save_json(response_path,out_response)
total_results_path = os.listdir(output_path)
total_results_path = [result for result in total_results_path if result.startswith("results_")]
if len(total_results_path) == num_chunks:
total_results = {}
for result in total_results_path:
results_path = os.path.join(output_path,result)
with open(results_path,"r") as f:
partial_results = json.load(f)
total_results.update(partial_results)
os.remove(results_path)
with open(os.path.join(output_path,"results.json"),"w") as f:
json.dump(total_results,f)
return evaluate(total_results,gold_answers)
else:
return
else:
raise ValueError("num_chunks must be greater than 0")
|