|
- import json
- import os
- import numpy as np
- from model_url import get_model_resp, get_url_tokenizer
-
-
- def run_predict(url, log_path, few_shot = True):
- tokenizer = get_url_tokenizer()
- id_label = {'0': '不完全一致', '1': '完全一致'}
- MAIN_DIR = os.path.dirname(os.path.abspath(__file__))
- file_dir = MAIN_DIR + "/task_dataset/AFQMC/test_public.json"
- count = 0
- correct_num = 0
- acc = 0
-
- pre = ""
- devfile = MAIN_DIR + "/task_dataset/AFQMC/dev.json"
- cnt = 0
- if few_shot:
- with open(devfile, "r", encoding="utf8") as f:
- for line in f.readlines():
- cnt += 1
- line = json.loads(line)
- sentence1, sentence2, label = line["sentence1"], line["sentence2"], line["label"]
- pre += f"语句一:“{sentence1}”\n语句二:“{sentence2}”\n语句一与语句二是关于蚂蚁金融产品的疑问,两者所询问的内容是否完全一致?\n{id_label[label]}\n"
- if cnt == 5:
- break
-
- with open(file_dir, "r", encoding="utf8") as f:
- for line in f.readlines():
- count += 1
- line = json.loads(line)
- sentence1, sentence2, label = line["sentence1"], line["sentence2"], line["label"]
- example = f"{pre}语句一:“{sentence1}”\n语句二:“{sentence2}”\n语句一与语句二是关于蚂蚁金融产品的疑问,两者所询问的内容是否完全一致?\n"
- input_str_one = f"{example}不完全一致"
- input_str_two = f"{example}完全一致"
-
- input_str = []
- input_str.append(input_str_one)
- input_str.append(input_str_two)
- mask_length_list = []
- input_length_list = []
- for pred in input_str:
- input_length_list.append(len(tokenizer.encode(pred)))
- mask_length_list.append(len(tokenizer.encode(example)))
- model_resp = get_model_resp(url=url, input_str=input_str, tokens_to_generate=0, top_k=1, logprobs=True)
- return_resp = []
- for resp_item, input_length, mask_length in zip(model_resp, input_length_list, mask_length_list):
- assert len(resp_item) == input_length - 1
- item = resp_item[mask_length - 1:input_length - 1]
- return_resp.append(item)
-
- pred_list = [sum(logprobs) / len(logprobs) for logprobs in return_resp]
- answers_pred = int(np.argmax(pred_list))
- if answers_pred == int(label):
- correct_num += 1
- acc = correct_num / count
-
- print(f"AFQMC, 准确率Acc:{acc}, number: {count}")
-
- if not few_shot:
- with open(log_path + '/AFQMC_zeroshot.txt', 'w') as file:
- file.write(f"AFQMC, zero shot , Acc: {acc}, number: {count}")
- else:
- with open(log_path + '/AFQMC_fewshot.txt', 'w') as file:
- file.write(f"AFQMC, few shot , Acc: {acc}, number: {count}")
|