| from transformers import pipeline |
| import torch |
| import pandas as pd |
| import sys |
| import os |
| sys.path.append(os.path.join(os.path.dirname(os.path.abspath(__file__)), '..', '..', '..', 'document_retrieval', 'Decompose_retrieval')) |
| from pipline import * |
|
|
|
|
|
|
| def ask_gpt(query): |
| messages = [{"role": "user", "content": query}] |
| result = client.chat.completions.create(messages=messages, model="meta-llama/Meta-Llama-3-8B-Instruct") |
| return result.choices[0].message.content |
|
|
|
|
| def get_accuracy_re(true_imgs, re_imgs): |
| |
| |
| cnt = 0 |
| for i in range(len(true_imgs)): |
| i_gt = true_imgs[i].rsplit('.', 1)[0] |
| image_retrieval = re_imgs[i].rsplit('.', 1)[0] |
| if i_gt == image_retrieval: |
| cnt += 1 |
| return cnt/len(true_imgs) |
|
|
|
|
| def get_rag_answers(q_list, sub_q_list, dataset_name = "multiqa"): |
| if dataset_name == "multiqa": |
| ans_img_list, re_im = eval_acc(q_list, sub_q_list, patch_emb_by_img_ls) |
| ans = [] |
| re_imgs = [] |
| for i in range(len(ans_img_list)): |
| ans.append(ans_img_list[i][0]) |
| re_imgs.append(re_im[i]) |
| return ans, re_imgs |
| |
|
|
| def exact_match(predictions, ground_truths): |
| |
| score = [] |
| for i in range(len(predictions)): |
| pred = predictions[i].lower() |
| gt = ground_truths[i].lower() |
| if gt == pred: |
| score.append(1) |
| else: |
| score.append(0) |
| return score |
|
|
|
|
| def get_accuracy_multiqa(predictions, ground_truths): |
| score = exact_match(predictions, ground_truths) |
| return score |
|
|
| dataset_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "multiqa_test.csv") |
| client = OpenAI(api_key="0",base_url="http://0.0.0.0:8000/v1") |
|
|
| single_task_df = pd.read_csv(dataset_path, header=0) |
| raw_data = single_task_df[['query', 'answer', 'image']] |
| raw_queries = list(raw_data['query']) |
| true_answers = list(raw_data['answer']) |
| true_images = list(raw_data['image']) |
|
|
|
|
| raw_op_prompts = [] |
| for i in range(len(raw_data)): |
| raw_pred = ask_gpt(raw_queries[i]) |
| raw_op_prompts.append(raw_pred) |
|
|
|
|
| pred_answers, re_imgs = get_rag_answers(raw_queries, raw_op_prompts, "multiqa") |
|
|
|
|
| accuracy = get_accuracy_multiqa( |
| predictions = pred_answers, |
| ground_truths = true_answers |
| ) |
| re_acc = get_accuracy_re(true_images , re_imgs) |
|
|
| accuracy = np.average(accuracy) |
|
|
| print(accuracy, re_acc) |
|
|
