File size: 6,672 Bytes
b0c0df0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import json
import os
from typing import Dict, List, Union

import datasets
import numpy as np
import pandas as pd
from datasets import Dataset, load_dataset
from PIL import Image
from tqdm import tqdm

EPS = 1e-6

RESULT_FEATURES = {
    "id": datasets.Value("int32"),
    "images": datasets.Sequence(datasets.Image()),
    "question": datasets.Value("string"),
    "ground_truth": datasets.Value("string"),
    "criteria": datasets.Value("string"),
    "subtask": datasets.Value("string"),
    "response": datasets.Value("string"),
    "score": datasets.Value("float32"),
    "reason": datasets.Value("string"),
}

SUBTASKS = ["Concrete Recognition", "Analytical Questions", "Evaluative Questions", "Divergent Thinking", "Real-world Assistance"]


def load_images(config) -> Dict[int, List[Image.Image]]:
    dataset = datasets.load_dataset(config["dataset_path"], config["dataset_name"], split=config["test_split"])
    images = {}
    for data in tqdm(dataset, desc="Loading images"):
        images[data["id"]] = data["images"]
    return images


def get_hf_results(results, detailed_results):
    live_bench_images = load_images(results["configs"]["live_bench"])

    def load_results():
        for result in tqdm(detailed_results["logs"], desc="Loading results"):
            doc = result["doc"]
            res = {}
            res["id"] = doc["id"]
            res["images"] = live_bench_images[doc["id"]]
            res["question"] = doc["question"]
            res["ground_truth"] = doc["answer"]
            res["criteria"] = doc["criteria"]
            res["subtask"] = doc["subtask"]
            res["response"] = result["filtered_resps"][0]
            res["score"] = result["gpt4_eval_score"]["rating"]
            res["reason"] = result["gpt4_eval_score"]["explanation"]
            yield res

    result_dataset = Dataset.from_generator(load_results, features=datasets.Features(RESULT_FEATURES))

    return result_dataset


def preview_results(results, heading: str):
    HEADING = "=" * 15 + " " + heading + " " + "=" * 15
    ENDING = "=" * len(HEADING)
    print(HEADING)
    print(results)
    print(ENDING)


def calculate_score(results: Dataset):
    results = results.to_pandas()

    sum_score, count = 0, 0
    score = {}
    for subtask in SUBTASKS:
        score[subtask] = []
    for index, result in tqdm(results.iterrows(), total=len(results), desc="Calculating score"):
        if result["score"] == -1:
            continue
        sum_score += result["score"] / 10
        count += 1
        subtask = result["subtask"]
        if subtask not in SUBTASKS:
            subtask = "Further Insights"
        score[result["subtask"]].append(result["score"] / 10)
    res = [(subtask, len(score[subtask]), np.mean(score[subtask]) * 100) for subtask in SUBTASKS]
    res.append(("Total", count, sum_score / count * 100))
    res = pd.DataFrame(res, columns=["Subtask", "Count", "Score"])

    return res


def get_results(folder):
    detailed_file = os.path.join(folder, "live_bench.json")
    results_file = os.path.join(folder, "results.json")

    with open(results_file, "r") as f:
        results = json.load(f)

    assert "live_bench" in results["configs"], "No live_bench config found in results.json"
    final_score = results["results"]["live_bench"]["gpt4_eval_score,none"]
    model_configs = results["model_configs"]
    version = results["configs"]["live_bench"]["metadata"]["version"]

    assert model_configs["limit"] is None, "Model limit is not None, please check if the model is tested on the full dataset"

    with open(detailed_file, "r") as f:
        detailed_results = json.load(f)

    hf_results = get_hf_results(results, detailed_results)
    preview_results(hf_results.to_pandas().iloc[0], "Detailed Results")
    score = calculate_score(hf_results)
    preview_results(score, "Final Score")

    assert (
        abs(score[score["Subtask"] == "Total"]["Score"] - final_score) <= EPS
    ).all(), f"Final score does not match the calculated score, the score calculated by the script is {score[score['Subtask'] == 'Total']['Score'].values[0]} and the final score in the log is {final_score}."

    return hf_results, score, version


def upload_results(

    hf_results: Dataset,

    score: pd.DataFrame,

    model_name,

    dataset_version,

    log_folder="logs",

):
    hf_results.push_to_hub(
        "lmms-lab/LiveBenchDetailedResults",
        config_name=dataset_version,
        split=model_name.replace("-", "_"),
    )
    if not os.path.exists(log_folder):
        os.makedirs(log_folder)
    score_path = os.path.abspath(os.path.join(log_folder, f"{dataset_version}_{model_name}.csv"))
    score.to_csv(score_path, index=False)
    print(f"Results saved to {score_path}")
    score_dict = {item["Subtask"]: item["Score"] for index, item in score.iterrows()}
    score_dict["Model Name"] = model_name
    try:
        hf_score = datasets.load_dataset("lmms-lab/LiveBenchResults", dataset_version, split="test")
    except:
        hf_score = Dataset.from_dict({subtask: [] for subtask in ["Model Name", "Total"] + SUBTASKS})
    hf_score = hf_score.add_item(score_dict)
    df_score = pd.DataFrame(hf_score)
    df_score = df_score.drop_duplicates(subset=["Model Name"], keep="last")
    df_score = df_score[["Model Name", "Total"] + SUBTASKS]
    hf_score = Dataset.from_pandas(df_score)
    hf_score.push_to_hub("lmms-lab/LiveBenchResults", dataset_version, split="test")


if __name__ == "__main__":
    argparse = argparse.ArgumentParser()
    argparse.add_argument("--folder", "-f", type=str, required=True, help="Results folder")
    argparse.add_argument("--name", "-m", type=str, required=True, help="Model name")
    argparse.add_argument("--log_folder", "-l", type=str, default="logs", help="Log folder")
    argparse.add_argument("--force", "-F", action="store_true", help="Force upload")
    args = argparse.parse_args()
    hf_results, score, version = get_results(args.folder)
    print(f"Results will be uploaded with model name {args.name} and model version {version}")
    if args.force is False:
        print("Are you sure you want to upload the results? (y/n)", end=" ")
        while True:
            choice = input().lower()
            if choice == "y":
                break
            elif choice == "n":
                exit()
            else:
                print("Invalid choice, please enter 'y' or 'n'")
    upload_results(hf_results, score, args.name, version, args.log_folder)