File size: 4,531 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
import datetime
import json
import os

from loguru import logger as eval_logger
from pycocoevalcap.eval import Bleu, Cider, COCOEvalCap, Meteor, Rouge, Spice
from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
from pycocotools.coco import COCO

from lmms_eval.tasks._task_utils.file_utils import generate_submission_file

dir_name = os.path.dirname(os.path.abspath(__file__))

FLICKR_METRICS = ["Bleu_4", "Bleu_3", "Bleu_2", "Bleu_1", "METEOR", "ROUGE_L", "CIDEr"]  # , "SPICE"]


def flickr_doc_to_visual(doc):
    return [doc["image"].convert("RGB")]


def flickr_doc_to_text(doc):
    # question = "Please carefully observe the image and come up with a caption for the image"
    return f"Provide a one-sentence caption for the provided image."


def flickr_process_result(doc, result):
    """
    Args:
        doc: a instance of the eval dataset
        results: [pred]
    Returns:
        a dictionary with key: metric name, value: metric value
    """
    pred = result[0] if len(result) > 0 else ""
    image_id = int(doc["img_id"])

    data_dict = {"answer": doc["caption"], "pred": pred, "image_id": image_id}

    return {f"flickr_{metric}": data_dict for metric in FLICKR_METRICS}


def flickr_aggregation_result(results, metric, args):
    scorers = [(Bleu(4), "Bleu_1"), (Bleu(4), "Bleu_2"), (Bleu(4), "Bleu_3"), (Bleu(4), "Bleu_4"), (Meteor(), "METEOR"), (Rouge(), "ROUGE_L"), (Cider(), "CIDEr")]  # , (Spice(), "SPICE")]
    scorers_dict = {s[1]: s for s in scorers}

    stored_results = []
    # In order to make the coco eval tools to successfully create index
    # We need at least two dict in the dataset
    # 'annotation' and 'images'
    # 'annotation' exactly reproduce the original annotation
    # 'images' however only need the image id which is contained in the file name
    dataset = {"annotations": [], "images": []}
    idx = 0
    for result in results:
        stored_results.append({"image_id": int(result["image_id"]), "caption": result["pred"]})
        for a in result["answer"]:
            dataset["annotations"].append({"image_id": int(result["image_id"]), "caption": a, "id": idx})
            idx += 1
        dataset["images"].append({"id": int(result["image_id"])})

    coco = COCO()
    # Manually create index here
    coco.dataset = dataset
    coco.createIndex()

    flickr_result = coco.loadRes(stored_results)
    flickr_eval = COCOEvalCap(coco, flickr_result)

    imgIds = flickr_eval.params["image_id"]
    gts = {}
    res = {}
    for imgId in imgIds:
        gts[imgId] = flickr_eval.coco.imgToAnns[imgId]
        res[imgId] = flickr_eval.cocoRes.imgToAnns[imgId]

    eval_logger.info("tokenization...")
    tokenizer = PTBTokenizer()
    gts = tokenizer.tokenize(gts)
    res = tokenizer.tokenize(res)

    eval_logger.info(f"Computing {metric} scores...")

    score, scores = scorers_dict[metric][0].compute_score(gts, res)
    # When metric is one of the Bleu, score will be a list
    if type(score) == list:
        n = int(metric.split("_")[-1])
        score = score[n - 1]

    path = generate_submission_file(f"flickr30k_captions_val2014_alg_results_{metric}.json", args)

    eval_logger.info("Storing prediction that can be submitted to the server ...")
    with open(path, "w") as f:
        json.dump(stored_results, f, indent=4)

    return score


def flickr_bleu4(results, args):
    return flickr_aggregation_result(results, "Bleu_4", args)


def flickr_bleu3(results, args):
    return flickr_aggregation_result(results, "Bleu_3", args)


def flickr_bleu2(results, args):
    return flickr_aggregation_result(results, "Bleu_2", args)


def flickr_bleu1(results, args):
    return flickr_aggregation_result(results, "Bleu_1", args)


def flickr_meteor(results, args):
    return flickr_aggregation_result(results, "METEOR", args)


def flickr_rougel(results, args):
    return flickr_aggregation_result(results, "ROUGE_L", args)


def flickr_cider(results, args):
    return flickr_aggregation_result(results, "CIDEr", args)


def flickr_spice(results, args):
    return flickr_aggregation_result(results, "SPICE", args)


def flickr_test_process_result(doc, result):
    """
    Args:
        doc: a instance of the eval dataset
        results: [pred]
    Returns:
        a dictionary with key: metric name (in this case flickr_passthrough), value: metric value
    """
    # The question id in our dataset is the image file itself
    image_id = doc["img_id"]
    return {"flickr_passthrough": {"pred": result, "image_id": image_id}}