File size: 4,563 Bytes
c5d2eae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json, os, hashlib
from pathlib import Path

# code dataset processor
# modality: image_depth, preprocessing: progressive


# --- real data source: code ---
TV_DATASET = None
HF_CANDIDATES = ['code_search_net', 'codeparrot/github-code']
IMAGE_FIELD = None
TEXT_FIELD = 'func_code_string'
LABEL_FIELD = 'func_documentation_string'
PROMPT_TEMPLATE = 'a photo of a {label}'
DATASET_URL = 'https://huggingface.co/datasets/code_search_net'

def fetch_real_samples(max_samples=5000, cache_dir='./_cache'):
    # 本地没有数据时自动下载真实公开数据集: torchvision -> HuggingFace -> 手动说明
    out = []
    if TV_DATASET is not None:
        try:
            import torchvision
            ctor = getattr(torchvision.datasets, TV_DATASET)
            try:
                ds = ctor(root=cache_dir, split='train', download=True)
            except TypeError:
                try:
                    ds = ctor(root=cache_dir, train=True, download=True)
                except TypeError:
                    ds = ctor(root=cache_dir, download=True)
            classes = getattr(ds, 'classes', None)
            os.makedirs(os.path.join(cache_dir, 'tv'), exist_ok=True)
            for i, item in enumerate(ds):
                if len(out) >= max_samples:
                    break
                img, label = item[0], item[1]
                name = classes[label] if classes else str(label)
                p = os.path.join(cache_dir, 'tv', str(i) + '.png')
                try:
                    img.save(p)
                except Exception:
                    continue
                out.append({'image': p, 'text': PROMPT_TEMPLATE.format(label=name)})
            if out:
                return out
        except Exception as e:
            print('torchvision load failed:', e)
    for repo in HF_CANDIDATES:
        try:
            from datasets import load_dataset
            try:
                ds = load_dataset(repo, split='train', streaming=True)
            except Exception:
                ds = load_dataset(repo, split='train')
            img_dir = os.path.join(cache_dir, 'hf_images')
            os.makedirs(img_dir, exist_ok=True)
            for i, ex in enumerate(ds):
                if len(out) >= max_samples:
                    break
                txt = None
                if TEXT_FIELD is not None and TEXT_FIELD in ex:
                    v = ex[TEXT_FIELD]
                    txt = v if isinstance(v, str) else ' '.join(map(str, v if isinstance(v, (list, tuple)) else [v]))
                if txt is None and LABEL_FIELD in ex:
                    txt = PROMPT_TEMPLATE.format(label=ex[LABEL_FIELD])
                if txt is None:
                    continue
                out.append({'text': txt})
            if out:
                return out
        except Exception as e:
            print('HF load failed for', repo, ':', e)
    print('Automatic download failed. Please get the data manually from:')
    print('   ' + DATASET_URL)
    return out

def build_dataset(src, dst, sz=224):
    samples = []
    for f in Path(src).glob('*.jsonl'):
        with open(f) as fp:
            for line in fp:
                if line.strip():
                    samples.append(json.loads(line))

    if not samples:
        samples = fetch_real_samples()

    # dedup
    seen = set()
    unique = []
    for s in samples:
        fp = s.get('image', s.get('audio', ''))
        if fp and os.path.exists(fp):
            h = hashlib.md5(open(fp, 'rb').read()).hexdigest()
            if h not in seen:
                seen.add(h)
                unique.append(s)
        else:
            unique.append(s)

    os.makedirs(dst, exist_ok=True)
    out = []
    for s in unique:
        item = {}
        if 'image' in s:
            from PIL import Image
            img = Image.open(s['image']).convert('RGB').resize((sz, sz))
            p = os.path.join(dst, os.path.basename(s['image']))
            img.save(p, 'JPEG', quality=95)
            item['image'] = p
        item['text'] = s.get('code', s.get('text', ''))
        item['domain'] = 'code'
        out.append(item)

    with open(os.path.join(dst, 'dataset.jsonl'), 'w') as f:
        for d in out:
            f.write(json.dumps(d, ensure_ascii=False) + '\n')
    return out

if __name__ == '__main__':
    import sys
    result = build_dataset(sys.argv[1], sys.argv[2])
    print(f'Processed {len(result)} samples')