Pytorch 交差検証について教えてください

前提・実現したいこと

以下のコードは、どのような動作を期待したものなのでしょうか。
train_inds, valid_inds = split_inds[flags.target_fold]
5分割交差検証を行うことを想定しているようなのですが、コードを見る限り、0～4の内、0foldのデータセットしか準備していないように感じております。
ご教示いただければ幸いです。

該当のソースコード

Python
1import cv2
2import numpy as np
3
4from dataclasses import dataclass, field
5from typing import Dict, Any, Tuple, Union, List
6
7
8@dataclass
9class Flags:
10    # General
11    debug: bool = True
12    outdir: str = "results/det"
13    device: str = "cuda:0"
14
15    # Data config
16    imgdir_name: str = "vinbigdata-chest-xray-resized-png-256x256"
17    # split_mode: str = "all_train"  # all_train or valid20
18    seed: int = 111
19    target_fold: int = 0  # 0~4
20    label_smoothing: float = 0.0
21    # Model config
22    model_name: str = "adv_inception_v3"#"resnet18"
23    model_mode: str = "normal"  # normal, cnn_fixed supported
24    # Training config
25    epoch: int = 20
26    batchsize: int = 8
27    valid_batchsize: int = 16
28    num_workers: int = 4
29    snapshot_freq: int = 5
30    ema_decay: float = 0.999  # negative value is to inactivate ema.
31    scheduler_type: str = ""
32    scheduler_kwargs: Dict[str, Any] = field(default_factory=lambda: {})
33    scheduler_trigger: List[Union[int, str]] = field(default_factory=lambda: [1, "iteration"])
34    aug_kwargs: Dict[str, Dict[str, Any]] = field(default_factory=lambda: {})
35    mixup_prob: float = -1.0  # Apply mixup augmentation when positive value is set.
36
37    def update(self, param_dict: Dict) -> "Flags":
38        # Overwrite by `param_dict`
39        for key, value in param_dict.items():
40            if not hasattr(self, key):
41                raise ValueError(f"[ERROR] Unexpected key for flag = {key}")
42            setattr(self, key, value)
43        return self
44
45flags_dict = {
46    "debug": False,  # Change to True for fast debug run!
47    "outdir": "results/tmp_debug",
48    # Data
49    "imgdir_name": "vinbigdata-chest-xray-resized-png-256x256",
50    # Model
51    "model_name": "adv_inception_v3",#"resnet18",
52    # Training
53    "num_workers": 4,
54    "epoch": 15,
55    "batchsize": 8,
56    "scheduler_type": "CosineAnnealingWarmRestarts",
57    "scheduler_kwargs": {"T_0": 28125},#{"T_0": 14063},#{"T_0": 7031},#{"T_0": 28125},  # 15000 * 15 epoch // (batchsize=8)
58    "scheduler_trigger": [1, "iteration"],
59    "aug_kwargs": {
60        "HorizontalFlip": {"p": 0.5},
61        "ShiftScaleRotate": {"scale_limit": 0.15, "rotate_limit": 10, "p": 0.5},
62        "RandomBrightnessContrast": {"p": 0.5},
63        "CoarseDropout": {"max_holes": 8, "max_height": 25, "max_width": 25, "p": 0.5},
64        "Blur": {"blur_limit": [3, 7], "p": 0.5},
65        "Downscale": {"scale_min": 0.25, "scale_max": 0.9, "p": 0.3},
66        "RandomGamma": {"gamma_limit": [80, 120], "p": 0.6},
67    }
68}
69
70class VinbigdataTwoClassDataset(DatasetMixin):
71    def __init__(self, dataset_dicts, image_transform=None, transform=None, train: bool = True,
72                 mixup_prob: float = -1.0, label_smoothing: float = 0.0):
73        super(VinbigdataTwoClassDataset, self).__init__(transform=transform)
74        self.dataset_dicts = dataset_dicts
75        self.image_transform = image_transform
76        self.train = train
77        self.mixup_prob = mixup_prob
78        self.label_smoothing = label_smoothing
79
80    def _get_single_example(self, i):
81        d = self.dataset_dicts[i]
82        filename = d["file_name"]
83
84        img = cv2.imread(filename)
85        if self.image_transform:
86            img = self.image_transform(img)
87        img = torch.tensor(np.transpose(img, (2, 0, 1)).astype(np.float32))
88
89        if self.train:
90            label = int(len(d["annotations"]) > 0)  # 0 normal, 1 abnormal
91            if self.label_smoothing > 0:
92                if label == 0:
93                    return img, float(label) + self.label_smoothing
94                else:
95                    return img, float(label) - self.label_smoothing
96            else:
97                return img, float(label)
98        else:
99            # Only return img
100            return img, None
101
102    def get_example(self, i):
103        img, label = self._get_single_example(i)
104        if self.mixup_prob > 0. and np.random.uniform() < self.mixup_prob:
105            j = np.random.randint(0, len(self.dataset_dicts))
106            p = np.random.uniform()
107            img2, label2 = self._get_single_example(j)
108            img = img * p + img2 * (1 - p)
109            if self.train:
110                label = label * p + label2 * (1 - p)
111
112        if self.train:
113            label_logit = torch.tensor([1 - label, label], dtype=torch.float32)
114            return img, label_logit
115        else:
116            # Only return img
117            return img
118
119    def __len__(self):
120        return len(self.dataset_dicts)
121
122dataset_dicts = get_vinbigdata_dicts(imgdir, train, debug=debug)
123dataset = VinbigdataTwoClassDataset(dataset_dicts)
124
125######以下のコードで何をやっているかの詳細を知りたいです######
126skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=flags.seed)
127y = np.array([int(len(d["annotations"]) > 0) for d in dataset_dicts])
128split_inds = list(skf.split(dataset_dicts, y))
129train_inds, valid_inds = split_inds[flags.target_fold]  # 0th fold
130train_dataset = VinbigdataTwoClassDataset(
131    [dataset_dicts[i] for i in train_inds],
132    image_transform=Transform(flags.aug_kwargs),
133    mixup_prob=flags.mixup_prob,
134    label_smoothing=flags.label_smoothing,
135)
136valid_dataset = VinbigdataTwoClassDataset([dataset_dicts[i] for i in valid_inds])
137
138

行動規範の内容に同意します

回答1件

StratifiedKFold.splitがやっているのは以下のようなことです。
ひとかたまりのデータをラベルがちゃんと入るように指定された個数のトレーニング用データとテスト用データのインデックスを生成しています。

python
1>>> from sklearn.model_selection import StratifiedKFold
2>>> import numpy as np
3>>>
4>>> X = np.array([[1,1,1],[2,2,2],[3,3,3],[4,4,4],[5,5,5],[6,6,6],[7,7,7],[8,8,8],[9,9,9]])
5>>> X = np.arange(18).reshape(9,2)
6>>> print(X)
7[[ 0  1]
8 [ 2  3]
9 [ 4  5]
10 [ 6  7]
11 [ 8  9]
12 [10 11]
13 [12 13]
14 [14 15]
15 [16 17]]
16>>> y = np.array([1,2,3,1,2,3,1,2,3])
17>>> skf = StratifiedKFold(n_splits=3)
18>>> for x in skf.split(X,y):
19...     print(x)
20...
21(array([3, 4, 5, 6, 7, 8]), array([0, 1, 2]))
22(array([0, 1, 2, 6, 7, 8]), array([3, 4, 5]))
23(array([0, 1, 2, 3, 4, 5]), array([6, 7, 8]))