부캠 TIL 0826

Binary classification, multi class classification

참고

참고

loss 함수 정리된 페이지

이진분휴시 마지막 노드 1개, bcewithlogitloss
다진분류시 마지막 노드 클래스 개수, crossentropy loss
pred, label 자료형과 shape은 위의 사진 참고

model.eval, torch.no_grad 차이

참고

train test split

num_train = len(train_dataset)
    indices = list(range(num_train))
    split = int(np.floor(valid_size * num_train))

    if shuffle:
        np.random.seed(random_seed)
        np.random.shuffle(indices)

    train_idx, valid_idx = indices[split:], indices[:split]
    train_sampler = SubsetRandomSampler(train_idx)
    valid_sampler = SubsetRandomSampler(valid_idx)

    train_loader = torch.utils.data.DataLoader(
        train_dataset, batch_size=batch_size, sampler=train_sampler,
        num_workers=num_workers, pin_memory=pin_memory,
    )
    valid_loader = torch.utils.data.DataLoader(
        valid_dataset, batch_size=batch_size, sampler=valid_sampler,
        num_workers=num_workers, pin_memory=pin_memory,
    )

strata를 줄 수 있음

from sklearn.model_selection import train_test_split

dataset = TotalDataset
dataset_size = len(dataset)

train_idx, val_idx = train_test_split(np.arange(dataset_size),
                                    test_size=validation_split,
                                    shuffle=True,
                                    random_state=random_seed,
                                    stratify=TotalDataset.labels)

train_sampler = SubsetRandomSampler(train_idx)
valid_sampler = SubsetRandomSampler(val_idx)

train_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, pin_memory=4,
                                           sampler=train_sampler)
validation_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, pin_memory=4,
                                                sampler=valid_sampler)