如何用fastai实现sklearns StratifiedKfold？

Question

我参加了APTOS 2019年的kaggle竞赛，并试图在5个折叠上进行集成，但我在正确实现StratifiedKFold方面遇到了问题。

我试着在谷歌上搜索快速讨论，但我看不到任何解决方案。我正在使用fastai库，并且有一个预先训练好的模型。

def get_df():
    base_image_dir = os.path.join('..', 'input/aptos2019-blindness- 
    detection/')
    train_dir = os.path.join(base_image_dir,'train_images/')
    df = pd.read_csv(os.path.join(base_image_dir, 'train.csv'))
    df['path'] = df['id_code'].map(lambda x: 
    os.path.join(train_dir,'{}.png'.format(x)))
    df = df.drop(columns=['id_code'])
    df = df.sample(frac=1).reset_index(drop=True) #shuffle dataframe
    test_df = pd.read_csv('../input/aptos2019-blindness- 
    detection/sample_submission.csv')
    return df, test_df

df, test_df = get_df()

random_state = np.random.seed(2019)
skf = StratifiedKFold(n_splits=5, random_state=random_state, shuffle=True)

X = df['path']
y = df['diagnosis']

#getting the splits
for train_index, test_index in skf.split(X, y):
   print('##')
   X_train, X_test = X[train_index], X[test_index]
   y_train, y_test = y[train_index], y[test_index]
   train = X_train, y_train
   test = X_test, y_test
   train_list = [list(x) for x in train]
   test_list  = [list(x) for x in test]


data = (ImageList.from_df(df=df,path='./',cols='path') 
    .split_by_rand_pct(0.2) 
    .label_from_df(cols='diagnosis',label_cls=FloatList) 
    .transform(tfms,size=sz,resize_method=ResizeMethod.SQUISH,padding_mode='zeros') 
    .databunch(bs=bs,num_workers=4) 
    .normalize(imagenet_stats)  
   )

learn = Learner(data, 
            md_ef, 
            metrics = [qk], 
            model_dir="models").to_fp16()
learn.data.add_test(ImageList.from_df(test_df,
                             '../input/aptos2019-blindness-detection',
                                  folder='test_images',
                                  suffix='.png'))

我想使用我从skf.split得到的折叠来训练我的模型，但我不确定如何做到这一点。

Answer 1

有两种方法可以做到这一点。

1. 将'split_by_idxs‘与索引

一起使用

    data = (ImageList.from_df(df=df,path='./',cols='path')
        .split_by_idxs(train_idx=train_index, valid_idx=test_index) 
        .label_from_df(cols='diagnosis',label_cls=FloatList) 
        .transform(tfms,size=sz,resize_method=ResizeMethod.SQUISH,padding_mode='zeros') 
        .databunch(bs=bs,num_workers=4) 
        .normalize(imagenet_stats)  
       )

使用'split_by_list'的

   il = ImageList.from_df(df=df,path='./',cols='path')

   data = (il.split_by_list(train=il[train_index], valid=il[test_index]) 
       .label_from_df(cols='diagnosis',label_cls=FloatList) 
       .transform(tfms,size=sz,resize_method=ResizeMethod.SQUISH,padding_mode='zeros') 
       .databunch(bs=bs,num_workers=4) 
       .normalize(imagenet_stats)  
      )

Answer 2

这是一段代码。希望这能有所帮助。

# creating a KFold object with 5 splits 
folds = KFold(n_splits = 5, shuffle = True, random_state = 10)

# specify range of hyperparameters
# Set the parameters by cross-validation
hyper_params = [ {'gamma': [1e-2, 1e-3, 1e-4],
                     'C': [5,10]}]


# specify model
model = SVC(kernel="rbf")

# set up GridSearchCV()
model_cv = GridSearchCV(estimator = model, 
                        param_grid = hyper_params, 
                        scoring= 'accuracy', 
                        cv = folds, 
                        verbose = 1,
                        return_train_score=True)      

# fit the model
model_cv.fit(X_train, y_train)