Python sklearn pipeline中分类的图像数组-ValueError:使用序列设置数组元素

我有一个图像，我想将其分类为a或B。为此，我加载并调整其大小为160x160，然后将2D数组转换为1D，并将其添加到数据帧：

from pandas import DataFrame
from scipy.misc import imread, imresize
rows = []
for product in products:
    try:
        relevant = product.categoryrelevant.all()[0].relevant
    except IndexError:
        relevant = False
    if relevant:
        relevant = "A"
    else:
        relevant = "B"
    # this exists for all pictures
    image_array = imread("{}/{}".format(MEDIA_ROOT, product.picture_file.url))
    image_array = imresize(image_array, (160, 160))
    image_array = image_array.reshape(-1)
    print(image_array)
    # [254 254 252 ..., 255 255 253]
    print(image_array.shape)
    # (76800,)
    rows.append({"id": product.pk, "image": image_array, "class": relevant})
    index.append(product)
df = DataFrame(rows, index=index)

我想要的不仅仅是用于以后分类的图像（例如，产品描述），因此我使用了一个带有FeatureUnion的管道（即使它现在只有图像）。ItemSelector取自此处：

它接受“图像”列中的值。或者，可以执行

train_X=df.iloc[train_index][“image”].值

，但我想稍后添加其他列

def randomforest_image_pipeline():
    """Returns a RandomForest pipeline."""
    return Pipeline([
        ("union", FeatureUnion(
            transformer_list=[
                ("image", Pipeline([
                    ("selector", ItemSelector(key="image")),
                ]))
            ],
            transformer_weights={
                "image": 1.0
            },
        )),
        ("classifier", RandomForestClassifier()),
    ])

然后使用KFold进行分类：

from sklearn.model_selection import KFold
kfold(tested_pipeline=randomforest_image_pipeline(), df=df)
def kfold(tested_pipeline=None, df=None, splits=6):
    k_fold = KFold(n_splits=splits)
    for train_indices, test_indices in k_fold.split(df):
        # training set
        train_X = df.iloc[train_indices]
        train_y = df.iloc[train_indices]['class'].values
        # test set
        test_X = df.iloc[test_indices]
        test_y = df.iloc[test_indices]['class'].values
        for val in train_X["image"]:
            print(len(val), val.dtype, val.shape)
            # 76800 uint8 (76800,) for all
        tested_pipeline.fit(train_X, train_y) # crashes in this call
        pipeline_predictions = tested_pipeline.predict(test_X)
        ...

但是对于

.fit

我得到以下错误：

Traceback (most recent call last):
  File "<path>/project/classifier/classify.py", line 362, in <module>
    best = best_pipeline(dataframe=data, f1_scores=f1_dict, get_fp=True)
  File "<path>/project/classifier/classify.py", line 351, in best_pipeline
    confusion_list=confusion_list, get_fp=get_fp)
  File "<path>/project/classifier/classify.py", line 65, in kfold
    tested_pipeline.fit(train_X, train_y)
  File "/usr/local/lib/python3.5/dist-packages/sklearn/pipeline.py", line 270, in fit
    self._final_estimator.fit(Xt, y, **fit_params)
  File "/usr/local/lib/python3.5/dist-packages/sklearn/ensemble/forest.py", line 247, in fit
    X = check_array(X, accept_sparse="csc", dtype=DTYPE)
  File "/usr/local/lib/python3.5/dist-packages/sklearn/utils/validation.py", line 382, in check_array
    array = np.array(array, dtype=dtype, order=order, copy=copy)
ValueError: setting an array element with a sequence.

数组

在崩溃的行中看起来像这样（从调试器复制）：

---

如何解决此问题？

此错误是因为您正在将图像的所有数据（即76800个特征）保存在列表中，而该列表保存在数据框的一列中

因此，当您使用ItemSelector选择该列时，该列的输出将是形状

（Train_len，）

的一维数组。76800的内部尺寸对于FeatureUnion或后续估计器不可见

更改ItemSelector的

transform（）

函数，以返回正确的具有形状的二维数据数组（Train_len，76800）。只有到那时它才会起作用

改为：

def transform(self, data_dict):
    return np.array([np.array(x) for x in data_dict[self.key]])

---

如果您不懂任何东西，请随时提问。

@Lomtrur太棒了！现在确保您在FeatureUnion中添加的其他变压器也返回一个二维阵列。只有这样，它们才能正确组合。

[array([ 255.,  255.,  255., ...,  255.,  255.,  255.])
 array([ 255.,  255.,  255., ...,  255.,  255.,  255.])
 array([ 255.,  255.,  255., ...,  255.,  255.,  255.]) ...,
 array([ 255.,  255.,  255., ...,  255.,  255.,  255.])
 array([ 255.,  255.,  255.

def transform(self, data_dict):
    return np.array([np.array(x) for x in data_dict[self.key]])