ValueError:输入数组的形状必须为== (..，..，[ ..，]3)，got (28，28，1)

Question

我正在尝试使用Keras的模型做一些程序，然后用AIX360的Lime explainer (这只是原始LIME的包装器)来解释它。所有数据都是MNIST灰度数字。但在我的例子中，我不能解释这个实例，因为我不知道该向解释器提供什么。

我的代码：

!pip install aix360
!pip install tensorflow==2.2.0

from __future__ import print_function
import warnings
# Supress jupyter warnings if required for cleaner output
warnings.simplefilter('ignore')

import numpy as np
import pandas as pd

import keras
import keras.layers

from keras.utils.np_utils import to_categorical # convert to one-hot-encoding
from keras.models import Sequential # Sequeantial layer addition

from aix360.algorithms.lime import LimeImageExplainer

print('Using keras:', keras.__version__)

# Load dataset
from keras.datasets import mnist
# Tuple of Numpy arrays: (x_train, y_train), (x_test, y_test).
(train, train_labels), (test, test_labels) = mnist.load_data()

# save input image dimensions
img_rows = train.shape[1]
img_cols = train.shape[2]

# Get classes and number of values
value_counts = pd.value_counts(train_labels).sort_index()
num_classes = value_counts.count()

train = train/255
test = test/255

train = train.reshape(train.shape[0], img_rows, img_cols, 1)
test = test.reshape(test.shape[0], img_rows, img_cols, 1)

model = keras.Sequential([
    keras.layers.Flatten(input_shape=(img_rows, img_cols,1)),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(num_classes)
])

model.compile(loss='sparse_categorical_crossentropy',
      optimizer='adam',
      metrics=['accuracy'])

batch_size = 128
epochs = 1

model.fit(train, train_labels,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(test, test_labels))

score = model.evaluate(test, test_labels, verbose=0)

print('Test loss:', score[0])
print('Test accuracy:', score[1])

limeExplainer = LimeImageExplainer()

limeExplainer.explain_instance(test[0], model.predict_proba)

最后一行是错误所在。不要注意模型是如何训练的，这不是问题。

编辑:编辑代码，希望它可以在codelab中运行(添加了第二行)

EDIT2:完成: tensorflow 2.2.0keras 2.4.3 aix360 0.2.0

Answer 1

我添加了这个转换，并在RGB图像上进行了训练：

def to_rgb(x):
    x_rgb = np.zeros((x.shape[0], 28, 28, 3))
    for i in range(3):
        x_rgb[..., i] = x[..., 0]
    return x_rgb

train_rgb = to_rgb(train)
test_rgb = to_rgb(test)

它起作用了：

limeExplainer.explain_instance(test_rgb[0], model.predict_proba)
100%
1000/1000 [00:00<00:00, 2598.51it/s]
<lime.lime_image.ImageExplanation at 0x7f8d20381f50>