Define a slice layer using Lamda layer
def slice(x, h1, h2, w1, w2):
 """ Define a tensor slice function
 """
 return x[:, h1:h2, w1:w2, :]

# Add slice layer
slice_1 = Lambda(slice, arguments={'h1': 0, 'h2': 6, 'w1': 0, 'w2': 6})(sliced)
# As for tensorfow backend, Lambda doesn't need output shape argument
slice_2 = Lambda(slice, arguments={'h1': 0, 'h2': 6, 'w1': 6, 'w2': 12})(sliced)

import tensorflow as tf

with tf.Session() as sess:
 a = tf.constant([1, 2, 3, 4, 5])
 b = tf.slice(a, [0], [2]) # 第一个维度从 0 开始，截 2 个
 c = tf.slice(a, [2], [3]) # 第一个维度从 2 开始，截 3 个
 print(a.eval())
 print(b.eval())
 print(c.eval())

from keras.layers import Lambda
from keras.models import Sequential
import numpy as np

a = np.array([[1, 2, 3, 4, 5]])
model = Sequential([
 Lambda(lambda a: a[:, :2], input_shape=[5]) # 第二维截前 2 个
])

print(model.predict(a))