- android - 多次调用 OnPrimaryClipChangedListener
- android - 无法更新 RecyclerView 中的 TextView 字段
- android.database.CursorIndexOutOfBoundsException : Index 0 requested, 光标大小为 0
- android - 使用 AppCompat 时,我们是否需要明确指定其 UI 组件(Spinner、EditText)颜色
给定一个形状张量(2,3,3,1),其中批量大小为 2,每个矩阵的形状为(3,3,1)。如何从每个矩阵中找到 k 个最大元素并创建一个 one-hot 矩阵,使得这 k 个位置的条目为 1,其他位置为 0。示例:(请注意,为了简单起见,每个条目都是 float ,使用整数)
input_tensor=[[[1, 5, 7],
[2, 8, 1],
[3, 9, 1],
],
[[0, 9, 5],
[6, 0, 4],
[3, 0, 8]
]
]
k=3 的单热张量:
output_tensor=[[[0, 0, 1],
[0, 1, 0],
[0, 1, 0],
],
[[0, 1, 0],
[1, 0, 0],
[0, 0, 1]
]
]
tf.nn.top_k 将仅返回最后一个维度中的 k 个最大元素。如何从 3d 张量中获取 k 个最大元素 ex:(3,3,1)。此外 tf.one_hot 会在指定深度和给定索引的每一行中放置一个 1,但此处情况并非如此。
最佳答案
如果你想从每个矩阵中找到k
个最大的元素,可以使用以下方法。
import tensorflow as tf
input_tensor = tf.constant([[[1, 5, 7],[2, 8, 1],[3, 9, 1]],
[[0, 9, 5],[6, 0, 4],[3, 0, 8]]],dtype=tf.int32)
k_tf = tf.placeholder(shape=(),dtype=tf.int32)
temp = tf.reshape(input_tensor,shape=(input_tensor.shape[0],-1))
# [[1 5 7 2 8 1 3 9 1]
# [0 9 5 6 0 4 3 0 8]]
result = tf.reduce_sum(tf.one_hot(indices=tf.nn.top_k(temp,k=k_tf)[1], depth=temp.shape[1]), axis=1)
# [[0. 0. 1. 0. 1. 0. 0. 1. 0.]
# [0. 1. 0. 1. 0. 0. 0. 0. 1.]]
result = tf.reshape(result,input_tensor.shape)
with tf.Session() as sess:
print('k=2:')
print(sess.run(result, feed_dict={k_tf: 2}))
print('k=3:')
print(sess.run(result,feed_dict={k_tf:3}))
k=2:
[[[0. 0. 0.]
[0. 1. 0.]
[0. 1. 0.]]
[[0. 1. 0.]
[0. 0. 0.]
[0. 0. 1.]]]
k=3:
[[[0. 0. 1.]
[0. 1. 0.]
[0. 1. 0.]]
[[0. 1. 0.]
[1. 0. 0.]
[0. 0. 1.]]]
关于python - 如何为 4d 张量中的 k 个最大元素创建一个单热张量?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/55665304/
很难说出这里问的是什么。这个问题是含糊的、模糊的、不完整的、过于宽泛的或修辞性的,无法以目前的形式得到合理的回答。如需帮助澄清此问题以便重新打开它,visit the help center 。 已关
我们可以说 O(K + (N-K)logK)相当于O(K + N logK)对于 1 < = K <= N ? 最佳答案 简短的回答是它们不等价,这取决于k 的值。如果k等于N,那么第一个复杂度是O(
我有以下解决方案,但我从其他评论者那里听说它是 O(N * K * K),而不是 O(N * K)其中 N 是 K 列表的(最大)长度,K 是列表的数量。例如,给定列表 [1, 2, 3] 和 [4,
我试图理解这些语法结构之间的语义差异。 if ((i% k) == (l % k) == 0) 和 if ((i % k) == 0 && (l % k) == 0) 最佳答案 您的特定表达式((i
我有时会使用一维数组: A = np.array([1, 2, 3, 4]) 或 2D 阵列(使用 scipy.io.wavfile 读取单声道或立体声信号): A = np.array([[1, 2
在文档聚类过程中,作为数据预处理步骤,我首先应用奇异向量分解得到U、S和Vt 然后通过选择适当数量的特征值,我截断了 Vt,这让我从阅读的内容中得到了很好的文档-文档相关性 here .现在我正在对矩
我问的是关于 Top K 算法的问题。我认为 O(n + k log n) 应该更快,因为……例如,如果您尝试插入 k = 300 和 n = 100000000,我们可以看到 O(n + k log
这个问题与另一个问题R:sample()密切相关。 。我想在 R 中找到一种方法来列出 k 个数字的所有排列,总和为 k,其中每个数字都是从 0:k 中选择的。如果k=7,我可以从0,1,...,7中
我目前正在评估基于隐式反馈的推荐系统。我对排名任务的评估指标有点困惑。具体来说,我希望通过精确度和召回率来进行评估。 Precision@k has the advantage of not requ
我在 Python 中工作,需要找到一种算法来生成所有可能的 n 维 k,k,...,k 数组,每个数组都沿轴有一行 1。因此,该函数接受两个数字 - n 和 k,并且应该返回一个数组列表,其中包含沿
我们有 N 对。每对包含两个数字。我们必须找到最大数 K,这样如果我们从给定的 N 对中取 J (1 2,如果我们选择三对 (1,2),我们只有两个不同的数字,即 1 和 2。 从一个开始检查每个可能
鉴于以下问题,我不能完全确定我当前的解决方案: 问题: 给定一个包含 n 元素的最大堆,它存储在数组 A 中,是否可以打印所有最大的 K 元素在 O(K*log(K)) 中? 我的回答: 是的,是的,
我明白了: val vector: RDD[(String, Array[String])] = [("a", {v1,v2,..}),("b", {u1,u2,..})] 想转换成: RDD[(St
我有 X 个正数,索引为 x_i。每个 x_i 需要进入 K 组之一(其中 K 是预先确定的)。令 S_j 为 K_j 中所有 x_i 的总和。我需要分配所有 x_i 以使所有 S_j 的方差最小化。
关闭。这个问题是not reproducible or was caused by typos .它目前不接受答案。 这个问题是由于错别字或无法再重现的问题引起的。虽然类似的问题可能是on-topi
我正在研究寻找原始数的算法,看到下面的语句,我不明白为什么。 while (k*k <= n) 优于 while (k <= Math.sqrt(n)) 是因为函数调用吗?该调用函数使用更多资源。 更
我想找到一种尽可能快的方法来将两个小 bool 矩阵相乘,其中小意味着 8x8、9x9 ... 16x16。这个例程会被大量使用,所以需要非常高效,所以请不要建议直截了当的解决方案应该足够快。 对于
有没有一种惯用的方法来获取 Set和 Function ,并获得 Map实时取景? (即 Map 由 Set 和 Function 组合支持,例如,如果将元素添加到 Set ,则相应的条目也存在于 M
这个问题在这里已经有了答案: Can a local variable's memory be accessed outside its scope? (20 个答案) returning addr
给定一个矩阵:- k = [1 2 3 ; 4 5 6 ; 7 8 NaN]; 如果我想用 0 替换一个数字,比如 2,我可以使用这个:k(k==2) =
我是一名优秀的程序员,十分优秀!