张量排序

目录

• Outline
• Sort/argsort
  • 一维
  • 二维
• Top_k
  • Top_one
  • Top-k accuracy
  • 示例

Outline

• Sort/argsort

• Topk

• Top-5 Acc.

Sort/argsort

一维

import tensorflow as tf

a = tf.random.shuffle(tf.range(5))
a

<tf.Tensor: id=59, shape=(5,), dtype=int32, numpy=array([4, 0, 3, 2, 1], dtype=int32)>

tf.sort(a, direction='DESCENDING')

<tf.Tensor: id=69, shape=(5,), dtype=int32, numpy=array([4, 3, 2, 1, 0], dtype=int32)>

# 返回索引
tf.argsort(a, direction='DESCENDING')

<tf.Tensor: id=81, shape=(5,), dtype=int32, numpy=array([0, 2, 3, 4, 1], dtype=int32)>

idx = tf.argsort(a, direction='DESCENDING')
tf.gather(a, idx)

<tf.Tensor: id=94, shape=(5,), dtype=int32, numpy=array([4, 3, 2, 1, 0], dtype=int32)>

二维

a = tf.random.uniform([3, 3], maxval=10, dtype=tf.int32)
a

<tf.Tensor: id=99, shape=(3, 3), dtype=int32, numpy=
array([[1, 9, 4],
       [2, 1, 4],
       [3, 6, 0]], dtype=int32)>

tf.sort(a)

<tf.Tensor: id=112, shape=(3, 3), dtype=int32, numpy=
array([[1, 4, 9],
       [1, 2, 4],
       [0, 3, 6]], dtype=int32)>

tf.sort(a, direction='DESCENDING')

<tf.Tensor: id=122, shape=(3, 3), dtype=int32, numpy=
array([[9, 4, 1],
       [4, 2, 1],
       [6, 3, 0]], dtype=int32)>

idx = tf.argsort(a)
idx

<tf.Tensor: id=146, shape=(3, 3), dtype=int32, numpy=
array([[0, 2, 1],
       [1, 0, 2],
       [2, 0, 1]], dtype=int32)>

Top_k

• Only return top-k values and indices

Top_one

a

<tf.Tensor: id=99, shape=(3, 3), dtype=int32, numpy=
array([[1, 9, 4],
       [2, 1, 4],
       [3, 6, 0]], dtype=int32)>

# 返回前2个值
res = tf.math.top_k(a, 2)
res

TopKV2(values=<tf.Tensor: id=160, shape=(3, 2), dtype=int32, numpy=
array([[9, 4],
       [4, 2],
       [6, 3]], dtype=int32)>, indices=<tf.Tensor: id=161, shape=(3, 2), dtype=int32, numpy=
array([[1, 2],
       [2, 0],
       [1, 0]], dtype=int32)>)

res.values

<tf.Tensor: id=160, shape=(3, 2), dtype=int32, numpy=
array([[9, 4],
       [4, 2],
       [6, 3]], dtype=int32)>

res.indices

<tf.Tensor: id=161, shape=(3, 2), dtype=int32, numpy=
array([[1, 2],
       [2, 0],
       [1, 0]], dtype=int32)>

Top-k accuracy

• Prob:[0.1,0.2,0.3,0.4]

• Lable:[2]

• Only consider top-1 prediction:[3]

• Only consider top-2 prediction:[3,2]

• Only consider top-3 prediction:[3,2,1]

prob = tf.constant([[0.1, 0.2, 0.7], [0.2, 0.7, 0.1]])
target = tf.constant([2, 0])

# 概率最大的索引在最前面
k_b = tf.math.top_k(prob, 3).indices
k_b

<tf.Tensor: id=190, shape=(2, 3), dtype=int32, numpy=
array([[2, 1, 0],
       [1, 0, 2]], dtype=int32)>

k_b = tf.transpose(k_b, [1, 0])
k_b

<tf.Tensor: id=193, shape=(3, 2), dtype=int32, numpy=
array([[2, 1],
       [1, 0],
       [0, 2]], dtype=int32)>

# 对真实值broadcast，与prod比较
target = tf.broadcast_to(target, [3, 2])
target

<tf.Tensor: id=196, shape=(3, 2), dtype=int32, numpy=
array([[2, 0],
       [2, 0],
       [2, 0]], dtype=int32)>

示例

def accuracy(output, target, topk=(1, )):
    maxk = max(topk)
    batch_size = target.shape[0]

    pred = tf.math.top_k(output, maxk).indices
    pred = tf.transpose(pred, perm=[1, 0])
    target_ = tf.broadcast_to(target, pred.shape)
    correct = tf.equal(pred, target_)

    res = []
    for k in topk:
        correct_k = tf.cast(tf.reshape(correct[:k], [-1]), dtype=tf.float32)
        correct_k = tf.reduce_sum(correct_k)
        acc = float(correct_k / batch_size)
        res.append(acc)

    return res

# 10个样本6类
output = tf.random.normal([10, 6])
# 使得所有样本的概率加起来为1
output = tf.math.softmax(output, axis=1)
# 10个样本对应的标记
target = tf.random.uniform([10], maxval=6, dtype=tf.int32)
print(f'prob: {output.numpy()}')
pred = tf.argmax(output, axis=1)
print(f'pred: {pred.numpy()}')
print(f'label: {target.numpy()}')

acc = accuracy(output, target, topk=(1, 2, 3, 4, 5, 6))
print(f'top-1-6 acc: {acc}')

prob: [[0.12232917 0.18645659 0.27771464 0.17322136 0.14854735 0.09173083]
 [0.02338449 0.01026637 0.11773597 0.69083494 0.03814701 0.11963127]
 [0.05774692 0.1926369  0.49359822 0.10262781 0.10738047 0.0460096 ]
 [0.21298195 0.02826484 0.1813868  0.06380058 0.06848615 0.44507968]
 [0.01364106 0.16782394 0.08621352 0.22500433 0.19081964 0.31649753]
 [0.02917767 0.15526605 0.6310118  0.11471876 0.05473462 0.0150911 ]
 [0.03684716 0.15286008 0.11792535 0.47401306 0.05833342 0.160021  ]
 [0.32859987 0.17415446 0.07394216 0.22221863 0.07559296 0.12549189]
 [0.02662764 0.5529567  0.06995299 0.02131662 0.08664025 0.2425058 ]
 [0.10253917 0.10178788 0.21553555 0.12878521 0.3788466  0.07250563]]
pred: [2 3 2 5 5 2 3 0 1 4]
label: [3 4 3 0 4 0 3 2 1 4]
top-1-6 acc: [0.30000001192092896, 0.4000000059604645, 0.6000000238418579, 0.800000011920929, 0.8999999761581421, 1.0]