如何在特定数量的列中找到 tensorflow 数据集批次中的最大值？

Posted 2023-02-16

【中文标题】如何在特定数量的列中找到 tensorflow 数据集批次中的最大值？

【英文标题】：How do I find the max value in a tensorflow dataset batch across a specific number of columns?

【发布时间】：2021-12-19 17:40:03

【问题描述】：

假设下面的代码：

import tensorflow as tf
import numpy as np
 
simple_data_samples = np.array([
         [1, 1, 1, 7, -1],
         [2, -2, 2, -2, -2],
         [3, 3, 3, -3, -3],
         [-4, 4, 4, -4, -4],
         [5, 5, 5, -5, -5],
         [6, 6, 6, -4, -6],
         [7, 7, 8, -7, -7],
         [8, 8, 8, -8, -8],
         [9, 4, 9, -9, -9],
         [10, 10, 10, -10, -10],
         [11, 5, 11, -11, -11],
         [12, 12, 12, -12, -12],
])


def print_dataset(ds):
    for inputs, targets in ds:
        print("---Batch---")
        print("Feature:", inputs.numpy())
        print("Label:", targets.numpy())
        print("")
 
    
def timeseries_dataset_multistep_combined(features, label_slice, input_sequence_length, output_sequence_length, sequence_stride, batch_size):
    feature_ds = tf.keras.preprocessing.timeseries_dataset_from_array(features, None, sequence_length=input_sequence_length + output_sequence_length, sequence_stride=sequence_stride ,batch_size=batch_size, shuffle=False)
     
    def split_feature_label(x):
        return x[:, :input_sequence_length, :]+ tf.reduce_max(x[:,:,:],axis=1), x[:, input_sequence_length:, label_slice]+ tf.reduce_max(x[:,:,:],axis=1)
         
    feature_ds = feature_ds.map(split_feature_label)
     
    return feature_ds
 
ds = timeseries_dataset_multistep_combined(simple_data_samples, slice(None, None, None), input_sequence_length=4, output_sequence_length=2, sequence_stride=2, batch_size=1)
print_dataset(ds)

让我解释一下上面的代码是做什么的。它创建了许多特征和标签。然后它从每列中获取最大值并将其添加到列中的各个值中。比如这个特征及其对应的标签：

Feature: [[[ 1  1  1  7 -1]
  [ 2 -2  2 -2 -2]
  [ 3  3  3 -3 -3]
  [-4  4  4 -4 -4]]]
Label: [[[ 5  5  5 -5 -5]
  [ 6  6  6 -4 -6]]]

在每列中具有以下最大值：

6,6,6,7,-1

然后将最大值添加到相应的列中，然后得到最终输出：

[[ 7  7  7 14 -2]
  [ 8  4  8  4 -3]
  [ 9  9  9  3 -4]
  [ 2 10 10  2 -5]]]
Label: [[[11 11 11  1 -6]
  [12 12 12  2 -7]]]

我想从每个特征的前三列和最后两列及其对应的标签中提取最大值，而不是从每一列中提取最大值。提取后，我想将最大值添加到相应列中的每个值。例如，在上面的示例中，前三列的最大值为 6，后两列的最大值为 7。之后，前三列中的每个值将添加 6，后两列中的每个值添加 7。第一批的最终输出如下所示：

Feature: [[[ 7  7  7  14 6]
  [ 8 4  8 5 5]
  [ 9  9  9 4 4]
  [ 2  10  10 3 3]]]
Label: [[[ 11  11  11 2 2]
  [ 12  12 12 3 1]]]

有人知道如何从每批的前三列和后两列中提取最大值吗？

【参考方案1】：

像这样使用tf.tile 和tf.reduce_max 对你有用吗：

import tensorflow as tf
import numpy as np
 
simple_data_samples = np.array([
         [1, 1, 1, 7, -1],
         [2, -2, 2, -2, -2],
         [3, 3, 3, -3, -3],
         [-4, 4, 4, -4, -4],
         [5, 5, 5, -5, -5],
         [6, 6, 6, -4, -6],
         [7, 7, 8, -7, -7],
         [8, 8, 8, -8, -8],
         [9, 4, 9, -9, -9],
         [10, 10, 10, -10, -10],
         [11, 5, 11, -11, -11],
         [12, 12, 12, -12, -12],
])


def print_dataset(ds):
    for inputs, targets in ds:
        print("---Batch---")
        print("Feature:", inputs.numpy())
        print("Label:", targets.numpy())
        print("")
 
    
def timeseries_dataset_multistep_combined(features, label_slice, input_sequence_length, output_sequence_length, sequence_stride, batch_size):
    feature_ds = tf.keras.preprocessing.timeseries_dataset_from_array(features, None, sequence_length=input_sequence_length + output_sequence_length, sequence_stride=sequence_stride ,batch_size=batch_size, shuffle=False)
     
    def split_feature_label(x):
        reduced_first_max_columns = tf.reduce_max(x[:,:,:3], axis=1, keepdims=True) 
        reduced_last_max_columns = tf.reduce_max(x[:,:,3:], axis=1, keepdims=True)
        reduced_first_max_columns = tf.tile(tf.reduce_max(reduced_first_max_columns, axis=-1), [1, 3])
        reduced_last_max_columns = tf.tile(tf.reduce_max(reduced_last_max_columns, axis=-1), [1, 2])
        reduced_x = tf.expand_dims(tf.concat([reduced_first_max_columns, reduced_last_max_columns], axis=1), axis=0)
        
        return x[:, :input_sequence_length, :] + reduced_x, x[:, input_sequence_length:, label_slice] + reduced_x
         
    feature_ds = feature_ds.map(split_feature_label)
     
    return feature_ds
 
ds = timeseries_dataset_multistep_combined(simple_data_samples, slice(None, None, None), input_sequence_length=4, output_sequence_length=2, sequence_stride=2, batch_size=1)
print_dataset(ds)

---Batch---
Feature: [[[ 7  7  7 14  6]
  [ 8  4  8  5  5]
  [ 9  9  9  4  4]
  [ 2 10 10  3  3]]]
Label: [[[11 11 11  2  2]
  [12 12 12  3  1]]]

---Batch---
Feature: [[[11 11 11 -6 -6]
  [ 4 12 12 -7 -7]
  [13 13 13 -8 -8]
  [14 14 14 -7 -9]]]
Label: [[[ 15  15  16 -10 -10]
  [ 16  16  16 -11 -11]]]
...