在tensorflow.python.data.ops.dataset_ops.BatchDataset中找到最大值

Question

假设下面的代码如下：

import tensorflow as tf
import numpy as np
 
simple_features = np.array([
         [1, 1, 1],
         [2, 2, 2],
         [3, 3, 3],
         [4, 4, 4],
         [5, 5, 5],
         [6, 6, 6],
         [7, 7, 7],
         [8, 8, 8],
         [9, 9, 9],
         [10, 10, 10],
         [11, 11, 11],
         [12, 12, 12],
])
 
simple_labels = np.array([
         [-1, -1],
         [-2, -2],
         [-3, -3],
         [-4, -4],
         [-5, -5],
         [-6, -6],
         [-7, -7],
         [-8, -8],
         [-9, -9],
         [-10, -10],
         [-11, -11],
         [-12, -12],
])
 
def print_dataset(ds):
    for inputs, targets in ds:
        print("---Batch---")
        print("Feature:", inputs.numpy())
        print("Label:", targets.numpy())
        print("")
 
    
ds = tf.keras.preprocessing.timeseries_dataset_from_array(simple_features, simple_labels, sequence_length=4, batch_size=32)
print_dataset(ds)

我想从每个simple_feature及其相应的simple_label中提取最大值。提取最大值后，我想将该值添加到simple_feature及其相应的simple_label中。例如，第一个simple_feature给我[1,1,1]，它相应的标签给我[-1,-1]。最大值为1，然后将1添加到[1,1,1]和[-1,-1]中，然后得到[2,2,2]和[0,0]。最后的数据集应保留为tensorflow.python.data.ops.dataset_ops.BatchDataset。

Answer 1

您可以使用tf.data.Dataset.from_tensor_slices和tf.data.Dataset.map解决问题。

import tensorflow as tf
import numpy as np
 
simple_features = np.array([
         [1, 1, 1],
         [2, 2, 2],
         [3, 3, 3],
         [4, 4, 4],
         [5, 5, 5],
         [6, 6, 6],
         [7, 7, 7],
         [8, 8, 8],
         [9, 9, 9],
         [10, 10, 10],
         [11, 11, 11],
         [12, 12, 12],
])
 
simple_labels = np.array([
         [-1, -1],
         [-2, -2],
         [-3, -3],
         [-4, -4],
         [-5, -5],
         [-6, -6],
         [-7, -7],
         [-8, -8],
         [-9, -9],
         [-10, -10],
         [-11, -11],
         [-12, -12],
])
 
def print_dataset(ds):
    for inputs, targets in ds:
        print("---Batch---")
        print("Feature: \n", inputs.numpy())
        print("Label: \n", targets.numpy())
        print("")

def map_max_values(x, y):
  max_values = tf.reduce_max(x, axis=1)
  temp_x = tf.reshape(tf.repeat(max_values, repeats=x.shape[1]), shape=(x.shape[0], x.shape[1]))
  temp_y = tf.reshape(tf.repeat(max_values, repeats=y.shape[1]), shape=(y.shape[0], y.shape[1]))

  x = x + temp_x
  y = y + temp_y
  return x, y

batch_size = 4
ds = tf.data.Dataset.from_tensor_slices((simple_features,simple_labels)).batch(batch_size, drop_remainder=True)
ds = ds.map(map_max_values)
print_dataset(ds)

---Batch---
Feature: 
 [[2 2 2]
 [4 4 4]
 [6 6 6]
 [8 8 8]]
Label: 
 [[0 0]
 [0 0]
 [0 0]
 [0 0]]

---Batch---
Feature: 
 [[10 10 10]
 [12 12 12]
 [14 14 14]
 [16 16 16]]
Label: 
 [[0 0]
 [0 0]
 [0 0]
 [0 0]]

---Batch---
Feature: 
 [[18 18 18]
 [20 20 20]
 [22 22 22]
 [24 24 24]]
Label: 
 [[0 0]
 [0 0]
 [0 0]
 [0 0]]

或者，如果您真的想使用tf.keras.preprocessing.timeseries_dataset_from_array，请尝试如下：

def map_max_values(x, y):
  max_values = tf.reduce_max(x, axis=2)
  temp_x = tf.reshape(tf.repeat(max_values, repeats=tf.shape(x)[2], axis=1), shape=tf.shape(x))
  temp_y = tf.reshape(tf.repeat(tf.expand_dims(max_values[:, 0], axis=1), repeats=tf.shape(y)[1], axis=1), shape=tf.shape(y))

  x = x + temp_x
  y = y + temp_y
  return x, y

ds = tf.keras.preprocessing.timeseries_dataset_from_array(simple_features, simple_labels, sequence_length=4, batch_size=32)
ds = ds.map(map_max_values)
print_dataset(ds)