python - tf.metrics.accuracy 和手写准确度函数给出不同的结果

Question

我正在尝试了解 tf.metrics.accuracy 的工作原理。我想比较下面给出的函数的批量准确度结果

with tf.name_scope('Accuracy1'):
        correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(y, 1))
        accuracy1 = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name="accuracy")

与

with tf.name_scope('Accuracy2'):
        accuracy2, accuracy_op = tf.metrics.accuracy(labels=tf.argmax(y, 1), predictions=tf.argmax(predictions, 1))

下面提供了最小的工作示例:

import numpy as np 
import pandas as pd 
import tensorflow as tf
import math

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

num_steps=28
num_inputs = 28
num_classes = 10
num_neurons = 128
num_layers = 3
batch_size = 500

graph = tf.Graph()
with graph.as_default():
    with tf.name_scope("graph_inputs"):
        X = tf.placeholder(tf.float32, [None, num_steps, num_inputs], name='input_placeholder')
        y = tf.placeholder(tf.float32, [None, num_classes], name='labels_placeholder')
       output_keep_prob = tf.placeholder_with_default(1.0, shape=(), name ="output_dropout")

def build_lstm_cell(num_neurons, output_keep_prob):
    """Returns a dropout-wrapped LSTM-cell.
    See https://stackoverflow.com/a/44882273/2628369 for why this local function is necessary.
    Returns:
    tf.contrib.rnn.DropoutWrapper: The dropout-wrapped LSTM cell.
    """
    initializer = tf.contrib.layers.xavier_initializer()
    lstm_cell = tf.contrib.rnn.LSTMCell(num_units=num_neurons, initializer=initializer, forget_bias=1.0, state_is_tuple=True, name='LSTM_cell')
    lstm_cell_drop = tf.contrib.rnn.DropoutWrapper(lstm_cell, output_keep_prob=output_keep_prob)
    return lstm_cell_drop

with tf.name_scope("LSTM"):
    with tf.name_scope("Cell"):
        multi_layer_cell = tf.contrib.rnn.MultiRNNCell([build_lstm_cell(num_neurons, output_keep_prob) for _ in range(num_layers)], state_is_tuple=True)
    with tf.name_scope("Model"):
        outputs, states = tf.nn.dynamic_rnn(cell=multi_layer_cell, inputs=X, swap_memory=False, time_major = False, dtype=tf.float32)#[Batch_size, time_steps, num_neurons]
    with tf.name_scope("Graph_Outputs"):
        outputs = tf.transpose(outputs, [1, 0, 2]) # [num_timesteps, batch_size, num_neurons]
        outputs = tf.gather(outputs, int(outputs.get_shape()[0]) - 1) # [batch_size, num_neurons]
    with tf.variable_scope('Softmax'):
        logits =  tf.layers.dense(inputs = outputs, units = num_classes, name="logits") #[Batch_size, num_classes]
    with tf.name_scope('Predictions'):
        predictions = tf.nn.softmax(logits, name="predictions")  #[Batch_size, num_classes]
    with tf.name_scope('Accuracy1'):
        correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(y, 1))
        accuracy1 = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name="accuracy")
    with tf.name_scope('Accuracy2'):
        accuracy2, accuracy_op = tf.metrics.accuracy(labels=tf.argmax(y, 1), predictions=tf.argmax(predictions, 1))
    with tf.name_scope('Loss'):
        xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y)
        loss = tf.reduce_mean(xentropy, name="loss")
    with tf.name_scope('Train'):
        optimizer= tf.train.AdamOptimizer(learning_rate=0.0001)
        trainer=optimizer.minimize(loss, name="training_op")

with tf.Session(graph = graph) as sess:
    tf.global_variables_initializer().run()
    total_batch = mnist.train.num_examples // batch_size
    for batch in range(total_batch):
        tf.local_variables_initializer().run()
        xBatch, yBatch = mnist.train.next_batch(batch_size)
        xBatch = xBatch.reshape((batch_size, num_steps, num_inputs))
        sess.run(trainer, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})
        miniBatchAccuracy1 = sess.run(accuracy1, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})
        print('[hand-written] Batch {} accuracy: {}'.format(batch, miniBatchAccuracy1))
        accuracy_op_val = sess.run(accuracy_op, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})
        miniBatchAccuracy2 = sess.run(accuracy2)
        print("[tf.metrics.accuracy] Batch {} accuracy: {}".format(batch, miniBatchAccuracy2))
    sess.close()

我使用这两种方法打印每个批处理的准确度值，它们是不同的。结果不应该是一样的吗？

[hand-written] Batch 0 accuracy: 0.09600000083446503
[tf.metrics.accuracy] Batch 0 accuracy: 0.09399999678134918

[hand-written] Batch 1 accuracy: 0.1120000034570694
[tf.metrics.accuracy] Batch 1 accuracy: 0.07800000160932541

[hand-written] Batch 2 accuracy: 0.10199999809265137
[tf.metrics.accuracy] Batch 2 accuracy: 0.09600000083446503

[hand-written] Batch 3 accuracy: 0.12999999523162842
[tf.metrics.accuracy] Batch 3 accuracy: 0.12800000607967377

[hand-written] Batch 4 accuracy: 0.1379999965429306
[tf.metrics.accuracy] Batch 4 accuracy: 0.10199999809265137

[hand-written] Batch 5 accuracy: 0.16200000047683716
[tf.metrics.accuracy] Batch 5 accuracy: 0.1340000033378601

[hand-written] Batch 6 accuracy: 0.1340000033378601
[tf.metrics.accuracy] Batch 6 accuracy: 0.12600000202655792

[hand-written] Batch 7 accuracy: 0.12999999523162842
[tf.metrics.accuracy] Batch 7 accuracy: 0.16200000047683716
...
...
...
...

Answer 1

在测量这两种情况的准确性时，您将辍学率传递为0.5。这就是它给出两个不同值的原因。将 dropout 值设置为 1.0，您应该会看到这两种情况的类似值。