Question

每一个人！从shuffle我得到图像和标签，带有conv的图像效果很好。

来自tfreocords的图片和标签

...
train_images, train_labels = shuffle(train_all_images, train_all_labels)
...

但是train_labels无效，如下所示：

numpy.sum(numpy.argmax(predictions, 1) ==  train_labels)

结果总是错误的，因为它根本无法从train_labels获取值。

以下是一些细节：

train_all_images, train_all_labels = read_and_decode("train")

train_images, train_labels = shuffle(train_all_images, train_all_labels)

......一些训练模式

optimizer = tf.train.MomentumOptimizer(learning_rate,
                                       0.9).minimize(loss,
                                                     global_step=batch)
train_prediction = tf.nn.softmax(logits)

with tf.Session() as sess:
    tf.global_variables_initializer().run()
    tf.train.start_queue_runners(sess)
    print('Initialized!')

    for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
        sess.run(optimizer)
        if step % EVAL_FREQUENCY == 0:              
            l, lr, predictions = sess.run([loss, learning_rate, train_prediction])

            print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
            print('Minibatch error: %.1f%%' % error_rate(predictions, train_labels))
            sys.stdout.flush()

def error_rate(predictions, labels):
    return 100.0 - ( 100.0 *
    numpy.sum(numpy.argmax(predictions, 1) == labels) /
    predictions.shape[0])

Answer 1

原因是你必须使用数不多的mothods而不是numpy，以下是适当的。

def correct_rate(out, labels):
  arg = tf.argmax(out, 1)
  arg = tf.cast(arg, tf.int32)
  correct_prediction = tf.equal(labels, arg)
  accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  return accuracy

accr = correct_rate(logits, train_labels)
print(sess.run(accr))

为什么来自tensorflow的shuffle的标签不起作用？

1 个答案: