tensorflow cifar-10评估示例softmax输出

Question

我想查看每个测试图像的概率，所以我修改了代码（cifar10_eval.py），如下

def eval_once(saver, summary_writer, logits, labels, top_k_op, summary_op):
...............
      while step < num_iter and not coord.should_stop():
        result1, result2  = sess.run([logits, labels])
        print('Step:', step, 'result',result1, 'Label:', result2)
...............

我像这样运行python代码。

# python cifar10_eval.py --batch_size=1 --run_once=True

屏幕结果如下

Step: 0 result [[ 0.01539493 -0.00109618 -0.00364288 -0.00898853 -0.00086198  0.00587899  0.00981337 -0.00785329 -0.00282823 -0.00171288]] Label: [4]
Step: 1 result [[ 0.01539471 -0.00109601 -0.00364273 -0.00898863 -0.00086192  0.005879  0.00981339 -0.00785322 -0.00282811 -0.00171296]] Label: [7]
Step: 2 result [[ 0.01539475 -0.00109617 -0.00364274 -0.00898876 -0.00086183  0.00587886  0.00981328 -0.00785333 -0.00282814 -0.00171295]] Label: [8]
Step: 3 result [[ 0.01539472 -0.00109597 -0.00364275 -0.0089886  -0.00086183  0.00587902  0.00981344 -0.00785326 -0.00282817 -0.00171299]] Label: [4]
Step: 4 result [[ 0.01539488 -0.00109631 -0.00364294 -0.00898863 -0.00086199  0.00587896  0.00981327 -0.00785329 -0.00282809 -0.00171307]] Label: [0]
Step: 5 result [[ 0.01539478 -0.00109607 -0.00364292 -0.00898858 -0.00086194  0.00587904  0.00981335 -0.0078533  -0.00282818 -0.00171321]] Label: [4]
Step: 6 result [[ 0.01539493 -0.00109627 -0.00364277 -0.00898873 -0.0008618   0.00587892  0.00981339 -0.00785325 -0.00282807 -0.00171289]] Label: [9]
Step: 7 result [[ 0.01539504 -0.00109619 -0.0036429  -0.00898865 -0.00086194  0.00587894  0.0098133  -0.00785331 -0.00282818 -0.00171294]] Label: [4]
Step: 8 result [[ 0.01539493 -0.00109627 -0.00364286 -0.00898867 -0.00086183  0.00587899  0.00981332 -0.00785329 -0.00282825 -0.00171283]] Label: [8]
Step: 9 result [[ 0.01539495 -0.00109617 -0.00364286 -0.00898852 -0.00086186  0.0058789  0.00981337 -0.00785326 -0.00282827 -0.00171287]] Label: [9]

Label的值看起来不错，但logits的输出似乎是相同的值！为什么？有人能告诉我原因吗？

这是新的cifar10_eval.py源代码。

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datetime import datetime
import math
import time
import numpy as np
import tensorflow as tf
#from tensorflow.models.image.cifar10 import cifar10
import cifar10
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string('eval_dir', '/tmp/cifar10_eval',
                           """Directory where to write event logs.""")
tf.app.flags.DEFINE_string('eval_data', 'test',
                           """Either 'test' or 'train_eval'.""")
tf.app.flags.DEFINE_string('checkpoint_dir', '/tmp/cifar10_train',
                           """Directory where to read model checkpoints.""")
tf.app.flags.DEFINE_integer('eval_interval_secs', 60 * 5,
                            """How often to run the eval.""")
tf.app.flags.DEFINE_integer('num_examples', 10000,
                            """Number of examples to run.""")
tf.app.flags.DEFINE_boolean('run_once', True,
                         """Whether to run eval only once.""")
def eval_once(saver, summary_writer, logits, labels, top_k_op, summary_op):
  with tf.Session() as sess:
    ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
    if ckpt and ckpt.model_checkpoint_path:
      saver.restore(sess, ckpt.model_checkpoint_path)
      global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
    else:
      print('No checkpoint file found')
      return
    # Start the queue runners.
    coord = tf.train.Coordinator()
    try:
      threads = []
      for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
        threads.extend(qr.create_threads(sess, coord=coord, daemon=True,
                                         start=True))
      #num_iter = int(math.ceil(FLAGS.num_examples / FLAGS.batch_size))
      #total_sample_count = num_iter * FLAGS.batch_size
      num_iter = FLAGS.num_examples
      total_sample_count =  FLAGS.num_examples
      print (num_iter, FLAGS.batch_size, total_sample_count)
      true_count = 0  # Counts the number of correct predictions.
      step = 0
      time.sleep(1)
      while step < num_iter and not coord.should_stop():
        result1, result2  = sess.run([logits, labels])
        #label = sess.run(labels)
        print('Step:', step, 'result',result1, 'Label:', result2)
        step += 1
      precision = true_count / step
      print('Summary -- Step:', step, 'Accurcy:',true_count * 100.0 / step * 1.0, )
      print('%s: total:%d true:%d precision @ 1 = %.3f' % (datetime.now(), total_sample_count, true_count, precision))
    except Exception as e:  # pylint: disable=broad-except
      coord.request_stop(e)
    coord.request_stop()
    coord.join(threads, stop_grace_period_secs=10)

def evaluate():
  """Eval CIFAR-10 for a number of steps."""
  with tf.Graph().as_default():
    # Get images and labels for CIFAR-10.
    eval_data = FLAGS.eval_data == 'test'
    images, labels = cifar10.inputs(eval_data=eval_data, )
    # Build a Graph that computes the logits predictions from the
    # inference model. logits is softmax
    logits = cifar10.inference(images)
    # Calculate predictions.
    top_k_op = tf.nn.in_top_k(logits, labels, 1)
    # Restore the moving average version of the learned variables for eval.
    variable_averages = tf.train.ExponentialMovingAverage(
        cifar10.MOVING_AVERAGE_DECAY)
    variables_to_restore = variable_averages.variables_to_restore()
    saver = tf.train.Saver(variables_to_restore)
    # Build the summary operation based on the TF collection of Summaries.
    summary_op = tf.merge_all_summaries()
    graph_def = tf.get_default_graph().as_graph_def()
    summary_writer = tf.train.SummaryWriter(FLAGS.eval_dir,
                                            graph_def=graph_def)
    while True:
      eval_once(saver, summary_writer, logits, labels,top_k_op, summary_op)
      if FLAGS.run_once:
        break
      time.sleep(FLAGS.eval_interval_secs)

def main(argv=None):  # pylint: disable=unused-argument
  cifar10.maybe_download_and_extract()
  if tf.gfile.Exists(FLAGS.eval_dir):
    tf.gfile.DeleteRecursively(FLAGS.eval_dir)
  tf.gfile.MakeDirs(FLAGS.eval_dir)
  print('Evaluate Start')
  evaluate()

if __name__ == '__main__':
  tf.app.run()

Answer 1

我已经用1k步（准确率10%左右）进行了训练。但在我用100k步（准确率86%左右）训练后，结果是非常好的

.................
Step: 9991 result: [[  1.30259633   0.71064955  -2.6035285   -1.30183697  -4.1291523   -3.00246906   0.30873945  -4.02916574  13.05054665  -0.42556083]] Label: 8
Step: 9992 result: [[-1.05670786 -1.86572766  0.28350741  1.78929067  0.03841069  1.23079467    2.97172165 -1.18722486 -1.17184007 -1.02505279]] Label: 6
Step: 9993 result: [[  1.50454926   2.34122658  -3.45632267  -0.55308843  -4.35214806    -2.28931832  -1.74908364  -4.71527719  11.44062901   1.72015083]] Label: 8
Step: 9994 result: [[ 1.96891284 -2.57139373  0.29864013  1.30923986  1.72708285  0.95571399   -0.49331608  0.49454236 -2.26134181 -1.39561605]] Label: 0
Step: 9995 result: [[-0.65523863  1.58577776  0.13226865  1.43122363 -2.34669352  0.18927786   -2.51019335 -1.70729315 -0.21297894  4.06098557]] Label: 9
Step: 9996 result: [[-2.17944765 -3.22895575  2.29571438  2.63287306  0.46685112  4.42715979   -0.76104468  2.39603662 -3.21783161 -2.8433671 ]] Label: 2
Step: 9997 result: [[  4.26957560e+00   1.95574760e-03   1.91038296e-01  -8.00723195e-01    -2.36319876e+00  -2.12906289e+00  -3.35138845e+00   7.97132492e-01   6.60009801e-01   2.73786736e+00]] Label: 0
Step: 9998 result: [[ 0.42694128 -2.07150149  0.47749567  2.62247086  1.11608386  3.05186462   -0.19805858  0.03386561 -2.87092948 -2.59781456]] Label: 5
Step: 9999 result: [[ 0.23629765 -3.21540785  1.01075113  0.46802399  3.44423246  0.25743011    4.71304989 -1.12128389 -3.07727337 -2.7076664 ]] Label: 6
2016-04-09 00:32:49.861650 Total:10000 True:8631: precision @ 1 = 0.863