SlimをTensorflow0.9.0で無理やり動かしてみた

Slimとは

TF-Slim is a lightweight library for defining, training and evaluating complex models in TensorFlow. Components of tf-slim can be freely mixed with native tensorflow, as well as other frameworks, such as tf.contrib.learn.

（意訳）
TF-SlimはTensorflowで複雑なモデルを定義・訓練・評価するための軽量ライブラリで、生のTensorflowやTF-Learn（tf.contrib.learn）と混ぜても安全。

Githubのリポジトリで開発が進んでいる。

r0.10のブランチからTensorflowをビルドしたら使えそうだけど、0.9.0でも一部機能は実装されていたので無理やり使ってみた。
題材はCIFAR10

インポート

from tensorflow.contrib import slim

データ読み込み

0.10だとslim.dataに色々なメソッドが実装されていそうだったけど、0.9ではslim.dataが存在しないので諦める。

モデル定義

conv*2, pool*1, norm*1のセットを2回→fc*3

with slim.arg_scope([slim.layers.conv2d, slim.layers.fully_connected],
                    activation_fn=tf.nn.relu,
                    weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                    weights_regularizer=slim.regularizers.l2_regularizer(0.0005)):
  with slim.arg_scope([slim.layers.max_pool2d], padding='SAME'):
      net = slim.layers.repeat(inputs, 2, slim.layers.conv2d, 64, [3, 3], scope='conv1')
      net = slim.layers.max_pool2d(net, [3, 3], scope='pool1')
      net = tf.nn.lrn(net, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm1')
      net = slim.layers.repeat(net, 2, slim.layers.conv2d, 64, [5, 5], scope='conv2')
      net = slim.layers.max_pool2d(net, [3, 3], scope='pool2')
      net = tf.nn.lrn(net, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm2')
      net = slim.layers.flatten(net, scope='flatten')
      net = slim.layers.stack(net, slim.layers.fully_connected, [384, 192], scope='fc1')
      net = slim.layers.fully_connected(net, NUM_CLASSES, activation_fn=None, scope='fc2')

repeatは0.9に含まれていないけど、これだけはないと不便なので0.10からコピー。

from tensorflow.python.framework import ops
from tensorflow.python.ops import variable_scope

def repeat(inputs, repetitions, layer, *args, **kwargs):
  scope = kwargs.pop('scope', None)
  with variable_scope.variable_op_scope([inputs], scope, 'Repeat'):
    inputs = ops.convert_to_tensor(inputs)
    if scope is None:
      if hasattr(layer, '__name__'):
        scope = layer.__name__
      elif hasattr(layer, 'func') and hasattr(layer.func, '__name__'):
        scope = layer.func.__name__  # In case layer is a functools.partial.
      else:
        scope = 'repeat'
    outputs = inputs
    for i in range(repetitions):
      kwargs['scope'] = scope + '_' + str(i+1)
      outputs = layer(outputs, *args, **kwargs)
    return outputs

slim.layers.repeat = repeat

損失関数

loss = slim.losses.softmax_cross_entropy(predictions, labels)

最適化

optimizer = tf.train.GradientDescentOptimizer(learning_rate=LEARNING_RATE)

学習

train_op = slim.learning.create_train_op(loss, optimizer)

result = slim.learning.train(train_op,
                             FLAGS.log_dir,
                             number_of_steps=1000,
                             save_summaries_secs=300,
                             save_interval_secs=600)

評価

slim.evaluationに用意されているが、うまく使えなかったので割愛。

感想

READMEに書いてあることも信用できないし、0.9だと未実装なものや0.10ではメソッド名が変わっているものもあってコードを直接見ることになるので、素直に0.10を待ったほうがいいかもしれない。
ウェイト・バイアスのリストアとモデル評価のところがうまくいかなかったので、誰かできたら教えて欲しい……
モデルの定義だけは利用すると簡潔でいい。変数のリストア方法がわからなかったけど……

コード

cifar10.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf

from tensorflow.contrib import slim

import os, sys, tarfile
from six.moves import urllib, range

from tensorflow.models.image.cifar10 import cifar10_input

FLAGS = tf.app.flags.FLAGS

tf.app.flags.DEFINE_string('tmp_data_dir', '/tmp/cifar10_data',
                           """Path to the CIFAR-10 data directory.""")
tf.app.flags.DEFINE_string('log_dir', '/tmp/cifar10_train',
                           """Directory where to write event logs """
                           """and checkpoint.""")

DATA_URL = 'http://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz'
NUM_CLASSES = cifar10_input.NUM_CLASSES
LEARNING_RATE = 0.1

def main(argv=None):  # pylint: disable=unused-argument
    maybe_download_and_extract()

    if tf.gfile.Exists(FLAGS.log_dir):
        tf.gfile.DeleteRecursively(FLAGS.log_dir)
    tf.gfile.MakeDirs(FLAGS.log_dir)

    with tf.Graph().as_default():
        images, labels = distorted_inputs()
        labels = slim.layers.one_hot_encoding(labels, NUM_CLASSES)

        predictions = my_model(images)
        loss = slim.losses.softmax_cross_entropy(predictions, labels)

        tf.scalar_summary('loss', loss)
        optimizer = tf.train.GradientDescentOptimizer(learning_rate=LEARNING_RATE)
        train_op = slim.learning.create_train_op(loss, optimizer)

        result = slim.learning.train(train_op,
                                     FLAGS.log_dir,
                                     number_of_steps=1000,
                                     save_summaries_secs=300,
                                     save_interval_secs=600)
        print('loss: %f' % result)

def my_model(inputs):
    with slim.arg_scope([slim.layers.conv2d, slim.layers.fully_connected],
                        activation_fn=tf.nn.relu,
                        weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                        weights_regularizer=slim.regularizers.l2_regularizer(0.0005)):
        with slim.arg_scope([slim.layers.max_pool2d], padding='SAME'):
            net = slim.layers.repeat(inputs, 2, slim.layers.conv2d, 64, [3, 3], scope='conv1')
            net = slim.layers.max_pool2d(net, [3, 3], scope='pool1')
            net = tf.nn.lrn(net, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm1')
            net = slim.layers.repeat(net, 2, slim.layers.conv2d, 64, [5, 5], scope='conv2')
            net = slim.layers.max_pool2d(net, [3, 3], scope='pool2')
            net = tf.nn.lrn(net, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm2')
            net = slim.layers.flatten(net, scope='flatten')
            net = slim.layers.stack(net, slim.layers.fully_connected, [384, 192], scope='fc1')
            net = slim.layers.fully_connected(net, NUM_CLASSES, activation_fn=None, scope='fc2')
    return net

def distorted_inputs():
    if not FLAGS.tmp_data_dir:
        raise ValueError('Please supply a tmp_data_dir')
    data_dir = os.path.join(FLAGS.tmp_data_dir, 'cifar-10-batches-bin')
    images, labels = cifar10_input.distorted_inputs(data_dir=data_dir,
                                                    batch_size=FLAGS.batch_size)
    return images, labels

def maybe_download_and_extract():
    """Download and extract the tarball from Alex's website."""
    dest_directory = FLAGS.tmp_data_dir
    if not os.path.exists(dest_directory):
        os.makedirs(dest_directory)
    filename = DATA_URL.split('/')[-1]
    filepath = os.path.join(dest_directory, filename)
    if not os.path.exists(filepath):
        def _progress(count, block_size, total_size):
            sys.stdout.write('\r>> Downloading %s %.1f%%' % (filename,
                             float(count * block_size) / float(total_size) * 100.0))
            sys.stdout.flush()
        filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, _progress)
        print()
        statinfo = os.stat(filepath)
        print('Successfully downloaded', filename, statinfo.st_size, 'bytes.')
        tarfile.open(filepath, 'r:gz').extractall(dest_directory)

# define slim.layers.repeat
from tensorflow.python.framework import ops
from tensorflow.python.ops import variable_scope

def repeat(inputs, repetitions, layer, *args, **kwargs):
    scope = kwargs.pop('scope', None)
    with variable_scope.variable_op_scope([inputs], scope, 'Repeat'):
        inputs = ops.convert_to_tensor(inputs)
        if scope is None:
          if hasattr(layer, '__name__'):
            scope = layer.__name__
          elif hasattr(layer, 'func') and hasattr(layer.func, '__name__'):
            scope = layer.func.__name__  # In case layer is a functools.partial.
          else:
            scope = 'repeat'
        outputs = inputs
        for i in range(repetitions):
          kwargs['scope'] = scope + '_' + str(i+1)
          outputs = layer(outputs, *args, **kwargs)
        return outputs

slim.layers.repeat = repeat

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