2018년 8월 21일 화요일

Session-based tensorflow training에 LMS 적용한 MNIST python code

이 MNIST training을 위한 tensorflow python code는 원래 /opt/DL/tensorflow/lib/python3.6/site-packages/tensorflow/examples/tutorials/mnist/mnist_deep.py 에 LMS를 적용한 것입니다.  보시다시피 graph라는 단어가 나오기 때문에 Session-based tensorflow training이고, 그에 따라 LMS가 적용되어 있습니다.  

실제 LMS 구현을 위한 부분은 굵은 파란색으로 표시했습니다.  의외로 간단하다는 것을 보실 수 있습니다.  해당 부분들을 제거하면 그냥 LMS 없는 평범한 MNIST training code가 됩니다.

이 example code도 PowerAI 5.2를 설치하면 딸려오는 /opt/DL/tensorflow/lib/python3.6/site-packages/tensorflow/contrib/lms/examples/mnist_deep_lms.py 을 그대로 가져다 놓은 것입니다.

실제 수행해보면 다음과 같이 동작하며, 27개의 tensor가 host 서버의 RAM으로 swap-out/in 되는 것을 보실 수 있습니다.



[bsyu@p57a22 examples]$ CUDA_VISIBLE_DEVICES=3 python mnist_deep_lms.py
...
INFO:tensorflow:[LMS][0] Editing model for LMS
INFO:tensorflow:[LMS][0] n_tensors: all tensors
INFO:tensorflow:[LMS][0] lb: 3
INFO:tensorflow:[LMS][0] Edited model is valid and logically equivalent to the original one
INFO:tensorflow:[LMS][0] Added 53 ops into the model
INFO:tensorflow:[LMS][0] Editing model for LMS, took: 179.62193489074707 ms
INFO:tensorflow:[LMS][0] 27 tensors will be swapped out(in) to(from) the host
Saving graph to: /tmp/tmpz5ozc9jr
2018-08-20 21:37:47.954552: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties:
....
step 19900, training accuracy 1
test accuracy 0.9918




[bsyu@p57a22 doc]$ cd /opt/DL/tensorflow/lib/python3.6/site-packages/tensorflow/contrib/lms/examples/


[bsyu@p57a22 examples]$ source /opt/DL/tensorflow/bin/tensorflow-activate

[bsyu@p57a22 examples]$ cat /opt/DL/tensorflow/lib/python3.6/site-packages/tensorflow/contrib/lms/examples/mnist_deep_lms.py

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""A deep MNIST classifier using convolutional layers.

See extensive documentation at
https://www.tensorflow.org/get_started/mnist/pros
"""
# Disable linter warnings to maintain consistency with tutorial.
# pylint: disable=invalid-name
# pylint: disable=g-bad-import-order

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

import argparse
import sys
import tempfile

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf

tf.logging.set_verbosity(tf.logging.INFO)
FLAGS = None


def deepnn(x):
  """deepnn builds the graph for a deep net for classifying digits.

  Args:
    x: an input tensor with the dimensions (N_examples, 784), where 784 is the
    number of pixels in a standard MNIST image.

  Returns:
    A tuple (y, keep_prob). y is a tensor of shape (N_examples, 10), with values
    equal to the logits of classifying the digit into one of 10 classes (the
    digits 0-9). keep_prob is a scalar placeholder for the probability of
    dropout.
  """
  # Reshape to use within a convolutional neural net.
  # Last dimension is for "features" - there is only one here, since images are
  # grayscale -- it would be 3 for an RGB image, 4 for RGBA, etc.
  with tf.name_scope('reshape'):
    x_image = tf.reshape(x, [-1, 28, 28, 1])

  # First convolutional layer - maps one grayscale image to 32 feature maps.
  with tf.name_scope('conv1'):
    W_conv1 = weight_variable([5, 5, 1, 32])
    b_conv1 = bias_variable([32])
    h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)

  # Pooling layer - downsamples by 2X.
  with tf.name_scope('pool1'):
    h_pool1 = max_pool_2x2(h_conv1)

  # Second convolutional layer -- maps 32 feature maps to 64.
  with tf.name_scope('conv2'):
    W_conv2 = weight_variable([5, 5, 32, 64])
    b_conv2 = bias_variable([64])
    h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)

  # Second pooling layer.
  with tf.name_scope('pool2'):
    h_pool2 = max_pool_2x2(h_conv2)

  # Fully connected layer 1 -- after 2 round of downsampling, our 28x28 image
  # is down to 7x7x64 feature maps -- maps this to 1024 features.
  with tf.name_scope('fc1'):
    W_fc1 = weight_variable([7 * 7 * 64, 1024])
    b_fc1 = bias_variable([1024])

    h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
    h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

  # Dropout - controls the complexity of the model, prevents co-adaptation of
  # features.
  with tf.name_scope('dropout'):
    keep_prob = tf.placeholder(tf.float32)
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  # Map the 1024 features to 10 classes, one for each digit
  with tf.name_scope('fc2'):
    W_fc2 = weight_variable([1024, 10])
    b_fc2 = bias_variable([10])

    y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
  return y_conv, keep_prob


def conv2d(x, W):
  """conv2d returns a 2d convolution layer with full stride."""
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
  """max_pool_2x2 downsamples a feature map by 2X."""
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')


def weight_variable(shape):
  """weight_variable generates a weight variable of a given shape."""
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)


def bias_variable(shape):
  """bias_variable generates a bias variable of a given shape."""
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)


def main(_):
  # Import data
  mnist = input_data.read_data_sets(FLAGS.data_dir)

  # Create the model
  x = tf.placeholder(tf.float32, [None, 784])

  # Define loss and optimizer
  y_ = tf.placeholder(tf.int64, [None])

  # Build the graph for the deep net
  y_conv, keep_prob = deepnn(x)

  with tf.name_scope('loss'):
    cross_entropy = tf.losses.sparse_softmax_cross_entropy(
        labels=y_, logits=y_conv)
  cross_entropy = tf.reduce_mean(cross_entropy)

  with tf.name_scope('adam_optimizer'):
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

  with tf.name_scope('accuracy'):
    correct_prediction = tf.equal(tf.argmax(y_conv, 1), y_)
    correct_prediction = tf.cast(correct_prediction, tf.float32)
  accuracy = tf.reduce_mean(correct_prediction)

  # Enable Large Model Support
  from tensorflow.contrib.lms import LMS
  lms_model = LMS({'adam_optimizer'},
                  excl_scopes = {'loss', 'accuracy', 'dropout'},
                  lb=3)
  lms_model.run(tf.get_default_graph())

  graph_location = tempfile.mkdtemp()
  print('Saving graph to: %s' % graph_location)
  train_writer = tf.summary.FileWriter(graph_location)
  train_writer.add_graph(tf.get_default_graph())

  with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(20000):    
      batch = mnist.train.next_batch(50)   
      if i % 100 == 0:
        train_accuracy = accuracy.eval(feed_dict={
            x: batch[0], y_: batch[1], keep_prob: 1.0})
        print('step %d, training accuracy %g' % (i, train_accuracy))
      train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    print('test accuracy %g' % accuracy.eval(feed_dict={
        x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

if __name__ == '__main__':
  parser = argparse.ArgumentParser()
  parser.add_argument('--data_dir', type=str,
                      default='/tmp/tensorflow/mnist/input_data',
                      help='Directory for storing input data')
  FLAGS, unparsed = parser.parse_known_args()

  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

댓글 없음:

댓글 쓰기