mnist.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.
# ==============================================================================

"""Simple, end-to-end, LeNet-5-like convolutional MNIST model example.
This should achieve a test error of 0.7%. Please keep this model as simple and
linear as possible, it is meant as a tutorial for simple convolutional models.
Run with --self_test on the command line to execute a short self-test.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import gzip
import os
import sys
import time

import numpy
from six.moves import urllib
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf
import numpy as np

SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
WORK_DIRECTORY = 'data'
IMAGE_SIZE = 28
NUM_CHANNELS = 1
PIXEL_DEPTH = 255
NUM_LABELS = 10
VALIDATION_SIZE = 5000  # Size of the validation set.
SEED = 66478  # Set to None for random seed.
BATCH_SIZE = 64
NUM_EPOCHS = 10
EVAL_BATCH_SIZE = 64
EVAL_FREQUENCY = 100  # Number of steps between evaluations.
FINETUNE = False
PARTIALLY_FINETUNE=True


tf.app.flags.DEFINE_boolean("self_test", False, "True if running a self test.")
tf.app.flags.DEFINE_boolean('use_fp16', False,
                            "Use half floats instead of full floats if True.")
FLAGS = tf.app.flags.FLAGS


def data_type():
  """Return the type of the activations, weights, and placeholder variables."""
  if FLAGS.use_fp16:
    return tf.float16
  else:
    return tf.float32


def maybe_download(filename):
  """Download the data from Yann's website, unless it's already here."""
  if not tf.gfile.Exists(WORK_DIRECTORY):
    tf.gfile.MakeDirs(WORK_DIRECTORY)
  filepath = os.path.join(WORK_DIRECTORY, filename)
  if not tf.gfile.Exists(filepath):
    filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
    with tf.gfile.GFile(filepath) as f:
      size = f.size()
    print('Successfully downloaded', filename, size, 'bytes.')
  return filepath


def extract_data(filename, num_images):
  """Extract the images into a 4D tensor [image index, y, x, channels].
  Values are rescaled from [0, 255] down to [-0.5, 0.5].
  """
  print('Extracting', filename)
  with gzip.open(filename) as bytestream:
    bytestream.read(16)
    buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images * NUM_CHANNELS)
    data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32)
    data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
    data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS)
    return data


def extract_labels(filename, num_images):
  """Extract the labels into a vector of int64 label IDs."""
  print('Extracting', filename)
  with gzip.open(filename) as bytestream:
    bytestream.read(8)
    buf = bytestream.read(1 * num_images)
    labels = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.int64)
  return labels


def fake_data(num_images):
  """Generate a fake dataset that matches the dimensions of MNIST."""
  data = numpy.ndarray(
      shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
      dtype=numpy.float32)
  labels = numpy.zeros(shape=(num_images,), dtype=numpy.int64)
  for image in xrange(num_images):
    label = image % 2
    data[image, :, :, 0] = label - 0.5
    labels[image] = label
  return data, labels


def error_rate(predictions, labels):
  """Return the error rate based on dense predictions and sparse labels."""
  return 100.0 - (
      100.0 *
      numpy.sum(numpy.argmax(predictions, 1) == labels) /
      predictions.shape[0])

def orthogonal_initializer(scale = 1.1):
    ''' From Lasagne and Keras. Reference: Saxe et al., http://arxiv.org/abs/1312.6120
    '''
    print('Warning -- You have opted to use the orthogonal_initializer function')
    def _initializer(shape, dtype=tf.float32, partition_info=None):
      flat_shape = (shape[0], np.prod(shape[1:]))
      a = np.random.normal(0.0, 1.0, flat_shape)
      u, _, v = np.linalg.svd(a, full_matrices=False)
      # pick the one with the correct shape
      q = u if u.shape == flat_shape else v
      q = q.reshape(shape) #this needs to be corrected to float32
      print('you have initialized one orthogonal matrix.')
      return tf.constant(scale * q[:shape[0], :shape[1]], dtype=tf.float32)
    return _initializer

def identity_initializer():
    def _initializer(shape, dtype=tf.float32, partition_info=None):
        if len(shape) == 1:
            return tf.constant_op.constant(0., dtype=dtype, shape=shape)
        elif len(shape) == 2 and shape[0] == shape[1]:
            return tf.constant(np.identity(shape[0], dtype=np.float32))
        elif len(shape) == 4 and shape[2] == shape[3]:
            array = np.zeros(shape, dtype=float)
            cx, cy = shape[0]/2, shape[1]/2
            for i in range(shape[2]):
                array[cx, cy, i, i] = 1
            return tf.constant(array, dtype=dtype)
        else:
            raise Exception("Shape error")
    return _initializer

def custom_initializer(type, shape):
    if type == 'orthogonal_initializer':
        return orthogonal_initializer(shape)
    elif type == 'identity_initializer':
        return identity_initializer(shape)

def custom_conv_layer(inputT, shape, name=None, reuse=False, initializer=None):
    with tf.variable_scope(name, reuse=reuse) as scope:
        kernel = tf.get_variable('weights', shape, initializer=initializer)

        conv = tf.nn.conv2d(inputT,
                            kernel,
                            strides=[1, 1, 1, 1],
                            padding='SAME')
        conv_biases = tf.Variable(tf.zeros([shape[-1]], dtype=data_type()))
        relu = tf.nn.relu(tf.nn.bias_add(conv, conv_biases))

        pool = tf.nn.max_pool(relu,
                              ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1],
                              padding='SAME')
        return pool
def custom_fc_layer(inputT, shape, activations=True, name=None, reuse=False, initializer=None):
    input_shape = inputT.get_shape().as_list()
    print(input_shape)
    if len(input_shape) > 2:
        reshape = tf.reshape(inputT, (input_shape[0], -1))
    else:
        reshape = inputT
    print('reshaped',reshape.get_shape().as_list() )
    with tf.variable_scope(name, reuse=reuse) as scope:
        fc_weights = tf.get_variable('weights', shape, initializer=initializer)
        fc_biases = tf.Variable(tf.constant(0.1, shape=[shape[-1]], dtype=data_type()))
        return tf.nn.relu(tf.matmul(reshape, fc_weights) + fc_biases) if activations==True else tf.matmul(reshape, fc_weights) + fc_biases

def build_model(data, reuse=False):
    pool1 = custom_conv_layer(data, [5, 5, NUM_CHANNELS, 32], name="conv1", reuse=reuse, initializer=None)
    pool2 = custom_conv_layer(pool1, [5, 5, 32, 64], name="conv2", reuse=reuse, initializer=None)

    fc1 = custom_fc_layer(pool2, [IMAGE_SIZE//4 * IMAGE_SIZE//4 * 64, 512], name="fc1", reuse=reuse, initializer=None)
    # fc2 = custom_fc_layer(fc1, [512, NUM_LABELS], name="fc2", activations=False, reuse=reuse)
    if FINETUNE:
        fc2_1 = custom_fc_layer(fc1, [512, 256], name="fc2_1", activations=False, reuse=reuse)
        fc2 = custom_fc_layer(fc2_1, [256, NUM_LABELS], name="fc2", activations=False, reuse=reuse)
    else:
        fc2 = custom_fc_layer(fc1, [512, NUM_LABELS], name="fc2", activations=False, reuse=reuse)

    return fc2

def main(argv=None):  # pylint: disable=unused-argument
  if FLAGS.self_test:
    print('Running self-test.')
    train_data, train_labels = fake_data(256)
    validation_data, validation_labels = fake_data(EVAL_BATCH_SIZE)
    test_data, test_labels = fake_data(EVAL_BATCH_SIZE)
    num_epochs = 1
  else:
    # Get the data.
    train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
    train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
    test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
    test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')

    # Extract it into numpy arrays.
    train_data = extract_data(train_data_filename, 60000)
    train_labels = extract_labels(train_labels_filename, 60000)
    test_data = extract_data(test_data_filename, 10000)
    test_labels = extract_labels(test_labels_filename, 10000)

    # Generate a validation set.
    validation_data = train_data[:VALIDATION_SIZE, ...]
    validation_labels = train_labels[:VALIDATION_SIZE]
    train_data = train_data[VALIDATION_SIZE:, ...]
    train_labels = train_labels[VALIDATION_SIZE:]
    num_epochs = NUM_EPOCHS
  train_size = train_labels.shape[0]

  # This is where training samples and labels are fed to the graph.
  # These placeholder nodes will be fed a batch of training data at each
  # training step using the {feed_dict} argument to the Run() call below.
  train_data_node = tf.placeholder(
      data_type(),
      shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
  train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
  eval_data = tf.placeholder(
      data_type(),
      shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))

  # The variables below hold all the trainable weights. They are passed an
  # initial value which will be assigned when we call:
  # {tf.initialize_all_variables().run()}
  """
  conv1_weights = tf.Variable(
      tf.truncated_normal([5, 5, NUM_CHANNELS, 32],  # 5x5 filter, depth 32.
                          stddev=0.1,
                          seed=SEED, dtype=data_type()))
  """
  initializer = tf.truncated_normal_initializer(stddev=0.1,seed=SEED, dtype=data_type())
  #initializer = orthogonal_initializer()
  #initializer = None

  conv1_weights = tf.get_variable('conv1', [5, 5, NUM_CHANNELS, 32], initializer=initializer)

  conv1_biases = tf.Variable(tf.zeros([32], dtype=data_type()))
  """
  conv2_weights = tf.Variable(tf.truncated_normal(
      [5, 5, 32, 64], stddev=0.1,
      seed=SEED, dtype=data_type()))
  """
  conv2_weights = tf.get_variable('conv2', [5, 5, 32, 64], initializer=initializer)
  conv2_biases = tf.Variable(tf.constant(0.1, shape=[64], dtype=data_type()))
  """
  fc1_weights = tf.Variable(  # fully connected, depth 512.
      tf.truncated_normal([IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512],
                          stddev=0.1,
                          seed=SEED,
                          dtype=data_type()))
  """
  fc1_weights = tf.get_variable('fc1', [IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512], initializer=initializer)
  fc1_biases = tf.Variable(tf.constant(0.1, shape=[512], dtype=data_type()))
  """
  fc2_weights = tf.Variable(tf.truncated_normal([512, NUM_LABELS],
                                                stddev=0.1,
                                                seed=SEED,
                                                dtype=data_type()))
  """
  fc2_weights = tf.get_variable('fc2', [512, NUM_LABELS], initializer=initializer)
  fc2_biases = tf.Variable(tf.constant(
      0.1, shape=[NUM_LABELS], dtype=data_type()))

  # We will replicate the model structure for the training subgraph, as well
  # as the evaluation subgraphs, while sharing the trainable parameters.
  def model(data, train=False):
    """The Model definition."""
    # 2D convolution, with 'SAME' padding (i.e. the output feature map has
    # the same size as the input). Note that {strides} is a 4D array whose
    # shape matches the data layout: [image index, y, x, depth].
    conv = tf.nn.conv2d(data,
                        conv1_weights,
                        strides=[1, 1, 1, 1],
                        padding='SAME')
    # Bias and rectified linear non-linearity.
    relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
    # Max pooling. The kernel size spec {ksize} also follows the layout of
    # the data. Here we have a pooling window of 2, and a stride of 2.
    pool = tf.nn.max_pool(relu,
                          ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1],
                          padding='SAME')
    conv = tf.nn.conv2d(pool,
                        conv2_weights,
                        strides=[1, 1, 1, 1],
                        padding='SAME')
    relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
    pool = tf.nn.max_pool(relu,
                          ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1],
                          padding='SAME')
    # Reshape the feature map cuboid into a 2D matrix to feed it to the
    # fully connected layers.
    pool_shape = pool.get_shape().as_list()
    reshape = tf.reshape(
        pool,
        [pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
    # Fully connected layer. Note that the '+' operation automatically
    # broadcasts the biases.
    hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
    # Add a 50% dropout during training only. Dropout also scales
    # activations such that no rescaling is needed at evaluation time.
    #if train:
      #hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
    return tf.matmul(hidden, fc2_weights) + fc2_biases

  # Training computation: logits + cross-entropy loss.

  # logits = model(train_data_node, True)

  logits = build_model(train_data_node)

  loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
      logits, train_labels_node))

  # L2 regularization for the fully connected parameters.
  regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
                  tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
  # Add the regularization term to the loss.
  loss += 5e-4 * regularizers

  # Optimizer: set up a variable that's incremented once per batch and
  # controls the learning rate decay.
  batch = tf.Variable(0, dtype=data_type())
  # Decay once per epoch, using an exponential schedule starting at 0.01.
  learning_rate = tf.train.exponential_decay(
      0.01,                # Base learning rate.
      batch * BATCH_SIZE,  # Current index into the dataset.
      train_size,          # Decay step.
      0.95,                # Decay rate.
      staircase=True)
  # Use simple momentum for the optimization.
  optimizer = tf.train.MomentumOptimizer(learning_rate,0.9).minimize(loss, global_step=batch)
  if PARTIALLY_FINETUNE:
      optimizer = tf.train.MomentumOptimizer(learning_rate,0.9)
      grads = optimizer.compute_gradients(
          loss,
          [v for v in tf.trainable_variables() if not (v.name.startswith("fc2"))]
      )
      optimizer = optimizer.apply_gradients(grads, global_step=batch)
  else:
      optimizer = tf.train.MomentumOptimizer(learning_rate,0.9)
      grads = optimizer.compute_gradients(
          loss,
      )
      optimizer = optimizer.apply_gradients(grads, global_step=batch)

  # Predictions for the current training minibatch.
  train_prediction = tf.nn.softmax(logits)

  # Predictions for the test and validation, which we'll compute less often.
  #eval_prediction = tf.nn.softmax(model(eval_data))
  eval_prediction = tf.nn.softmax(build_model(eval_data, reuse=True))

  # Small utility function to evaluate a dataset by feeding batches of data to
  # {eval_data} and pulling the results from {eval_predictions}.
  # Saves memory and enables this to run on smaller GPUs.
  def eval_in_batches(data, sess):
    """Get all predictions for a dataset by running it in small batches."""
    size = data.shape[0]
    if size < EVAL_BATCH_SIZE:
      raise ValueError("batch size for evals larger than dataset: %d" % size)
    predictions = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.float32)
    for begin in xrange(0, size, EVAL_BATCH_SIZE):
      end = begin + EVAL_BATCH_SIZE
      if end <= size:
        predictions[begin:end, :] = sess.run(
            eval_prediction,
            feed_dict={eval_data: data[begin:end, ...]})
      else:
        batch_predictions = sess.run(
            eval_prediction,
            feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
        predictions[begin:, :] = batch_predictions[begin - size:, :]
    return predictions

  # Create a local session to run the training.
  start_time = time.time()
  if FINETUNE:
    saver = tf.train.Saver([v for v in tf.trainable_variables() if not (v.name.startswith("fc2"))], max_to_keep=500)
    #saver = tf.train.Saver(tf.all_variables(), max_to_keep=500)
  else:
    saver = tf.train.Saver(tf.all_variables(), max_to_keep=500)
  with tf.Session() as sess:
    # Run all the initializers to prepare the trainable parameters.

    if FINETUNE:
        tf.initialize_all_variables().run()
        saver.restore(sess, "./mnist.ckpt-1900")
    elif PARTIALLY_FINETUNE:
        tf.initialize_all_variables().run()
        saver.restore(sess, "./mnist_1.ckpt-3000")
    else:
        tf.initialize_all_variables().run()
    print('Initialized!')
    # Loop through training steps.
    prev_fc2 = []
    prev_other = []
    for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
      if PARTIALLY_FINETUNE:
          fc2_weights_for_print = [v for v in tf.trainable_variables() if v.name.startswith("fc2")][0]
          other_weights_for_print = [v for v in tf.trainable_variables() if v.name.startswith("fc1")][0]
          if len(prev_fc2) != 0 and step % 100 == 0:
              print(np.allclose(fc2_weights_for_print.eval(), prev_fc2))
              print(np.allclose(other_weights_for_print.eval(), prev_other))
          prev_fc2 = fc2_weights_for_print.eval()
          prev_other = other_weights_for_print.eval()
      # Compute the offset of the current minibatch in the data.
      # Note that we could use better randomization across epochs.
      offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
      batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
      batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
      # This dictionary maps the batch data (as a numpy array) to the
      # node in the graph it should be fed to.
      feed_dict = {train_data_node: batch_data,
                   train_labels_node: batch_labels}
      # Run the graph and fetch some of the nodes.
      _, l, lr, predictions = sess.run(
          [optimizer, loss, learning_rate, train_prediction],
          feed_dict=feed_dict)
      if step % EVAL_FREQUENCY == 0:
        elapsed_time = time.time() - start_time
        start_time = time.time()
        print('Step %d (epoch %.2f), %.1f ms' %
              (step, float(step) * BATCH_SIZE / train_size,
               1000 * elapsed_time / EVAL_FREQUENCY))
        print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
        print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
        print('Validation error: %.1f%%' % error_rate(
            eval_in_batches(validation_data, sess), validation_labels))
        test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
        if not FINETUNE and not PARTIALLY_FINETUNE:
            if (test_error<1.0):
                print('stop! iter = ', step)
                saver.save(sess, "./mnist_1.ckpt", global_step=step)
                break

        sys.stdout.flush()
    # Finally print the result!
    test_error = error_rate(eval_in_batches(test_data, sess), test_labels)

    print('Test error: %.1f%%' % test_error)
    if FLAGS.self_test:
      print('test_error', test_error)
      assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % (
          test_error,)


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