ResNet && DenseNet(实践篇)

上篇博客说了ResNet和DenseNet的原理，这次说说具体实现

---

ResNet

def basic_block(input, in_features, out_features, stride, is_training, keep_prob):
    """Residual block"""
  if stride == 1:
    shortcut = input
  else:
    shortcut = tf.nn.avg_pool(input, [ 1, stride, stride, 1 ], [1, stride, stride, 1 ], 'VALID')
    shortcut = tf.pad(shortcut, [[0, 0], [0, 0], [0, 0],
      [(out_features-in_features)//2, (out_features-in_features)//2]])
  current = conv2d(input, in_features, out_features, 3, stride)
  current = tf.nn.dropout(current, keep_prob)
  current = tf.contrib.layers.batch_norm(current, scale=True, is_training=is_training, updates_collections=None)
  current = tf.nn.relu(current)
  current = conv2d(current, out_features, out_features, 3, 1)
  current = tf.nn.dropout(current, keep_prob)
  current = tf.contrib.layers.batch_norm(current, scale=True, is_training=is_training, updates_collections=None)
  return current + shortcut

def block_stack(input, in_features, out_features, stride, depth, is_training, keep_prob):
    """Stack Residual block"""
  current = basic_block(input, in_features, out_features, stride, is_training, keep_prob)
  for _d in xrange(depth - 1):
    current = basic_block(current, out_features, out_features, 1, is_training, keep_prob)
  return current

DenseNet

def conv2d(input, in_features, out_features, kernel_size, with_bias=False):
  W = weight_variable([ kernel_size, kernel_size, in_features, out_features ])
  conv = tf.nn.conv2d(input, W, [ 1, 1, 1, 1 ], padding='SAME')
  if with_bias:
    return conv + bias_variable([ out_features ])
  return conv

def batch_activ_conv(current, in_features, out_features, kernel_size, is_training, keep_prob):
    """BatchNorm+Relu+conv+dropout"""
  current = tf.contrib.layers.batch_norm(current, scale=True, is_training=is_training, updates_collections=None)
  current = tf.nn.relu(current)
  current = conv2d(current, in_features, out_features, kernel_size)
  current = tf.nn.dropout(current, keep_prob)
  return current

def block(input, layers, in_features, growth, is_training, keep_prob):
    """Dense Block"""
  current = input
  features = in_features
  for idx in xrange(layers):
    tmp = batch_activ_conv(current, features, growth, 3, is_training, keep_prob)
    current = tf.concat(3, (current, tmp))
    features += growth
  return current, features

def model():
    """DenseNet on ImageNet"""
    current = tf.reshape(xs, [ -1, 32, 32, 3 ])  # Input
    current = conv2d(current, 3, 16, 3)

    current, features = block(current, layers, 16, 12, is_training, keep_prob)
    current = batch_activ_conv(current, features, features, 1, is_training, keep_prob)
    current = avg_pool(current, 2)
    current, features = block(current, layers, features, 12, is_training, keep_prob)
    current = batch_activ_conv(current, features, features, 1, is_training, keep_prob)
    current = avg_pool(current, 2)
    current, features = block(current, layers, features, 12, is_training, keep_prob)

    current = tf.contrib.layers.batch_norm(current, scale=True, is_training=is_training, updates_collections=None)
    current = tf.nn.relu(current)
    current = avg_pool(current, 8)
    final_dim = features
    current = tf.reshape(current, [ -1, final_dim ])
    Wfc = weight_variable([ final_dim, label_count ])
    bfc = bias_variable([ label_count ])
    ys_ = tf.nn.softmax( tf.matmul(current, Wfc) + bfc )

---

代码不是完整的，只是表达最navie的思想核心部分