Python layers.batch_norm方法代碼示例

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def basic_residual_block(input, planes, scope, kernel_size=3, stride=1, downsamplefn=None):
    '''

    :param planes:   kernel number , output channel
    :param stride:   stride
    :param downsample:  downsample fn
    :return:
    '''

    residual = input

    with tf.variable_scope(scope):
        _out = slim.conv2d(input, num_outputs=planes, kernel_size=[kernel_size, kernel_size],
                           stride=stride, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

        _out = slim.conv2d(_out, num_outputs=planes, kernel_size=[kernel_size, kernel_size],
                           stride=stride, activation_fn=None, normalizer_fn=batch_norm)

        if downsamplefn is not None:
            residual = downsamplefn(residual)

        out = _out + residual
        out = tf.nn.relu(out)

    return out 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def upsample_block(input, ratio, planes, scope):
    '''

    :param input:
    :param planes:
    :param scope:
    :param kernel_size:
    :param stride:
    :return:
    '''

    with tf.variable_scope(scope):
        _out = slim.conv2d(input, num_outputs=planes, kernel_size=[1, 1], stride=1,
                           activation_fn=None, normalizer_fn=batch_norm, padding='SAME')
        shape = _out.shape
        _out = tf.image.resize_nearest_neighbor(_out, (shape[1] * ratio, shape[1] * ratio))

    return _out 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def forward(self, input):
        with tf.variable_scope(self.scope):
            # conv1 + bn1 + relu1
            _out = slim.conv2d(input, num_outputs=self.num_channels, kernel_size=[3, 3],
                               stride=2, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

            # conv2 + bn2 + relu2
            _out = slim.conv2d(_out, num_outputs=self.num_channels, kernel_size=[3, 3],
                               stride=2, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

            # bottlenect
            for i in range(self.num_blocks):
                _out = bottleneck_block(_out, planes=self.bottleneck_channels,
                                        scope='_BN' + str(i), downsamplefn=trans_block if i == 0 else None)

            # one 3x3 keep same resolution and one 3x3 to 1/2x resolution
            _same_res = slim.conv2d(_out, num_outputs=self.output_channels[0], kernel_size=[3, 3],
                                    stride=1, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

            _half_res = slim.conv2d(_out, num_outputs=self.output_channels[1], kernel_size=[3, 3],
                                    stride=2, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

        return [_same_res, _half_res] 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def normalize(inp, activation, reuse, scope):
    """The function to forward the normalization.
    Args:
      inp: the input feature maps.
      reuse: whether reuse the variables for the batch norm.
      scope: the label for this conv layer.
      activation: the activation function for this conv layer.
    Return:
      The processed feature maps.
    """
    if FLAGS.norm == 'batch_norm':
        return tf_layers.batch_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'layer_norm':
        return tf_layers.layer_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'None':
        if activation is not None:
            return activation(inp)        
        return inp
    else:
        raise ValueError('Please set correct normalization.')

## Loss functions 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_normalization(input_data, name, activation_method="relu", is_train=True):
    """
    BN layer
    :param input_data: the input data
    :param name: name
    :param activation_method: the method of activation function
    :param is_train: if False, skip this layer, default is True
    :return:
        output: output after batch normalization
    """
    output = batch_norm(inputs=input_data, decay=0.9, center=True, scale=True, epsilon=1e-5, scope=name, updates_collections=None,
                        reuse=tf.AUTO_REUSE, is_training=is_train, zero_debias_moving_mean=True)
    output = activation_layer(input_data=output,
                              activation_method=activation_method)

    print("name: %s, activation: %s, is_train: %r" % (name, activation_method, is_train))

    return output 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_norm_layer(x,train_phase,scope_bn):
    outputs = tf.contrib.layers.batch_norm(x, is_training=train_phase, center=False, scale=False, activation_fn=tf.nn.relu, updates_collections=None, scope='batch_norm')
    return outputs
    # def batch_norm_layer(x,train_phase,scope_bn):
#     bn_train = batch_norm(x, decay=0.999, center=True, scale=True,
#     updates_collections=None,
#     is_training=True,
#     reuse=None, # is this right?
#     trainable=True,
#     scope=scope_bn)
#     bn_inference = batch_norm(x, decay=0.999, center=True, scale=True,
#     updates_collections=None,
#     is_training=False,
#     reuse=True, # is this right?
#     trainable=True,
#     scope=scope_bn)
#     z = tf.cond(train_phase, lambda: bn_train, lambda: bn_inference)
#     return z 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_norm_layer(x,train_phase,scope_bn):
    bn_train = batch_norm(x, decay=0.999, center=True, scale=True,
    updates_collections=None,
    is_training=True,
    reuse=None, # is this right?
    trainable=True,
    scope=scope_bn)
    bn_inference = batch_norm(x, decay=0.999, center=True, scale=True,
    updates_collections=None,
    is_training=False,
    reuse=True, # is this right?
    trainable=True,
    scope=scope_bn)
    z = tf.cond(train_phase, lambda: bn_train, lambda: bn_inference)
    return z

#def batch_norm_layer(x,train_phase,scope_bn):
#    outputs = tf.contrib.layers.batch_norm(x, is_training=train_phase, center=False, scale=False, activation_fn=tf.nn.relu, updates_collections=None, scope='scope_bn) 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_norm_layer(self, x, train_phase, scope_bn):
        bn_train = batch_norm(x, decay=self.batch_norm_decay, center=True, scale=True, updates_collections=None,
                              is_training=True, reuse=None, trainable=True, scope=scope_bn)
        bn_inference = batch_norm(x, decay=self.batch_norm_decay, center=True, scale=True, updates_collections=None,
                                  is_training=False, reuse=True, trainable=True, scope=scope_bn)
        z = tf.cond(train_phase, lambda: bn_train, lambda: bn_inference)
        return z 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def normalize(inp, activation, reuse, scope):
    if FLAGS.norm == 'batch_norm':
        return tf_layers.batch_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'layer_norm':
        return tf_layers.layer_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'None':
        if activation is not None:
            return activation(inp)
        else:
            return inp


# Loss functions 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_normal(input , scope="scope" , reuse=False):
    return batch_norm(input , epsilon=1e-5, decay=0.9 , scale=True, scope=scope , reuse=reuse, fused=True, updates_collections=None) 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def normalize(inp, activation, reuse, scope):
    if FLAGS.norm == 'batch_norm':
        return tf_layers.batch_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'layer_norm':
        return tf_layers.layer_norm(inp, activation_fn=activation, reuse=reuse, scope=scope)
    elif FLAGS.norm == 'None':
        if activation is not None:
            return activation(inp)
        else:
            return inp

## Loss functions 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def bottleneck_block(input, planes, scope, stride=1, downsamplefn=None):
    '''

    :param planes:   kernel number , output channel
    :param stride:   stride
    :param downsample:  downsample fn
    :return:
    '''

    expansion = 4
    residual = input

    with tf.variable_scope(scope):
        # conv1x1 + bn1 + relu
        _out = slim.conv2d(input, num_outputs=planes // expansion, kernel_size=[1, 1],
                           stride=1, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

        # conv2 3x3 + bn2 + relu
        _out = slim.conv2d(_out, num_outputs=planes // expansion, kernel_size=[3, 3],
                           stride=stride, activation_fn=tf.nn.relu, normalizer_fn=batch_norm)

        # conv3 1x1 + bn3
        _out = slim.conv2d(_out, num_outputs=planes, kernel_size=[1, 1],
                           stride=1, activation_fn=None, normalizer_fn=batch_norm)

        if downsamplefn is not None:
            residual = downsamplefn(residual, planes)

        out = _out + residual
        out = tf.nn.relu(out)

    return out 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def trans_block(input, planes):
    # conv 1x1 +  bn
    _out = slim.conv2d(input, num_outputs=planes, kernel_size=[1, 1],
                       stride=1, activation_fn=None, normalizer_fn=batch_norm)

    return _out 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def downsample_block(input, planes, scope, has_relu):
    '''

    :param input:
    :param planes:
    :param scope:
    :return:
    '''

    with tf.variable_scope(scope):
        _out = slim.conv2d(input, num_outputs=planes, kernel_size=[3, 3], stride=2,
                           activation_fn=tf.nn.relu if has_relu else None,
                           normalizer_fn=batch_norm, padding='SAME')

    return _out 

# 需要導入模塊: from tensorflow.contrib.layers.python import layers [as 別名]
# 或者: from tensorflow.contrib.layers.python.layers import batch_norm [as 別名]
def batch_normal(input, scope="scope", reuse=False):
    return batch_norm(input, epsilon=1e-5, decay=0.9, scale=True, scope=scope, reuse=reuse, fused=True, updates_collections=None)