tf.nn.conv2d_transpose和slim.conv2d_transpose之间的区别

Question

What is the difference between this two function in Tensorflow

tf.nn.conv2d_transpose(
    value,
    filter,
    output_shape,
    strides,
    padding='SAME',
    data_format='NHWC',
    name=None
)

It full defination is in tf.nn.conv2d_transpose slim.conv2d_transpose is define as follow:

tf.layers.conv2d_transpose(
    inputs,
    filters,
    kernel_size,
    strides=(1, 1),
    padding='valid',
    data_format='channels_last',
    activation=None,
    use_bias=True,
    kernel_initializer=None,
    bias_initializer=tf.zeros_initializer(),
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
    trainable=True,
    name=None,
    reuse=None
)

full defination is in slim.conv2d_transpose

how can I define the output shape in slim.conv2d_transpose

Answer 1

There is a significant difference between them. While tf.nn.conv2d_transpose represents an operation in the computational graph, tf.layers.conv2d_transpose defines the entire layer.

Being more precise tf.nn.conv2d_transpose applies a convolutional filter to the inputs .

tf.layers.conv2d_transpose , on the other hand, first creates trainable variables that will serve as filter according to the arguments given, and then it internally calls some conv2d_transpose operation. Based on the arguments, it also applies some other operations as adding bias, applying non-linearity, or normalizing the weights or inputs.

With tf.layers.conv2d_transpose you do not specify output shape as it is computed from filter size, input size, and stride. Here is the formula.