Created by: songyouwei

Motivation

In dygraph, most Parameters of Layers are created when the __call__ is first invoked, thus part of the shape of the Parameters can be derived from the input, which is similar to paddle's static graph API. This design pattern introduces many problems. The model depends on data, when a complex model (a subclass of layer) is inited, there are no parameters in it, and we can't load the previously saved checkpoint before a __call__. It also brings troubles to the initialization of custom Parameters. The creation of Layer also requires the name_scope parameter, which is unnecessary and troublesome for users.

This PR removes build_once function and name_scope parameter.

API changes

fluid.dygraph.Layer

1. name_scope parameter is not required in Layer's __init__
2. simplify create_parameter api, only shape is required

__init__ of fluid.dygraph.nn.*

1. For all subclasses of Layer: name_scopes are removed, all missing optional dtype parameters are added.
2. Conv2D, Conv2DTranspose, Conv3D and Conv3DTranspose: add num_channels (int): The number of channels in the input image.
3. LayerNorm: remove begin_norm_axis, add normalized_shape (int or list or tuple): Input shape from an expected input of size :math:[*, normalized_shape[0], normalized_shape[1], ..., normalized_shape[-1]]. If it is a single integer, this module will normalize over the last dimension which is expected to be of that specific size.
4. NCE: add dim (int): Dimension of input (possibly embedding dim).
5. PRelu: add input_shape(list or tuple, optional): The shape of input. This parameter is required when mode is not "all". Default: None.
6. BilinearTensorProduct: remove size. add input1_dim (int): The dimension of each first input. add input2_dim (int): The dimension of each second input.
7. GroupNorm: add channels (int): The number of channels of input.
8. SpectralNorm: add weight_shape (list or tuple): The shape of weight parameter.
9. TreeConv: add feature_size (int): last dimension of nodes_vector.

Migrating

For existing subclasses of Layer:

1. remove build_once function, and move Parameter creations into __init__.
2. avoid positional arguments for Layer.create_parameter function, and only shape is required.

For constructions of Layer's subclasses in dygraph models:

1. remove name_scope parameter of __init__
2. add possible new parameters of __init__ (these parameters are input shape related)

Example (Conv2D)

before

import paddle.fluid as fluid
import numpy as np
data = np.random.uniform(-1, 1, [10, 3, 32, 32]).astype('float32')
with fluid.dygraph.guard():
    data = fluid.to_variable(data)
    conv2d = fluid.Conv2D("conv2d", 2, 3)
    conv = conv2d(data)

after

import paddle.fluid as fluid
import numpy as np
data = np.random.uniform(-1, 1, [10, 3, 32, 32]).astype('float32')
with fluid.dygraph.guard():
    data = fluid.to_variable(data)
    conv2d = fluid.Conv2D(3, 2, 3)  # "conv2d" removed, and 3 is num_channels
    conv = conv2d(data)