Created by: songyouwei

Currently, dygraph Embedding layer uses the outdated lookup_table op, which forces the last dimension of the input to be 1. But fluid.embedding uses lookup_table_v2 op.

This PR changes the underlying op to lookup_table_v2, the shape of output Tensor is generated by appending an emb_size dimension to the last dimension of the input Tensor shape.

Example:

before

import paddle.fluid as fluid
import numpy as np
dict_size = 100
emb_dim = 32
with fluid.dygraph.guard():
    x_data = np.arange(12).reshape(4, 3).astype('int64')
    x = fluid.dygraph.to_variable(x_data)
    x = fluid.layers.unsqueeze(input=x, axes=[-1])  # need unsqueeze the last dim
    emb = fluid.dygraph.Embedding(
            size=[dict_size, emb_dim],
            param_attr='emb.w',
            is_sparse=False)
    out = emb(x)
    out.shape

after

import paddle.fluid as fluid
import numpy as np
dict_size = 100
emb_dim = 32
with fluid.dygraph.guard():
    x_data = np.arange(12).reshape(4, 3).astype('int64')
    x = fluid.dygraph.to_variable(x_data)
    emb = fluid.dygraph.Embedding(
            size=[dict_size, emb_dim],
            param_attr='emb.w',
            is_sparse=False)
    out = emb(x)
    out.shape