SetTransformer编码器
- class dgl.nn.pytorch.glob.SetTransformerEncoder(d_model, n_heads, d_head, d_ff, n_layers=1, block_type='sab', m=None, dropouth=0.0, dropouta=0.0)[source]
Bases:
Module来自Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks的编码器模块
- Parameters:
d_model (int) – 模型的隐藏大小。
n_heads (int) – 头的数量。
d_head (int) – 每个头的隐藏大小。
d_ff (int) – FFN（位置前馈网络）层中的内核大小。
n_layers (int) – The number of layers.
block_type (str) – 构建块类型：'sab'（集合注意力块）或 'isab'（诱导集合注意力块）。
m (int 或 None) – ISAB 块中诱导向量的数量。如果块类型为 'sab'，则设置为 None。
dropouth (float) – 每个子层的丢弃率。
dropouta (float) – 注意力头的丢弃率。
示例
>>> import dgl >>> import torch as th >>> from dgl.nn import SetTransformerEncoder >>> >>> g1 = dgl.rand_graph(3, 4) # g1 is a random graph with 3 nodes and 4 edges >>> g1_node_feats = th.rand(3, 5) # feature size is 5 >>> g1_node_feats tensor([[0.8948, 0.0699, 0.9137, 0.7567, 0.3637], [0.8137, 0.8938, 0.8377, 0.4249, 0.6118], [0.5197, 0.9030, 0.6825, 0.5725, 0.4755]]) >>> >>> g2 = dgl.rand_graph(4, 6) # g2 is a random graph with 4 nodes and 6 edges >>> g2_node_feats = th.rand(4, 5) # feature size is 5 >>> g2_node_feats tensor([[0.2053, 0.2426, 0.4111, 0.9028, 0.5658], [0.5278, 0.6365, 0.9990, 0.2351, 0.8945], [0.3134, 0.0580, 0.4349, 0.7949, 0.3891], [0.0142, 0.2709, 0.3330, 0.8521, 0.6925]]) >>> >>> set_trans_enc = SetTransformerEncoder(5, 4, 4, 20) # create a settrans encoder.
案例1：输入单个图形
>>> set_trans_enc(g1, g1_node_feats) tensor([[ 0.1262, -1.9081, 0.7287, 0.1678, 0.8854], [-0.0634, -1.1996, 0.6955, -0.9230, 1.4904], [-0.9972, -0.7924, 0.6907, -0.5221, 1.6211]], grad_fn=<NativeLayerNormBackward>)
案例2：输入一批图形
构建一批DGL图并将所有图的节点特征连接成一个张量。
>>> batch_g = dgl.batch([g1, g2]) >>> batch_f = th.cat([g1_node_feats, g2_node_feats]) >>> >>> set_trans_enc(batch_g, batch_f) tensor([[ 0.1262, -1.9081, 0.7287, 0.1678, 0.8854], [-0.0634, -1.1996, 0.6955, -0.9230, 1.4904], [-0.9972, -0.7924, 0.6907, -0.5221, 1.6211], [-0.7973, -1.3203, 0.0634, 0.5237, 1.5306], [-0.4497, -1.0920, 0.8470, -0.8030, 1.4977], [-0.4940, -1.6045, 0.2363, 0.4885, 1.3737], [-0.9840, -1.0913, -0.0099, 0.4653, 1.6199]], grad_fn=<NativeLayerNormBackward>)
注释
SetTransformerEncoder 不是一个读出层，它返回的张量是节点级别的表示，而不是图级别的表示，而 SetTransformerDecoder 会返回一个图读出张量。