import torch import torch.nn as nn import torch.nn.functional as F from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence from common.utils import HiddenData, ClassifierOutputData from model.decoder.interaction import BaseInteraction class LSTMEncoder(nn.Module): """ Encoder structure based on bidirectional LSTM. """ def __init__(self, embedding_dim, hidden_dim, dropout_rate): super(LSTMEncoder, self).__init__() # Parameter recording. self.__embedding_dim = embedding_dim self.__hidden_dim = hidden_dim // 2 self.__dropout_rate = dropout_rate # Network attributes. self.__dropout_layer = nn.Dropout(self.__dropout_rate) self.__lstm_layer = nn.LSTM( input_size=self.__embedding_dim, hidden_size=self.__hidden_dim, batch_first=True, bidirectional=True, dropout=self.__dropout_rate, num_layers=1 ) def forward(self, embedded_text, seq_lens): """ Forward process for LSTM Encoder. (batch_size, max_sent_len) -> (batch_size, max_sent_len, word_dim) -> (batch_size, max_sent_len, hidden_dim) :param embedded_text: padded and embedded input text. :param seq_lens: is the length of original input text. :return: is encoded word hidden vectors. """ # Padded_text should be instance of LongTensor. dropout_text = self.__dropout_layer(embedded_text) # Pack and Pad process for input of variable length. packed_text = pack_padded_sequence(dropout_text, seq_lens.cpu(), batch_first=True, enforce_sorted=False) lstm_hiddens, (h_last, c_last) = self.__lstm_layer(packed_text) padded_hiddens, _ = pad_packed_sequence(lstm_hiddens, batch_first=True) return padded_hiddens class GraphAttentionLayer(nn.Module): """ Simple GAT layer, similar to https://arxiv.org/abs/1710.10903 """ def __init__(self, in_features, out_features, dropout, alpha, concat=True): super(GraphAttentionLayer, self).__init__() self.dropout = dropout self.in_features = in_features self.out_features = out_features self.alpha = alpha self.concat = concat self.W = nn.Parameter(torch.zeros(size=(in_features, out_features))) nn.init.xavier_uniform_(self.W.data, gain=1.414) self.a = nn.Parameter(torch.zeros(size=(2 * out_features, 1))) nn.init.xavier_uniform_(self.a.data, gain=1.414) self.leakyrelu = nn.LeakyReLU(self.alpha) def forward(self, input, adj): h = torch.matmul(input, self.W) B, N = h.size()[0], h.size()[1] a_input = torch.cat([h.repeat(1, 1, N).view(B, N * N, -1), h.repeat(1, N, 1)], dim=2).view(B, N, -1, 2 * self.out_features) e = self.leakyrelu(torch.matmul(a_input, self.a).squeeze(3)) zero_vec = -9e15 * torch.ones_like(e) attention = torch.where(adj > 0, e, zero_vec) attention = F.softmax(attention, dim=2) attention = F.dropout(attention, self.dropout, training=self.training) h_prime = torch.matmul(attention, h) if self.concat: return F.elu(h_prime) else: return h_prime class GAT(nn.Module): def __init__(self, nfeat, nhid, nclass, dropout, alpha, nheads, nlayers=2): """Dense version of GAT.""" super(GAT, self).__init__() self.dropout = dropout self.nlayers = nlayers self.nheads = nheads self.attentions = [GraphAttentionLayer(nfeat, nhid, dropout=dropout, alpha=alpha, concat=True) for _ in range(nheads)] for i, attention in enumerate(self.attentions): self.add_module('attention_{}'.format(i), attention) if self.nlayers > 2: for i in range(self.nlayers - 2): for j in range(self.nheads): self.add_module('attention_{}_{}'.format(i + 1, j), GraphAttentionLayer(nhid * nheads, nhid, dropout=dropout, alpha=alpha, concat=True)) self.out_att = GraphAttentionLayer(nhid * nheads, nclass, dropout=dropout, alpha=alpha, concat=False) def forward(self, x, adj): x = F.dropout(x, self.dropout, training=self.training) input = x x = torch.cat([att(x, adj) for att in self.attentions], dim=2) if self.nlayers > 2: for i in range(self.nlayers - 2): temp = [] x = F.dropout(x, self.dropout, training=self.training) cur_input = x for j in range(self.nheads): temp.append(self.__getattr__('attention_{}_{}'.format(i + 1, j))(x, adj)) x = torch.cat(temp, dim=2) + cur_input x = F.dropout(x, self.dropout, training=self.training) x = F.elu(self.out_att(x, adj)) return x + input def normalize_adj(mx): """ Row-normalize matrix D^{-1}A torch.diag_embed: https://github.com/pytorch/pytorch/pull/12447 """ mx = mx.float() rowsum = mx.sum(2) r_inv = torch.pow(rowsum, -1) r_inv[torch.isinf(r_inv)] = 0. r_mat_inv = torch.diag_embed(r_inv, 0) mx = r_mat_inv.matmul(mx) return mx class GLGINInteraction(BaseInteraction): def __init__(self, **config): super().__init__(**config) self.intent_embedding = nn.Parameter( torch.FloatTensor(self.config["intent_label_num"], self.config["intent_embedding_dim"])) # 191, 32 nn.init.normal_(self.intent_embedding.data) self.adj = None self.__slot_lstm = LSTMEncoder( self.config["input_dim"] + self.config["intent_label_num"], config["output_dim"], config["dropout_rate"] ) self.__slot_graph = GAT( config["output_dim"], config["hidden_dim"], config["output_dim"], config["dropout_rate"], config["alpha"], config["num_heads"], config["num_layers"]) self.__global_graph = GAT( config["output_dim"], config["hidden_dim"], config["output_dim"], config["dropout_rate"], config["alpha"], config["num_heads"], config["num_layers"]) def generate_global_adj_gat(self, seq_len, index, batch, window): global_intent_idx = [[] for i in range(batch)] global_slot_idx = [[] for i in range(batch)] for item in index: global_intent_idx[item[0]].append(item[1]) for i, len in enumerate(seq_len): global_slot_idx[i].extend( list(range(self.config["intent_label_num"], self.config["intent_label_num"] + len))) adj = torch.cat([torch.eye(self.config["intent_label_num"] + max(seq_len)).unsqueeze(0) for i in range(batch)]) for i in range(batch): for j in global_intent_idx[i]: adj[i, j, global_slot_idx[i]] = 1. adj[i, j, global_intent_idx[i]] = 1. for j in global_slot_idx[i]: adj[i, j, global_intent_idx[i]] = 1. for i in range(batch): for j in range(self.config["intent_label_num"], self.config["intent_label_num"] + seq_len[i]): adj[i, j, max(self.config["intent_label_num"], j - window):min(seq_len[i] + self.config["intent_label_num"], j + window + 1)] = 1. if self.config["row_normalized"]: adj = normalize_adj(adj) adj = adj.to(self.intent_embedding.device) return adj def generate_slot_adj_gat(self, seq_len, batch, window): slot_idx_ = [[] for i in range(batch)] adj = torch.cat([torch.eye(max(seq_len)).unsqueeze(0) for i in range(batch)]) for i in range(batch): for j in range(seq_len[i]): adj[i, j, max(0, j - window):min(seq_len[i], j + window + 1)] = 1. if self.config["row_normalized"]: adj = normalize_adj(adj) adj = adj.to(self.intent_embedding.device) return adj def forward(self, encode_hidden: HiddenData, pred_intent: ClassifierOutputData = None, intent_index=None): seq_lens = encode_hidden.inputs.attention_mask.sum(-1) slot_lstm_out = self.__slot_lstm(torch.cat([encode_hidden.slot_hidden, pred_intent.classifier_output], dim=-1), seq_lens) global_adj = self.generate_global_adj_gat(seq_lens, intent_index, len(seq_lens), self.config["slot_graph_window"]) slot_adj = self.generate_slot_adj_gat(seq_lens, len(seq_lens), self.config["slot_graph_window"]) batch = len(seq_lens) slot_graph_out = self.__slot_graph(slot_lstm_out, slot_adj) intent_in = self.intent_embedding.unsqueeze(0).repeat(batch, 1, 1) global_graph_in = torch.cat([intent_in, slot_graph_out], dim=1) encode_hidden.update_slot_hidden_state(self.__global_graph(global_graph_in, global_adj)) return encode_hidden