GLMP 代码 详细注释

GLMP 缩写来自论文 GLOBAL-TO-LOCAL MEMORY POINTER NETWORKS FOR TASK-ORIENTED DIALOGUE 下面是它代码的详细注释（已跑通）

3.1 模型

3.1.1 ContextRNN

class ContextRNN(nn.Module): def __init__(self, input_size, hidden_size, dropout, n_layers=1): #初始化设置参数 super(ContextRNN, self).__init__() self.input_size = input_size self.hidden_size = hidden_size self.n_layers = n_layers self.dropout = dropout #nn.Dropout：参数为float类型，将元素置0的概率 self.dropout_layer = nn.Dropout(dropout) #nn.Embedding:参数分别为（单词个数 词向量维度 遇到PAD_token输出0） #此处可看出embedding词向量的维度和hidden的维度相同 self.embedding = nn.Embedding(input_size, hidden_size, padding_idx=PAD_token) #调用pytorch中的GRU模块，设置网络为双向GRU self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=dropout, bidirectional=True) self.W = nn.Linear(2*hidden_size, hidden_size) def get_state(self, bsz): """Get cell states and hidden states.""" return _cuda(torch.zeros(2, bsz, self.hidden_size)) def forward(self, input_seqs, input_lengths, hidden=None): # Note: we run this all at once (over multiple batches of multiple sequences) #contiguous函数返回一个内存连续的tensor #view函数返回一个tensor，必须有与原tensor相同的数据和相同数目的元素，但可以有不同的大小。 #一个tensor必须是连续的contiguous()才能被查看。 #两个函数联合作用将embedding的维度调整成一句一行 embedded = self.embedding(input_seqs.contiguous().view(input_seqs.size(0), -1).long()) embedded = embedded.view(input_seqs.size()+(embedded.size(-1),)) embedded = torch.sum(embedded, 2).squeeze(2) embedded = self.dropout_layer(embedded) #初始化hidden hidden = self.get_state(input_seqs.size(1)) if input_lengths: embedded = nn.utils.rnn.pack_padded_sequence(embedded, input_lengths, batch_first=False) outputs, hidden = self.gru(embedded, hidden) if input_lengths: outputs, _ = nn.utils.rnn.pad_packed_sequence(outputs, batch_first=False) hidden = self.W(torch.cat((hidden[0], hidden[1]), dim=1)).unsqueeze(0) outputs = self.W(outputs) return outputs.transpose(0,1), hidden

3.1.2 ExternalKnowledge

class ExternalKnowledge(nn.Module): def __init__(self, vocab, embedding_dim, hop, dropout): #ExternalKnowledge的工作原理类似于MemoryNetwork super(ExternalKnowledge, self).__init__() self.max_hops = hop#跳数 self.embedding_dim = embedding_dim self.dropout = dropout self.dropout_layer = nn.Dropout(dropout) for hop in range(self.max_hops+1):#针对每一跳都初始化 #nn.Embedding:单词个数 词向量维度 遇到PAD_token输出0 C = nn.Embedding(vocab, embedding_dim, padding_idx=PAD_token) #将值用均值为0，方差为0.1的正态分布填充 C.weight.data.normal_(0, 0.1) ##将一个child module添加到当前model,被添加的module可以通过name属性来获取。 self.add_module("C_{}".format(hop), C) #定义查询C的方法？通过C_{i}来查询 self.C = AttrProxy(self, "C_") #定义softmax为单维度的softmax self.softmax = nn.Softmax(dim=1) self.sigmoid = nn.Sigmoid() #一维卷积函数Conv1d：参数分别为(进 出通道 卷积核的大小 输入的每一条边补充0的层数) self.conv_layer = nn.Conv1d(embedding_dim, embedding_dim, 5, padding=2) def add_lm_embedding(self, full_memory, kb_len, conv_len, hiddens): #将hiddens按照kb_len的顺序加入full_memory矩阵 for bi in range(full_memory.size(0)): start, end = kb_len[bi], kb_len[bi]+conv_len[bi] full_memory[bi, start:end, :] = full_memory[bi, start:end, :] + hiddens[bi, :conv_len[bi], :] return full_memory def load_memory(self, story, kb_len, conv_len, hidden, dh_outputs): # Forward multiple hop mechanism #squeeze函数：把第一个维度（维度为1）挤掉 u = [hidden.squeeze(0)] story_size = story.size() self.m_story = [] for hop in range(self.max_hops):#循环K跳来计算Attention权重 #求c^k_i #把外部知识的三元组挤成一个向量 embed_A = self.C[hop](story.contiguous().view(story_size[0], -1))#.long()) # b * (m * s) * e embed_A = embed_A.view(story_size+(embed_A.size(-1),)) # b * m * s * e #sum函数：返回输入张量给定维度上每行的和 embed_A = torch.sum(embed_A, 2).squeeze(2) # b * m * e if not args["ablationH"]: embed_A = self.add_lm_embedding(embed_A, kb_len, conv_len, dh_outputs) embed_A = self.dropout_layer(embed_A) if(len(list(u[-1].size()))==1): u[-1] = u[-1].unsqueeze(0) ## used for bsz = 1. #调整q^k到可以与C_i相乘的维度 u_temp = u[-1].unsqueeze(1).expand_as(embed_A) #将结果调整成可以进行单维度softmax的向量 prob_logit = torch.sum(embed_A*u_temp, 2) #按照论文所给公式求出p^k_i prob_ = self.softmax(prob_logit) #重复求c^k_i的步骤求出c^k+1_i embed_C = self.C[hop+1](story.contiguous().view(story_size[0], -1).long()) embed_C = embed_C.view(story_size+(embed_C.size(-1),)) embed_C = torch.sum(embed_C, 2).squeeze(2) if not args["ablationH"]: embed_C = self.add_lm_embedding(embed_C, kb_len, conv_len, dh_outputs) #调整p的维度来与C^k+1_i相乘 prob = prob_.unsqueeze(2).expand_as(embed_C) #求和得到o^k o_k = torch.sum(embed_C*prob, 1) #q^k+1 = q^k + o^k u_k = u[-1] + o_k u.append(u_k) self.m_story.append(embed_A) self.m_story.append(embed_C) #返回p^k(?)和q^k+1 return self.sigmoid(prob_logit), u[-1] def forward(self, query_vector, global_pointer): #U为查询向量 u = [query_vector] #循环k跳来得出最后的查询结果 for hop in range(self.max_hops): #从m_story中取出load_memory中存好的c^k m_A = self.m_story[hop] if not args["ablationG"]: #global_pointer更新Global contextual representation m_A = m_A * global_pointer.unsqueeze(2).expand_as(m_A) if(len(list(u[-1].size()))==1): u[-1] = u[-1].unsqueeze(0) ## used for bsz = 1. u_temp = u[-1].unsqueeze(1).expand_as(m_A) prob_logits = torch.sum(m_A*u_temp, 2) prob_soft = self.softmax(prob_logits) m_C = self.m_story[hop+1] if not args["ablationG"]: m_C = m_C * global_pointer.unsqueeze(2).expand_as(m_C) prob = prob_soft.unsqueeze(2).expand_as(m_C) o_k = torch.sum(m_C*prob, 1) u_k = u[-1] + o_k u.append(u_k) return prob_soft, prob_logits

3.1.3 LocalMemoryDecoder

class LocalMemoryDecoder(nn.Module): def __init__(self, shared_emb, lang, embedding_dim, hop, dropout): #初始化网络 super(LocalMemoryDecoder, self).__init__() self.num_vocab = lang.n_words self.lang = lang self.max_hops = hop self.embedding_dim = embedding_dim self.dropout = dropout self.dropout_layer = nn.Dropout(dropout) #将shared_emb保存为C（此C不同于ExternalKnowledge中的self.C) #根据GLMP中代码得出shared_emb为encoder的embedding self.C = shared_emb self.softmax = nn.Softmax(dim=1) #sketch RNN用于跑出没有槽值信息但是有sketch tag的response self.sketch_rnn = nn.GRU(embedding_dim, embedding_dim, dropout=dropout) self.relu = nn.ReLU() self.projector = nn.Linear(2*embedding_dim, embedding_dim) self.conv_layer = nn.Conv1d(embedding_dim, embedding_dim, 5, padding=2) self.softmax = nn.Softmax(dim = 1) def forward(self, extKnow, story_size, story_lengths, copy_list, encode_hidden, target_batches, max_target_length, batch_size, use_teacher_forcing, get_decoded_words, global_pointer): # Initialize variables for vocab and pointer #初始化输入输出矩阵 all_decoder_outputs_vocab = _cuda(torch.zeros(max_target_length, batch_size, self.num_vocab)) all_decoder_outputs_ptr = _cuda(torch.zeros(max_target_length, batch_size, story_size[1])) decoder_input = _cuda(torch.LongTensor([SOS_token] * batch_size)) #mask矩阵用来防止生成相同的槽 memory_mask_for_step = _cuda(torch.ones(story_size[0], story_size[1])) decoded_fine, decoded_coarse = [], [] hidden = self.relu(self.projector(encode_hidden)).unsqueeze(0) # Start to generate word-by-word for t in range(max_target_length): #hidden的生成在前四行，不同循环的不同变量只有decoder_input embed_q = self.dropout_layer(self.C(decoder_input)) # b * e if len(embed_q.size()) == 1: embed_q = embed_q.unsqueeze(0) _, hidden = self.sketch_rnn(embed_q.unsqueeze(0), hidden) #取sketch_RNN 的第一个hidden query_vector = hidden[0] #求p^vocab p_vocab = self.attend_vocab(self.C.weight, hidden.squeeze(0)) all_decoder_outputs_vocab[t] = p_vocab #topk函数：得到前k个元素，返回两个tensor，第一个为数值，第二个为下标 #此处是得到数值最大的元素的下标 #所以过不过softmax意义不大，这可能是代码把softmax注释掉的原因 _, topvi = p_vocab.data.topk(1) # query the external konwledge using the hidden state of sketch RNN #通过sketch RNN的hidden来向外部知识查询 #extKnow为ExternalKnowledge的forward函数 prob_soft, prob_logits = extKnow(query_vector, global_pointer) #得到L_t：本地内存指针的位置标志 all_decoder_outputs_ptr[t] = prob_logits if use_teacher_forcing:#是否使用标准答案来改变输入以改变sketchRNN生成的hidden #使用预先存入的生成 decoder_input = target_batches[:,t] else: #使用sketchRNN上次生成的output decoder_input = topvi.squeeze() if get_decoded_words: search_len = min(5, min(story_lengths)) prob_soft = prob_soft * memory_mask_for_step #取前search_len元素，作为填入槽的object预备 _, toppi = prob_soft.data.topk(search_len) temp_f, temp_c = [], [] for bi in range(batch_size): token = topvi[bi].item() #topvi[:,0][bi].item()#取下标 temp_c.append(self.lang.index2word[token])#取单词 if '@' in self.lang.index2word[token]:#判断是否是槽 #如果是槽且符合条件，那么可以将代表的宾语加入输出 cw = 'UNK' for i in range(search_len): if toppi[:,i][bi] < story_lengths[bi]-1: cw = copy_list[bi][toppi[:,i][bi].item()] break temp_f.append(cw) if args['record']: #mask矩阵标0，防止生成相同的槽 memory_mask_for_step[bi, toppi[:,i][bi].item()] = 0 else: #不是槽的话就直接输出词语 temp_f.append(self.lang.index2word[token]) decoded_fine.append(temp_f) decoded_coarse.append(temp_c) return all_decoder_outputs_vocab, all_decoder_outputs_ptr, decoded_fine, decoded_coarse def attend_vocab(self, seq, cond): scores_ = cond.matmul(seq.transpose(1,0))#首先对输入的矩阵转置，然后进行矩阵乘法 #论文中写的公式是带softmax的……但注释掉了就很秀 # scores = F.softmax(scores_, dim=1) return scores_

3.2 GLMP

3.2.1 Encoder&Decoder

def encode_and_decode(self, data, max_target_length, use_teacher_forcing, get_decoded_words): # Build unknown mask for memory #初始化mask矩阵 if args['unk_mask'] and self.decoder.training: story_size = data['context_arr'].size() rand_mask = np.ones(story_size) bi_mask = np.random.binomial([np.ones((story_size[0],story_size[1]))], 1-self.dropout)[0] rand_mask[:,:,0] = rand_mask[:,:,0] * bi_mask conv_rand_mask = np.ones(data['conv_arr'].size()) for bi in range(story_size[0]): start, end = data['kb_arr_lengths'][bi], data['kb_arr_lengths'][bi] + data['conv_arr_lengths'][bi] conv_rand_mask[:end-start,bi,:] = rand_mask[bi,start:end,:] rand_mask = self._cuda(rand_mask) conv_rand_mask = self._cuda(conv_rand_mask) conv_story = data['conv_arr'] * conv_rand_mask.long() story = data['context_arr'] * rand_mask.long() else: story, conv_story = data['context_arr'], data['conv_arr'] # Encode dialog history and KB to vectors #encoder为modules中的ContextRNN，extKnow为modules中的ExternalKonwledge dh_outputs, dh_hidden = self.encoder(conv_story, data['conv_arr_lengths']) global_pointer, kb_readout = self.extKnow.load_memory(story, data['kb_arr_lengths'], data['conv_arr_lengths'], dh_hidden, dh_outputs) #cat函数会在给定维度上对输入的张量序列进行连接操作。 #这里将kb_memory和对话历史的信息联合到一起，组成外部知识 encoded_hidden = torch.cat((dh_hidden.squeeze(0), kb_readout), dim=1) # Get the words that can be copy from the memory #准备kbmemory中可以拷贝的宾语列表 batch_size = len(data['context_arr_lengths']) self.copy_list = [] for elm in data['context_arr_plain']: elm_temp = [ word_arr[0] for word_arr in elm ] self.copy_list.append(elm_temp) #decoder(LocalMemoryDecoder)生成output语句 outputs_vocab, outputs_ptr, decoded_fine, decoded_coarse = self.decoder.forward( self.extKnow, story.size(), data['context_arr_lengths'], self.copy_list, encoded_hidden, data['sketch_response'], max_target_length, batch_size, use_teacher_forcing, get_decoded_words, global_pointer) return outputs_vocab, outputs_ptr, decoded_fine, decoded_coarse, global_pointer

3.2.2 Evaluate

def evaluate(self, dev, matric_best, early_stop=None): print("STARTING EVALUATION") # Set to not-training mode to disable dropout #因为train函数的默认为True，即开启dropout self.encoder.train(False) self.extKnow.train(False) self.decoder.train(False) ref, hyp = [], [] acc, total = 0, 0 dialog_acc_dict = {} F1_pred, F1_cal_pred, F1_nav_pred, F1_wet_pred = 0, 0, 0, 0 F1_count, F1_cal_count, F1_nav_count, F1_wet_count = 0, 0, 0, 0 pbar = tqdm(enumerate(dev),total=len(dev)) new_precision, new_recall, new_f1_score = 0, 0, 0 #读入数据 if args['dataset'] == 'kvr': with open('data/KVR/kvret_entities.json') as f: global_entity = json.load(f) global_entity_list = [] for key in global_entity.keys(): if key != 'poi': global_entity_list += [item.lower().replace(' ', '_') for item in global_entity[key]] else: for item in global_entity['poi']: global_entity_list += [item[k].lower().replace(' ', '_') for k in item.keys()] global_entity_list = list(set(global_entity_list)) for j, data_dev in pbar: # Encode and Decode _, _, decoded_fine, decoded_coarse, global_pointer = self.encode_and_decode(data_dev, self.max_resp_len, False, True) decoded_coarse = np.transpose(decoded_coarse) decoded_fine = np.transpose(decoded_fine) for bi, row in enumerate(decoded_fine):#各种计算 st = '' for e in row: if e == 'EOS': break else: st += e + ' ' st_c = '' for e in decoded_coarse[bi]: if e == 'EOS': break else: st_c += e + ' ' pred_sent = st.lstrip().rstrip() pred_sent_coarse = st_c.lstrip().rstrip() gold_sent = data_dev['response_plain'][bi].lstrip().rstrip() ref.append(gold_sent) hyp.append(pred_sent) if args['dataset'] == 'kvr': # compute F1 SCORE #计算F-评论的结果 single_f1, count = self.compute_prf(data_dev['ent_index'][bi], pred_sent.split(), global_entity_list, data_dev['kb_arr_plain'][bi]) F1_pred += single_f1 F1_count += count single_f1, count = self.compute_prf(data_dev['ent_idx_cal'][bi], pred_sent.split(), global_entity_list, data_dev['kb_arr_plain'][bi]) F1_cal_pred += single_f1 F1_cal_count += count single_f1, count = self.compute_prf(data_dev['ent_idx_nav'][bi], pred_sent.split(), global_entity_list, data_dev['kb_arr_plain'][bi]) F1_nav_pred += single_f1 F1_nav_count += count single_f1, count = self.compute_prf(data_dev['ent_idx_wet'][bi], pred_sent.split(), global_entity_list, data_dev['kb_arr_plain'][bi]) F1_wet_pred += single_f1 F1_wet_count += count else: # compute Dialogue Accuracy Score #计算对话准确性结果 current_id = data_dev['ID'][bi] if current_id not in dialog_acc_dict.keys(): dialog_acc_dict[current_id] = [] if gold_sent == pred_sent: dialog_acc_dict[current_id].append(1) else: dialog_acc_dict[current_id].append(0) # compute Per-response Accuracy Score #计算每个回应的准确性 total += 1 if (gold_sent == pred_sent): acc += 1 if args['genSample']: self.print_examples(bi, data_dev, pred_sent, pred_sent_coarse, gold_sent) # Set back to training mode #开启dropput防止过拟合 self.encoder.train(True) self.extKnow.train(True) self.decoder.train(True) bleu_score = moses_multi_bleu(np.array(hyp), np.array(ref), lowercase=True) acc_score = acc / float(total) print("ACC SCORE:\t"+str(acc_score)) if args['dataset'] == 'kvr': F1_score = F1_pred / float(F1_count) print("F1 SCORE:\t{}".format(F1_pred/float(F1_count))) print("\tCAL F1:\t{}".format(F1_cal_pred/float(F1_cal_count))) print("\tWET F1:\t{}".format(F1_wet_pred/float(F1_wet_count))) print("\tNAV F1:\t{}".format(F1_nav_pred/float(F1_nav_count))) print("BLEU SCORE:\t"+str(bleu_score)) else: dia_acc = 0 for k in dialog_acc_dict.keys(): if len(dialog_acc_dict[k])==sum(dialog_acc_dict[k]): dia_acc += 1 print("Dialog Accuracy:\t"+str(dia_acc*1.0/len(dialog_acc_dict.keys()))) if (early_stop == 'BLEU'): if (bleu_score >= matric_best): self.save_model('BLEU-'+str(bleu_score)) print("MODEL SAVED") return bleu_score elif (early_stop == 'ENTF1'): if (F1_score >= matric_best): self.save_model('ENTF1-{:.4f}'.format(F1_score)) print("MODEL SAVED") return F1_score else: if (acc_score >= matric_best): self.save_model('ACC-{:.4f}'.format(acc_score)) print("MODEL SAVED") return acc_score

3.2.3 Compute_prf

def compute_prf(self, gold, pred, global_entity_list, kb_plain): local_kb_word = [k[0] for k in kb_plain] TP, FP, FN = 0, 0, 0 if len(gold)!= 0: count = 1 for g in gold: if g in pred: TP += 1 else: FN += 1 for p in set(pred): if p in global_entity_list or p in local_kb_word: if p not in gold: FP += 1 #计算准确率 precision = TP / float(TP+FP) if (TP+FP)!=0 else 0 #计算召回率 recall = TP / float(TP+FN) if (TP+FN)!=0 else 0 #F-评价，综合准确率和召回率的评价指标 F1 = 2 * precision * recall / float(precision + recall) if (precision+recall)!=0 else 0 else: precision, recall, F1, count = 0, 0, 0, 0 return F1, count