def update(self, batch):
# Train mode
self.network.train()
self.network.drop_emb = True
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert, x_bert_mask, x_bert_offsets,\
query, query_mask, query_char, query_char_mask, query_bert, query_bert_mask, query_bert_offsets, \
ground_truth, context_str, context_words, _, _, _, _ = batch
# Run forward
# score_s, score_e: batch x context_word_num
# score_yes, score_no, score_no_answer: batch x 1
score_s, score_e, score_yes, score_no, score_no_answer = self.network(
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert,
x_bert_mask, x_bert_offsets, query, query_mask, query_char,
query_char_mask, query_bert, query_bert_mask, query_bert_offsets,
len(context_words))
max_len = self.opt['max_len'] or score_s.size(1)
batch_size = score_s.shape[0]
context_len = score_s.size(1)
expand_score = gen_upper_triangle(score_s, score_e, max_len,
self.use_cuda)
scores = torch.cat(
(expand_score, score_no, score_yes, score_no_answer),
dim=1) # batch x (context_len * context_len + 3)
targets = []
# print("target length:{}".format(len(targets)))
span_idx = int(context_len * context_len)
for i in range(ground_truth.shape[0]):
if ground_truth[i][0] == -1 and ground_truth[i][
1] == -1: # no answer
targets.append(span_idx + 2)
if ground_truth[i][0] == 0 and ground_truth[i][1] == -1: # no
targets.append(span_idx)
if ground_truth[i][0] == -1 and ground_truth[i][1] == 0: # yes
targets.append(span_idx + 1)
if ground_truth[i][0] != -1 and ground_truth[i][
1] != -1: # normal span
targets.append((ground_truth[i][0] * context_len +
ground_truth[i][1]).item())
targets = torch.LongTensor(np.array(targets))
if self.use_cuda:
targets = targets.cuda()
loss = self.loss_func(scores, targets)
self.train_loss.update(loss.item(), 1)
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(self.network.parameters(),
self.opt['grad_clipping'])
self.optimizer.step()
self.updates += 1
if 'TUNE_PARTIAL' in self.opt:
self.network.vocab_embed.weight.data[
self.opt['tune_partial']:] = self.network.fixed_embedding
def predict(self, batch):
self.network.eval()
self.network.drop_emb = False
# Run forward
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert, x_bert_mask, x_bert_offsets, query, query_mask, \
query_char, query_char_mask, query_bert, query_bert_mask, query_bert_offsets, ground_truth, context_str, context_words, \
context_word_offsets, answers, context_id, turn_ids, x_prev_ans_mask = batch
context_len = len(context_words)
score_s, score_e, score_yes, score_no, score_no_answer = self.network(x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert, x_bert_mask, x_bert_offsets,
query, query_mask, query_char, query_char_mask, query_bert, query_bert_mask, query_bert_offsets, len(context_words), x_prev_ans_mask)
batch_size = score_s.shape[0]
max_len = self.opt['max_len'] or score_s.size(1)
expand_score = gen_upper_triangle(score_s, score_e, max_len, self.use_cuda)
scores = torch.cat((expand_score, score_no, score_yes, score_no_answer), dim=1) # batch x (context_len * context_len + 3)
prob = F.softmax(scores, dim = 1).data.cpu() # Transfer to CPU/normal tensors for numpy ops
# Get argmax text spans
predictions = []
confidence = []
pred_json = []
for i in range(batch_size):
_, ids = torch.sort(prob[i, :], descending=True)
idx = 0
best_id = ids[idx]
confidence.append(float(prob[i, best_id]))
if best_id < context_len * context_len:
st = best_id / context_len
ed = best_id % context_len
st = context_word_offsets[st][0]
ed = context_word_offsets[ed][1]
predictions.append(context_str[st:ed])
if best_id == context_len * context_len:
predictions.append('no')
if best_id == context_len * context_len + 1:
predictions.append('yes')
if best_id == context_len * context_len + 2:
predictions.append('unknown')
pred_json.append({
'id': context_id,
'turn_id': turn_ids[i],
'answer': predictions[-1]
})
return (predictions, confidence, pred_json) # list of strings, list of floats, list of jsons
def predict(self, batch):
"""
SDNet每更新1500个batch就会利用测试函数predict()在验证集上预测答案并计算准确率得分。predict()函数
的流程与update()函数类似，也要进行一次前向计算得到网络输出结果。之后，模型在所有可能的答案中选择概率
最大的作为预测结果。最终输出包括预测答案和对应的概率，并按照CoQA的要求输出JSON格式的结果。
"""
self.network.eval() # 将网络设置成测试模式，即不计算导数、不进行Dropout等操作
self.network.drop_emb = False
# 与update()函数类似，前向计算得到网络预测结果
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert, x_bert_mask, x_bert_offsets, \
query, query_mask, query_char, query_char_mask, query_bert, query_bert_mask, query_bert_offsets, \
ground_truth, context_str, context_words, context_word_offsets, answers, context_id, turn_ids = batch
context_len = len(context_words)
score_s, score_e, score_yes, score_no, score_no_answer = self.network(
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert,
x_bert_mask, x_bert_offsets, query, query_mask, query_char,
query_char_mask, query_bert, query_bert_mask, query_bert_offsets,
len(context_words))
batch_size = score_s.shape[0]
max_len = self.opt['max_len'] or score_s.size(1)
# 与update()函数类似，得到大小为m*m+3的一维概率向量
expand_score = gen_upper_triangle(score_s, score_e, max_len,
self.use_cuda)
scores = torch.cat(
(expand_score, score_no, score_yes, score_no_answer),
dim=1) # batch * (m * m + 3)
prob = F.softmax(scores, dim=1).data.cpu() # 将结果存入CPU，方便NumPy操作
predictions = [] # 存储预测的答案字符串
confidence = [] # 存储预测的概率
pred_json = [] # 存储JSON格式的答案
for i in range(batch_size):
_, ids = torch.sort(prob[i, :],
descending=True) # 对第i个答案的所有可能解的概率从大到小排序，只取索引
idx = 0
best_id = ids[idx] # best_id是概率最大答案的下标，在0到m*m+2之间
confidence.append(float(prob[i, best_id]))
# 处理答案是区间的情况，将best_id还原成开始位置st和结束位置ed
if best_id < context_len * context_len:
st = best_id / context_len
ed = best_id % context_len
# context_word_offsets提供每个词的第一个字符和最后一个字符在文章中的位置
st = context_word_offsets[st][0]
ed = context_word_offsets[ed][1]
# 获得预测的答案字符串
predictions.append(context_str[st:ed])
# 处理答案为“否”的情况
if best_id == context_len * context_len:
predictions.append('no')
# 处理答案为“是”的情况
if best_id == context_len * context_len + 1:
predictions.append('yes')
# 处理“没有答案”的情况
if best_id == context_len * context_len + 2:
predictions.append('unknown')
# 记录JSON格式的输出
pred_json.append({
'id': context_id,
'turn_id': turn_ids[i],
'answer': predictions[-1]
})
return (predictions, confidence, pred_json
) # list of strings, list of floats, list of jsons
def update(self, batch):
"""
训练函数train()最后调用了前向计算函数update()。该函数根据批次数据batch中的内容直接与SDNet网络代码对接，
进行一次前向计算，然后计算交叉熵损失函数，利用PyTorch自带的反向传播函数backward求导并更新参数。由于
CoQA任务的答案可能是文章中的一段区间，也有可能是“是/否/没有答案”，因此update()对所有概率进行统一处理：
如果文章中有m个单词，update()根据网络输出层结果生成一个长度为m^2+3的向量SCOYes，表示答案是各种可能的
文章区间与3种特殊情况的概率。
"""
self.network.train(1) # 进入训练模式
self.network.drop_emb = True
# 从batch中获得文章、问题、答案的所有信息，包括单词编号、词性标注、BERT分词编号等
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert, x_bert_mask, x_bert_offsets, \
query, query_mask, query_bert, query_char, query_char_mask, query_bert_mask, query_bert_offsets, \
ground_truth, context_str, context_words, _, _, _, _ = batch
# 进行前向计算，获得模型预测答案
# 1) 在文本每个位置开始和结束的概率score_s, score_e
# 2) 是Yes/No/No answer的概率为score_yes, score_no, score_no_answer
# 其中score_s和score_e的维度为batch * context_word_num, score_yes, score_no, score_no_answer的维度均为batch * 1
score_s, score_e, score_yes, score_no, score_no_answer = self.network(
x, x_mask, x_char, x_char_mask, x_features, x_pos, x_ent, x_bert,
x_bert_mask, x_bert_offsets, query, query_mask, query_char,
query_char_mask, query_bert, query_bert_mask, query_bert_offsets,
len(context_words))
# 答案最长长度在配置文件中有定义
max_len = self.opt['max_len'] or score_s.size(1)
batch_size = score_s.shape[0]
context_len = score_s.size(1)
expand_score = gen_upper_triangle(score_s, score_e, max_len,
self.use_cuda) # 区间答案的概率
# 将区间答案的概率与否/是/没有答案进行拼接
scores = torch.cat(
(expand_score, score_no, score_yes, score_no_answer),
dim=1) # batch * (context_len * context_len + 3)
# 标准答案的位置为转化成一维坐标，与expand_score对齐。例如：答案区间[3, 5]变成3*m+5, 否变成m*m, 是变成m*m+1, 没有答案变成m*m+2
targets = []
span_idx = int(context_len * context_len)
for i in range(ground_truth.shape[0]):
if ground_truth[i][0] == -1 and ground_truth[i][1] == -1: # 没有答案
targets.append(span_idx + 2)
if ground_truth[i][0] == 0 and ground_truth[i][1] == -1: # 否
targets.append(span_idx)
if ground_truth[i][0] == -1 and ground_truth[i][1] == 0: # 是
targets.append(span_idx + 1)
if ground_truth[i][0] != -1 and ground_truth[i][1] != -1: # 区间
targets.append(ground_truth[i][0] * context_len +
ground_truth[i][1])
targets = torch.LongTensor(np.array(targets))
if self.use_cuda:
targets = targets.cuda()
loss = self.loss_func(input=scores, target=targets) # 计算交叉熵损失函数
self.train_loss.update(loss.data[0], 1)
self.optimizer.zero_grad() # 优化器将所有导数清零
loss.backward() # 利用PyTorch自带的反向传播函数求导
torch.nn.utils.clip_grad_norm(parameters=self.network.parameters(),
max_norm=self.opt['grad_clipping'])
self.optimizer.step() # 更新参数
self.updates += 1
if 'TUNE_PARTIAL' in self.opt:
self.network.vocab_embed.weight.data[
self.opt['tune_partial']:] = self.network.fixed_embedding
