crf.py里维特比back_points初始化问题

[Coga8]

你好，我有一些疑问

back_points, index_points = [self.transitions[self.start_tag, :self.start_tag].cpu()], [torch.LongTensor([-1]).expand_as(forward)] 在这里back_points初始化时只考虑了start_tag到各个tag的转移概率，是不是少加了1st word到各个tag的发射概率

还有就是

        for i in range(1, time_steps):  # START_TAG -> 1st word -> 2nd word ->...->END_TAG
            emission_start = forward.expand(self.num_tag, self.num_tag).t()
            emission_end = features[i, :].expand(self.num_tag, self.num_tag)

i = 1时，因为forward为全0，所以emission_start在这里为全0，但我感觉应该先给forward赋值back_points[0]吧