[深度学习] loss=inf

原因一：data underflow

原因: 在使用log函数时出现 data underflow

解决方法：增加一个bias

# original
g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
g_wh = torch.log(g_wh) / variances[1]

# current
eps = 1e-5
g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
g_wh = torch.log(g_wh+eps) / variances[1]

https://github.com/amdegroot/ssd.pytorch/issues/162

https://www.zhihu.com/question/49346370

原因二：CTC length不符合要求

原因：input length < target CTC length

ctc loss要求input length > target CTC length, 注意 target CTC length 不同于target length，两者区别如下:

def ctc_len(label):
    add_len = 0
    label_len = len(label)
    for i in range(label_len - 1):
        if label[i] == label[i + 1]:
            add_len += 1  # 这里+1是因为ctc会在重复字符之间填充blank
    return label_len + add_len

target_length = len(label)
target_ctc_len = ctc_len(label)

https://github.com/xingchensong/ASR-Wavnet/blob/master/datafeeder.py#L114