[Hackathon 7th] 修复 vctk 的 ernie_sat 训练时出现的类型提升问题

[megemini]

PR types

Bug fixes

PR changes

Others

Describe

修复 vctk 的 ernie_sat 训练时出现的类型提升问题：

aistudio@jupyter-942478-8626068:~/PaddleSpeech/examples/vctk/ernie_sat$ CUDA_VISIBLE_DEVICES=0,1 ./local/train.sh conf/default.yaml ./output
...

Traceback (most recent call last):
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddle/distributed/spawn.py", line 385, in _func_wrapper
    result = func(*args)
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/exps/ernie_sat/train.py", line 167, in train_sp
    trainer.run()
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/training/trainer.py", line 203, in run
    six.reraise(*exc_info)
  File "/usr/lib/python3/dist-packages/six.py", line 703, in reraise
    raise value
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/training/trainer.py", line 149, in run
    update()
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/training/updaters/standard_updater.py", line 110, in update
    self.update_core(batch)
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/models/ernie_sat/ernie_sat_updater.py", line 70, in update_core
    mlm_loss, text_mlm_loss = self.criterion(
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddle/nn/layer/layers.py", line 1531, in __call__
    return self.forward(*inputs, **kwargs)
  File "/home/aistudio/PaddleSpeech/paddlespeech/t2s/modules/losses.py", line 1121, in forward
    mlm_loss = paddle.sum((loss * paddle.reshape(
TypeError: (InvalidType) Type promotion only support calculations between floating-point numbers and between complex and real numbers. But got different data type x: float32, y: bool. (at ../paddle/phi/common/type_promotion.h:234)

@zxcd @Liyulingyue @GreatV @enkilee @yinfan98

[paddle-bot]

Thanks for your contribution!

[zxcd]

convert mlm_loss_pos dtype first to reduce duplicate codes.

[zxcd]

LGTM