Some small bugs when training videocrafter2

[Vincent-luo]

Thansks for this great project!

After doing some training with videocrafter2, I found some bugs.

1. can't train text encoder even if set cond_stage_trainable: true in config. Refer to these lines, text encoding process(get_learned_conditioning) is under torch.no_grad(), so no grad is computed. Simple fix is just remove the @torch.no_grad()

    @torch.no_grad()
    def get_batch_input(self, batch, random_uncond, return_first_stage_outputs=False, return_original_cond=False, is_imgbatch=False):
        ...

        if isinstance(cond, dict) or isinstance(cond, list):
            cond_emb = self.get_learned_conditioning(cond)
        else:
            cond_emb = self.get_learned_conditioning(cond.to(self.device))
        ...

2. in some shell scripts there is line like --auto_resume True(like this line), actually we don't need "True" here. auto_resume is a flag with no arguement, if it exist it's true, otherwise it's false.

3. same logging message printed twice, seems like the mainlogger inherits the root logger's handler. Simple fix is set " logger.propagate = False" in set_logger.