A multimodal network trains unstably with very large gradient norms. Which action is appropriate to evaluate?

1. Removing the validation set
2. Learning-rate adjustment and justified gradient clipping
3. Increasing every loss term without diagnosis
4. Selecting the model by its filename