Hi, thanks for your code.
When set train_input to noise, I find that all the predicted results with teacher forcing are \<s> during inference. It seems that the training process crashed.
After trying, I found that changing the random seed can improve this situation. The random seed has a great impact on the training process. Some random seeds will cause training failure.
Hi, thanks for your code. When set train_input to noise, I find that all the predicted results with teacher forcing are \<s> during inference. It seems that the training process crashed.
After trying, I found that changing the random seed can improve this situation. The random seed has a great impact on the training process. Some random seeds will cause training failure.