Does reducing batch size affects convergence?

[pinakinathc]

Due to a limit of 11GB GPU resource, I am forced to reduce the batch size for image2text to 8 instead of 128.
After training for a while, I was unable to get a converged trained model that replicated results in the paper.

Hence, my question is:

1. How much GPU memory did you use?
2. Does using smaller batch size leads to failure in training?

[pinakinathc]

@s-mahajan can you upload the trained checkpoints as mentioned in the README.md file?