Accuracy decreases with iterations

[mhasnat]

Hi @wanji

Thank you for uploading your contributions in this repository. With the given examples, they are quite easy to understand. I would like to know your feedback on my experiment where I observe the decrease in accuracy with iterations.

I started the training with the "naive" version (given all triplets by myself) and fine-tune from an existing caffemodel. Results shows that the accuracy decreases with iterations.

It would be helpful to know if you have encountered similar issue. Moreover, do you think going to the "batch" version from the "naive" version can be a solution to this problem?

Thanks a lot in advance.