Jan Zdenek

It seems that you need gcc 6 or higher for `-fno-plt` option. I was running the code on two computers, Ubuntu 14.04 on which it's supposedly difficult to upgrade gcc...

Yeah, I would also like to know what is the F-score that you've achieved with the model

Thanks! I understand it now. ~~There seems to be an issue that made me confused about the blank label and space characters. When I transform the original ground truth using...

One more question. Does the argument `--timesteps`, which defaults to `3`, do anything when I use a curriculum where I have let's say paths to ground truth files containing 2...

Thanks for updating the README. I've found another likely issue in the code. In `baby_step_curriculum.py`, you use `deque` collection with maximum length set to 5 to store the validation losses,...

> Yes, you are right. This definitely is an issue in the code (it also explains why it did not work for me as expected)! I was thinking `pop` would...

> Sure, I used a batch size of 20 per GPU and a start learning rate of `1e-6`. You can also see those values, if you have a look at...

> In the code snippet you show here, you'll also need to be careful to check whether the last split has the same length as the first split, as the...

> Thanks for having a look at the code and playing around with it! Highly appreciate that. You're welcome. I might use it for my own research so I just...

Hi! I have a question. Did you actually use the full validation when training the model? I experience a weird bug caused by full validation. When I set the conditions...