Mehran Maghoumi

Hi @toinsson Sorry for delays in my responses and PR reviews, I've been tied up with my day job over the past few weeks. It's great to see you took...

Thanks for this great contribution, really appreciate it! :) I need some time to study and verify it. In the mean time, could you explain this a bit more? >...

Thanks for creating this PR! The goal is to basically expand `(x-y)^2` to `x^2 + y^2 - 2xy`, right? I noticed an error when running this change. What is `y.mT`?...

Hello, and thanks for your interest in my work! Yes, the large memory footprint is expected. The issue partly stems from the way the cost map is calculated, since we'd...

Thanks for posting and example @educationunion! This issue was previously reported in #8 but I didn't have a good minimal example to try. I had a quick look, and I...

Yes, this was previously discussed in #11. I already have an implementation available locally (which uses `PackedSequence`) to handle inputs of different lengths, but the performance is quite poor and...

@taolusi I haven't tried distributed training myself, so I'm unsure whether this works. What kind of issues are you facing? Is it running into CUDA errors?

I will investigate this when I get a chance. I have a hunch the problem is due to CUDA context across different devices. Not sure how Numba handles multiple contexts.

Thanks for reporting this issue @LuckerYi. Can you provide a bit more details? What are the lengths of your sequences? Does it happen with all bandwidth sizes, or just particular...

Thanks for your response. What was confusing to me originally was the fact that your CRF layer is actually a loss that one can minimize, whereas other PyTorch implementations had...