Interested in working on the project through GSoC and more

[gucifer]

Hello there,
I am highly interested in working on the metrics module of this project and have started reading through the code base. I am a final year student in computer science.
I have been doing research work in Image classification and data noise reduction for over a year. Currently, I am doing a thesis on Text Summarization and NLP. So I am fairly familiar with ROUGE, BLEU, METEOR and others.
I am also reading up on distributed implementation for the same. Any good resource other than the pytorch documentation that might help?
I am facing issue with string operations. Since pytorch does not have a tensor representation for list of strings or similar, the step of counting common strings(n-grams) does not seem to have a parallel alternative, unless each character is encoded using corresponding ASCII value as mentioned here.

[vfdev-5]

Thanks for your interest in the project and GSoC application we submit !
Working on metics module would be really awesome ! Apart of GSoC way to collaborate we could also explore other ways (please, contact me on discord vfdev-5#4458 to discuss those modalities).

I am facing issue with string operations. Since pytorch does not have a tensor representation for list of strings or similar, the step of counting common strings(n-grams) does not seem to have a parallel alternative, unless each character is encoded using corresponding ASCII value as mentioned here.

I think string encoding into numbers is a good approach. Currently, in ignite we provide a helper broadcast and all_gather methods to work on strings:

• https://pytorch.org/ignite/distributed.html#ignite.distributed.utils.broadcast
• https://pytorch.org/ignite/distributed.html#ignite.distributed.utils.all_gather

[gucifer]

Hi there,
I send you a friend request on discord, my id is kana#7979. Let us discuss further on there!

[sdesrozis]

Hi @gucifer ! Thank you for your interest 😊 and good luck for your thesis, we definitely need guys on NLP !

[gucifer]

Hello there,
I have done a basic implementation of ROUGE Score, commit for the same can be seen on my fork of ignite. It is a simple implementation and does not follow DDP, will add support for it once I learn how to.
Also, what is the mathematical notation being followed for the comments? Anything else I should be aware of before i open a PR?