One Static Graph Feature May be Required
- Dominant language
- C++
- Stars
- 3.4k
- Forks
- 701
- PR merge metrics
- No merged PRs in 30d
Description
When I am working on #1290, I find that it takes ~27 times to construct a cnn network than real calculation (forward and backward) on 1 GPU. I do not know whether it is a normal case. I may compare with some other frameworks later.
But I think it will be good to modify dy.pick API so that static graph can be built. Currently, the 'idx' field in pick can not be a placeholder like inputTensor, so loss related to pick can not be built statically.
Contributor guide
No contributing guide indexed for this repository
Research direction
Start with the dy.pick API and trace how its idx argument differs from inputTensor during static graph construction. Check whether existing tests cover pick with dynamic inputs; done means idx can be represented as a placeholder and pick-related loss computation can be built statically.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100