Open izaakniksan opened 4 years ago
[ ] Go over rubric
[ ] Explain that sometimes cache size < sum of all the totals because tensors come in and out of existence at different times. So all won't exist at once.
[ ] Explain why in resnet, the current cache is much smaller than the max cache. Also, CUDA driver takes up hundreds of MB when initialized (https://github.com/facebookresearch/maskrcnn-benchmark/issues/182)
[ ] Explain stack trace of building from source, show how I searched
[x] Acknowledgements : how to format it? Put it right after abstract?
[x] make sure all listings are labelled
[x] Conclusion
[x] In section 4.5, change activation to feature map
[x] Put all gathered data tables. Remove "iteration #, epoch#' from the table titles, and just explain that I waited until steady state.
[x] Future work (profile more models, extend to multi-gpu [say that I couldn't do this because I don't have access to such a setup])
[x] Talk about why gradients increase in mem usage (what are intermediate gradients?)
[x] Cite the resnet paper/implementation/imagenet dataset in appendix
[x] Explain memory profiler (step by step, how hooks work, etc.)
[x] Change tables to isolate intermediate grads on their own, since i didn't include them in the bar graphs
[x] Get more recommendation data
[x] System diagram on powerpoint
[x] Add official cover page
Looks good man
tygamvrelis
commented
4 years ago
[ ] Go over rubric
[ ] Explain that sometimes cache size < sum of all the totals because tensors come in and out of existence at different times. So all won't exist at once.
[ ] Explain why in resnet, the current cache is much smaller than the max cache. Also, CUDA driver takes up hundreds of MB when initialized (https://github.com/facebookresearch/maskrcnn-benchmark/issues/182)
[ ] Explain stack trace of building from source, show how I searched
[x] Acknowledgements : how to format it? Put it right after abstract?
[x] make sure all listings are labelled
[x] Conclusion
[x] In section 4.5, change activation to feature map
[x] Put all gathered data tables. Remove "iteration #, epoch#' from the table titles, and just explain that I waited until steady state.
[x] Future work (profile more models, extend to multi-gpu [say that I couldn't do this because I don't have access to such a setup])
[x] Talk about why gradients increase in mem usage (what are intermediate gradients?)
[x] Cite the resnet paper/implementation/imagenet dataset in appendix
[x] Explain memory profiler (step by step, how hooks work, etc.)
[x] Change tables to isolate intermediate grads on their own, since i didn't include them in the bar graphs
[x] Get more recommendation data
[x] System diagram on powerpoint
[x] Add official cover page