Avoid repeating kernel - Githubissues

f-dangel / unfoldNd

(N=1,2,3)-dimensional unfold (im2col) and fold (col2im) in PyTorch

MIT License

82 stars 6 forks source link

Avoid repeating kernel #12

Closed f-dangel closed 3 years ago

f-dangel commented 3 years ago

Resolves #11.

[x] Merge, then re-run benchmarks
[x] Depending on the outcome, think about reverting back to the old strategy. Maybe adapt Performance section in README

f-dangel commented 3 years ago

In contrast to my expectations, both RAM consumption and run time increased quite significantly. Before reverting to the old strategy, it might make sense to try an alternative approach which treats the batch dimension like a channel dimension and leads to the convolution using even more groups.