Weighted graph preprocessing to enable biased sampling #38
Conversation
Codecov Report
@@ Coverage Diff @@
## master #38 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 7 9 +2
Lines 116 194 +78
=========================================
+ Hits 116 194 +78
Continue to review full report at Codecov.
|
| TORCH_CHECK(rowptr.is_cpu(), "'rowptr' must be a CPU tensor"); | ||
| TORCH_CHECK(bias.is_cpu(), "'bias' must be a CPU tensor"); | ||
|
|
||
| auto cdf = at::empty_like(bias); |
There was a problem hiding this comment.
| auto cdf = at::empty_like(bias); | |
| const auto cdf = at::empty_like(bias); |
There was a problem hiding this comment.
We need to write data to cdf so it's not a const here
There was a problem hiding this comment.
The tensor will always be a const since we only change its data, not its metadata. It‘s not so important though :)
| avg /= len; | ||
|
|
||
| // The sets for index with a bias lower or higher than average | ||
| std::vector<std::pair<int64_t, scalar_t>> high, low; |
There was a problem hiding this comment.
One more optimization we can do is to use std::vector<size_t> high, low instead of pair.
Later you can do
scalar_t low_bias = out_beg[low_idx];
out_beg[low_idx] = low_bias / avg;
out_beg[high_idx] -= (avg - low_bias);
Create pairs are much more costly than just use int.
No description provided.