Add scalar tensor operations #3127
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #3127 +/- ##
==========================================
- Coverage 81.13% 81.11% -0.02%
==========================================
Files 821 821
Lines 117950 117962 +12
==========================================
- Hits 95696 95688 -8
- Misses 22254 22274 +20 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
| let data = TensorData::new(alloc::vec![self], [1]); | ||
| let numerator = Tensor::<B, D, K>::from_data(data, &tensor.device()).unsqueeze(); | ||
| Tensor::div(numerator, tensor) |
There was a problem hiding this comment.
What about using recip and scalar multiply instead?
tensor.recip().mul_scalar(self)There was a problem hiding this comment.
recip() is only for Float tensors, but then that's probably fine! I've changed it to use that now and implemented it for f32 and f64
There was a problem hiding this comment.
Yeah I think it makes more sense for fractions anyway. We could always add int support back if warranted but it could be ambiguous.
|
Probably we should mention this in https://burn.dev/burn-book/building-blocks/tensor.html |
I've added a few lines. Didn't add notes that adding and multiplication are symmetric now as it just read a bit redundant, but lmk! |
While tensor scalar generally already existed, scalar tensor didn't.
This is especially annoying for operations like 1.0 / tensor or a common case like (1.0 - tensor).
Due to Rust's orphan rules these sadly can't be implemented for every E: ElementConversion, but manually implementing them for primitives seems to work fine!