multi-outcome tabnet-fit and predict
#118
Conversation
switch target `y` from vector to array
manage loss cases for multi_output lint and refactor code
encode output_dim for multi-outcome improve multi-outcome classification loss split predict based on `is_multi_outcome`
switch to hardhat v1.3.0
fix tests vqlues add mixed-outcome and multi-outcome with valid test
move multi-output tests is a dedicated file
fix multi-outcome classification
cregouby
changed the title
tabnet-fit and predicttabnet-fit and predict
increase the GPU timeout
get rid of `library()` in tests remove typo in vignette
R/model.R
Outdated
| # if target is_multi_outcome, loss has to be applied to each label-group | ||
| if (max(batch$output_dim$shape) > 1) { | ||
| # TODO maybe torch_stack here would help loss$backward and better to shift right torch_sum at the end ? | ||
| outcome_nlevels <- as.numeric(batch$output_dim) |
There was a problem hiding this comment.
It seems that batch$output_dim is placed on the model device, thus as.numeric() doesn't work if the model is on CUDA.
| outcome_nlevels <- as.numeric(batch$output_dim) | |
| outcome_nlevels <- as.numeric(batch$output_dim$to(device="cpu")) |
|
GPU tests are recovered now that they run in a larger instance. I think we should move the batch_dim to cpu before |
|
You might need to re-add the env check in the testthat.R file when submitting to CRAN, so tests doesn't run on CRAN machines where torch is usually not installed. |
Codecov Report
@@ Coverage Diff @@
## main #118 +/- ##
==========================================
+ Coverage 87.50% 88.02% +0.52%
==========================================
Files 10 10
Lines 1072 1127 +55
==========================================
+ Hits 938 992 +54
- Misses 134 135 +1
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
tabnet_fitandpredict, relying on the fresh hardhat::spruce_*_multiple