Closed cnedwards closed 1 year ago
Hi Carl, the assumed input for the y_pred in group.evaluate(y_pred) has the same order of the fixed test split in the test
variable from the benchmark group. TDC fixes the test set for a fair comparison.
I suspect that the MLP is worse than the constant is due to the bad performance of MLP on this task.
Closing for now! Feel free to reopen if you have any further question!
Describe the bug Ordering of input to group evaluation for drug synergy tasks is unclear.
This effects the MLP example (https://github.com/mims-harvard/TDC/blob/main/examples/multi_pred/drugcombo/train_MLP.py), since it uses a dataloader with shuffling on.
To Reproduce The following code should show the behavior:
This outputs the following on my machine:
Note that I modified the code in the library to round to more decimal places. By default it is 3, which makes the function look permutation invariant (e.g. both 29854.6109218136 and 29854.610926591966 round to 29854.611).
Expected behavior The input to the evaluate method should require the permutation to be specified so it can be correctly compared to the ground truth.
Environment:
If I just missed the correct usage of the group.evaluate() method please let me know. Thanks!