Incorrect number of dimensions issue

egr95 / R-codacore

An R package for learning log-ratio biomarkers from high-throughput sequencing data.

Other

21 stars 3 forks source link

Incorrect number of dimensions issue #18

Open DrinnanSante opened 1 year ago

DrinnanSante commented 1 year ago

Hi, I really like this idea, I have been reading about for a few weeks now, thanks for all the work!!

I have been trying to run it on our data and I keep getting this error:

model <- codacore(codafinal, z) Error in x[, cdbl$hard$numerator, drop = FALSE] : incorrect number of dimensions In addition: Warning messages: 1: glm.fit: fitted probabilities numerically 0 or 1 occurred 2: glm.fit: fitted probabilities numerically 0 or 1 occurred

I get it whether I increase or decrease the dataset, no else?

Thanks!

egr95 commented 1 year ago

You are welcome! It looks like the optimization might be running into some numerical issues on your dataset. Some pointers that can help you debug:

Check that the inputs are in the right format. What happens if you replace z with a random vector of 0s and 1s?
Tinker with the optimization parameters optParams. For instance, does the error occur if you use just 1 epoch, or if you decrease the learning rate?
Tinker with the cross-validation parameters cvParams. You can see more details on these in help(codacore) or directly in the codebase.

I would also be happy to take a quick look and help debug if you are open to sending me the data.

Glfrey commented 1 year ago

Hi @egr95,

I also found I had this issue and found it was only solvable by labelling an extra negative instance (i.e. incorrectly labelling a positive instance as negative). I couldn't solve it through any other permutation of parameters and I ensured the inputs were in the correct format. Could this error potentially be symptom of something wrong with either class imbalance or too few instances?