Error "Attempt to call single feature writer on packed feature writer" with embedding features on GPU
highly0 opened this issue · 2 comments
Problem:
I'm trying to train a regression task with a dataset that includes numerical, text, and embedding features (which includes sentence embeddings using MPNET and TFIDF vectors; all in numpy arrays format). Previously, I was able to train the same regression model without the embedding features (If I dropped all embedding features from X_train and X_test, training is successful and very fast). However, when I included the embedding features, I get this error "Attempt to call single feature writer on packed feature writer" (it took 20 minutes for training on GPU to produce this error on verbose=1
). Initially, I had the paramverbose=200
, so training failed hours after I started it. Any insights as to why this error popped up?
UPDATE: training on CPU on the dataset with numerical, text, and embedding features was succesful. It seems like there is a bug with the GPU training for embedding features?
catboost version: 1.2.2
GPU: NVIDIA GeForce RTX 3090
Having the same problem. Asked question on stackoverflow yesterday: https://stackoverflow.com/questions/77987029/getting-catboosterror-attempt-to-call-single-feature-writer-on-packed-feature
Version: 1.2.2
GPU: A100/V100
Unfortunately, I don't have access to the older versions of catboost now, but I definitely remember that there was no such problem on one of the older versions. If you have the opportunity, I advise you to try to install other version of catboost package (for example - 1.1)