Closed l0renor closed 2 years ago
Thanks for the quick reply. This unfortunately doesnβt fix the issue for me. In #29 the problem got resolved by setting pytorch-lightning verision to 1.1.7. I was already using this version as specified in the requirements.txt.
As I ran into the same issue as #28 I changed the requirements.txt. For me it only starts training with the following file:
omegaconf==2.0.6
pytorch-lightning==1.1.7
hydra-core==1.0.6
transformers
neptune-client==0.5.1
psutil==5.8.0
datasets==1.3.0
rouge-score==0.0.4
sacrebleu==1.5.0
wandb==0.10.26
streamlit==0.82.0
Which versions would you recommend @LittlePea13?
with the latest commit the issue with the labels should be fixed, I will update the reqs with a newer version of datasets but it needs a small change to the dataset files.
Thank you!
Hi, I am trying to train the model on the doc red dataset in order to test the effects of labeling the entities with an additional special token.
At the moment I am still trying to get the code to run with the original dataset.
In the first epoch after 56% i get the
KeyError: 'labels'
in line 48, inon_train_batch_end labels = batch.pop("labels")
I checked the dataset for empty labels and found 27 empty arrays in the doc red data. Deleting data points didn't solve the problem. I also tested only using the first 50% of the dataset. The error still occurred at 56%.
full console output with print(batch) before the error: