Closed ambroser53 closed 8 months ago
Our training pipeline calls the batchify
to collate the batch of data (refer to the CustomTrainer
). It seems that your training pipeline calls the default data_collator of huggingface trainer.
This fixed the issue thank you very much.
I'm trying to use your training script but struggle to get past a single loop in the trainer because the trainer within the collation function stacks every element in the batch it collates. However, the patch positions being passed are of varying sizes (most are [13,2] but some are [16,2]). Here is the full error:
Is this a known issue? I am training on custom data but have not changed any of the original code. What can I change to make this work? Will I simply have to run with a batch size of 1?