AzureML-BERT icon indicating copy to clipboard operation
AzureML-BERT copied to clipboard

Bert Data for Pretraining: No such file or directory: 'bert_data/validation_512_only'

Open nigaregr opened this issue 6 years ago • 5 comments

Hi, I have Pretraining running but it fails after 1st Epoch with the following error: File "/AzureML-BERT/pretrain/PyTorch/dataset.py", line 100, in init path = get_random_partition(self.dir_path, index) File "/AzureML-BERT/pretrain/PyTorch/dataset.py", line 33, in get_random_partition for x in os.listdir(data_directory)] FileNotFoundError: [Errno 2] No such file or directory: 'bert_data/validation_512_only'

I have the created the Wiki pretraining data using create_pretraining script. I do not see validation_512_only being generated?

nigaregr avatar Sep 18 '19 17:09 nigaregr

I think you should create another subfolder in bert_data/validation_512_only with the validation data (i.e .bin files generated by create_pretraining) in it

kishorepv avatar Sep 24 '19 17:09 kishorepv

Thanks @nigaregr for reporting this. @jingyanwangms can you update the tar file mentioned in https://github.com/microsoft/AzureML-BERT/blob/master/docs/artifacts.md#preprocessed-data with the newly generated wikipedia dataset and the validation folder?

skaarthik avatar Sep 25 '19 05:09 skaarthik

For now I created bert_data/validation_512_only folder and moved wikipedia_segmented_part_98.bin and it seems the training pipeline is working fine.

Still would be great to use the updated files @jingyanwangms

usuyama avatar Jan 06 '20 03:01 usuyama

Hi @skaarthik, have you decided to update the zip-dataset or the data prep instruction? Besides, I wonder what if I did as @usuyama suggested? Will there be any performance influence/drop? Thanks!

Howal avatar Apr 13 '20 11:04 Howal

Hi @Howal, what @usuyama did is a reasonable workaround in the absence of some other validation set.

skaarthik avatar Apr 18 '20 23:04 skaarthik