search

nyu-dl / dl4marco-bert

BSD 3-Clause "New" or "Revised" License
476 stars 87 forks source link

Error in eval mode #35

Closed cristinadece closed 4 years ago

cristinadece commented 4 years ago

I am trying to run the "run_msmarco.py" script in eval mode (on my server, so not in Colab). I have replaced the uncased_L-24_H-1024_A-16 model with your BERT_Base_trained_on_MSMARCO model on the MSMARCO_tfrecord.tar.gz data, in plain words I'm trying to reproduce your results. However, I get the following error: ValueError: Shape of variable bert/embeddings/LayerNorm/beta:0 ((1024,)) doesn’t match with shape of tensor bert/embeddings/LayerNorm/beta ([768]) from checkpoint reader. at line 393 for item in result:.

It seems like the expected dimension doesn't match the one received. Do you have any suggestions of what might be the problem and how it can be fixed?

rodrigonogueira4 commented 4 years ago

Hi, did you replace --bert_config_file=${DATA_DIR}/uncased_L-24_H-1024_A-16/bert_config.json, accordingly? The correct json config file can be found here: https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip

(Sorry, I will add this file to BERT_Base_trained_on_MSMARCO.zip)

rodrigonogueira4 commented 4 years ago

I added bert_config.json to BERT_Base_trained_on_MSMARCO.zip. You should not see the error anymore if you pass it to --bert_config_file

cristinadece commented 4 years ago

That was the problem. Thanks for fixing it.

rodrigonogueira4 commented 4 years ago

Great, thanks

cristinadece commented 4 years ago

I added bert_config.json to BERT_Base_trained_on_MSMARCO.zip. You should not see the error anymore if you pass it to --bert_config_file

Could you please add the config and vocab files to the BERT_Large_MSMARCO and BERT_Large_trained_on_TREC_CAR models?