What should I do if I want to use Bert-Base rather than Bert-Large? about language

[yanchlu]

Change the "bert-joint-baseline" to "uncased_L-12_H-768_A-12"?

[yanchlu]

InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Assign requires shapes of both tensors to match. lhs shape= [768] rhs shape= [1024]

---

Where does the '1024' comes from?!

python -m language.question_answering.bert_joint.prepare_nq_data
--logtostderr
--input_jsonl=tiny-data/nq-train-??.jsonl.gz
--output_tfrecord=tiny_output_dir/nq-train.tfrecords-00000-of-00001
--max_seq_length=512
--include_unknowns=0.02
--vocab_file=uncased_L-12_H-768_A-12/vocab.txt

python -m language.question_answering.bert_joint.run_nq
--logtostderr
--bert_config_file=uncased_L-12_H-768_A-12/bert_config.json
--vocab_file=uncased_L-12_H-768_A-12/vocab.txt
--train_precomputed_file=tiny_output_dir/nq-train.tfrecords-00000-of-00001
--train_num_precomputed=494670
--learning_rate=3e-5
--num_train_epochs=1
--max_seq_length=512
--save_checkpoints_steps=5000
--init_checkpoint=uncased_L-12_H-768_A-12/bert_model.ckpt
--do_train
--output_dir=bert_model_output

[yanchlu]

Oh, I forget to clean the "bert_model_output" file.