Hi,
I would like to load the parameters I trained from a fine-tuned Sentence Pair Classification with BERT.
I followed the scripts from https://gluon-nlp.mxnet.io/examples/sentence_embedding/bert.html, and save the model by
net =bert_classifier
net.save_parameters(output_dir +‘model_epoch_’+str(epoch_id)+’.params’)
net.export(output_dir+‘epoch_’+str(epoch_id))
However, when i tried to load it back to the system by
base_model, voca = nlp.model.get_model(‘bert_12_768_12’,
dataset_name=‘book_corpus_wiki_en_uncased’,
pretrained=output_dir+‘model_epoch_’+str(epoch_id)+’.params’,
ctx=ctx, use_pooler=False,
use_decoder=False, use_classifier=False)
base_model.load_parameters(output_dir +‘model_epoch_’+str(epoch_id)+’.params’, ctx=ctx, ignore_extra=True)
I got the error
AssertionError: Parameter ‘encoder.position_weight’ is missing in file ‘/home/ec2-user/SageMaker/wh/scripts/finetune_Bert/finetune_model/model_epoch_2.params’, which contains parameters: ‘bert.encoder.position_weight’, ‘bert.encoder.layer_norm.gamma’, ‘bert.encoder.layer_norm.beta’, …, ‘bert.pooler.weight’, ‘bert.pooler.bias’, ‘classifier.1.weight’, ‘classifier.1.bias’. Set allow_missing=True to ignore missing parameters.
I wonder if you guys have similar situations? thanks!