Test QAPipeline with GPU

@fmikaelian I have implemented the changes as discussed. I tested the fit() method for the retriever and the predict() method in a notebook included in examples. Everything is working fine.

Could you please test the fit() method for the reader in GPU to check if everything is ok?

Please note that the current implementation of QAPipeline still uses the TfidfRetriever as you have implemented (passing the dataframe column as input to TfidfRetriever.fit(df['content'])). It should be changed once you will have implemented the improvement I proposed in #95.

_Originally posted by @andrelmfarias in https://github.com/fmikaelian/cdQA/pull/101#issuecomment-488628180_

[fmikaelian]

Using a pre-trained reader model on GPU:

qa_pipe = QAPipeline(metadata=df, model='../models/bert_qa_squad_v1.1_sklearn/bert_qa_squad_v1.1_sklearn.joblib')
qa_pipe.fit()
qa_pipe.model.output_dir = '../logs/bert_qa_squad_v1.1_sklearn'

X = 'Since when does the Excellence Program of BNP Paribas exist?'
prediction = qa_pipe.predict(X)

print('question: {}'.format(X))
print('answer: {}'.format(prediction))

question: Since when does the Excellence Program of BNP Paribas exist?
answer: January 2016

[andrelmfarias]

I tested BertQA.fit() with the following code on Colab:

processor = BertProcessor(bert_model='bert-base-uncased', do_lower_case=True, is_training=True)
train_examples, train_features = processor.fit_transform('./data/dev-v1.1.json')

reader = BertQA(train_batch_size=8, num_train_epochs=1, output_dir='test', fp16=True)
reader.fit(X=(train_examples, train_features))

And got the following error:

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-23-ab02e6a6f1de> in <module>()
----> 1 reader.fit(X=(train_examples, train_features))

<ipython-input-21-c31ed17ed5ea> in fit(self, X, y)
   1144         torch.save(model_to_save.state_dict(), output_model_file)
   1145         model_to_save.config.to_json_file(output_config_file)
-> 1146         tokenizer.save_vocabulary(self.output_dir)
   1147 
   1148         self.model.to(self.device)

NameError: name 'tokenizer' is not defined

The problem occurs in https://github.com/fmikaelian/cdQA/blob/0dce89f48ab53a69e8fdb8b76f39029f465f5bbc/cdqa/reader/bertqa_sklearn.py#L1165

It probably comes from the direct adaptation of run_squad.py from Hugging Face to our BertQA class:

https://github.com/huggingface/pytorch-pretrained-BERT/blob/3fc63f126ddf883ba9659f13ec046c3639db7b7e/examples/run_squad.py#L1036

For now, I will delete the line in the fit() method of BertQA. If needed, we might include this saving in the BertProcessor class (when the attribute is_training is True), where we process the text and create a vocabulary.