Some background knowledge:
BERT can generate 2 kind of output: 1. two dimensional sequence_output(word embeddings) . 2. one dimensional pooled output(cls or sentence embedding) I recommend to use sequence_output of BERT(Many paper show that this is better than cls embedding).

Python 3.6
Required Library:
PyTorch
transformers: pip install transformers
tensorflow2.0

I recommend to use Bert_classification_tf2.py or ipynb

I set max sequence length to 512. You can change it in code.

For tensorflow 2.0: Bert_classification_tf2.py or ipynb
I implemented 3 models for fine-tuning BERT:

1. create_model_sequence_output(trainable=True): create a fine-tune model that uses two dimensional sequence_output of BERT(word embeddings)
2. create_model_cls_output(trainable=True): create a fine-tune model that uses one dimensional pooled output(cls or sentence embedding)
3. create_model_2(trainable=True): as same as create_model_cls_output(trainable=True)
   if trainable=True: --> the weights of pre-trained BERT will also be adjusted during fine-tuning.
   if trainable=False: --> the weights of pre-trained BERT will not be adjusted during fine-tuning.

For pytorch: Bert_classification_pytorch.py or ipynb
I only implemented 1 model for fine-tuning BERT: create a fine-tune model that uses one dimensional pooled output(cls or sentence embedding)

I put the example codes inside each python file. I used part of data in IMDB datatset as an example.

Reference:
Chris McCormick and Nick Ryan, "BERT Fine-Tuning Tutorial with PyTorch", https://mccormickml.com/2019/07/22/BERT-fine-tuning/