bert_sa (bert sentiment analysis tensorflow serving with RESTful API)

based on bert including training, online predicting and serving with REST

Fine tune a sentiment analysis model based on BERT

1. Add a SAProcessor and include it within main function in run_classifier.py
2. Prepare train, dev and test files; adapat _create_examples method in SAProcessor based on your own datasets (pandas may not be required)
3. Specify BERT_BASE_DIR, SA_DIR and output_dir in run_sa.sh and run

Test

1. For file based test, change output_predict_file in run_classifier.py, specify TRAINED_CLASSIFIER and output_dir path, run predict_sa.sh
2. For online prediction, refer to run_classifier_predict_online (modified based on bert_language_understanding)

Export your model

Refer to sa_predict_saved_model.py

KIND NOTICE: some graph definition and input placeholder is imported from run_classifier_predict_online.py

Serve the model with TensorFlow Serving

1. See TensorFlow Serving for details about installing docker and pulling a serving image
2. Running a serving image

docker run -p 8501:8501 --name 'bert_sa_serving' --mount type=bind,source=/data/notebooks/xff/bert/output/sa_output/saved_model,target=/models/bert_sa -e MODEL_NAME=bert_sa -t tensorflow/serving:latest-devel-gpu &

docker exec -it bert_sa_serving bash

tensorflow_model_server --port=8500 --rest_api_port=8501 \
  --model_name=bert_sa --model_base_path=/models/bert_sa

3. Sample request

line=u'建立了完善的质量体系并持续有效运行'
# preprocess is defined in run_classifier_predict_online.py
dict_data = preprocess(line)
resp = requests.post('http://172.17.0.1:8501/v1/models/bert_sa:predict', json=dict_data)
print(resp.json())

Results look like this: {'outputs': {'label_predict': 1, 'possibility': [0.00738544, 0.992615]}}