This repo contains the master solution for Assignment #2 in the Summer 2021 edition of MGMT 590 (Production Scale Data Products) at Purdue University. Specifically, this repo includes code and config for running a REST API for question answering using pre-trained models from Hugging Face Transformers.
The REST API depends on the following:
The API has been tested with Python version 3.6.9 and TensorFlow 2.5.0. See requirements.txt and our Dockerfile for more information about supported versions of these dependencies.
The Dockerized version of the API can be built as run as follows:
$ docker build -t mgmt590 .
$ docker run -it -p 8080:8080 mgmt590
Once the API is running, you can do a health check via cURL as follows:
$ curl http://<host>:8080
<p>The question answering API is healthy!</p>
This route allows a user to obtain a list of the models currently loaded into the server and available for inference.
Method and path:
GET /models
Expected Response Format:
[
{
"name": "distilled-bert",
"tokenizer": "distilbert-base-uncased-distilled-squad",
"model": "distilbert-base-uncased-distilled-squad"
},
{
"name": "deepset-roberta",
"tokenizer": "deepset/roberta-base-squad2",
"model": "deepset/roberta-base-squad2"
}
]
This route allows a user to add a new model into the server and make it available for inference.
Method and path:
PUT /models
Expected Request Body Format:
{
"name": "bert-tiny",
"tokenizer": "mrm8488/bert-tiny-5-finetuned-squadv2",
"model": "mrm8488/bert-tiny-5-finetuned-squadv2"
}
Expected Response Format (updated list of available models):
[
{
"name": "distilled-bert",
"tokenizer": "distilbert-base-uncased-distilled-squad",
"model": "distilbert-base-uncased-distilled-squad"
},
{
"name": "deepset-roberta",
"tokenizer": "deepset/roberta-base-squad2",
"model": "deepset/roberta-base-squad2"
},
{
"name": "bert-tiny",
"tokenizer": "mrm8488/bert-tiny-5-finetuned-squadv2",
"model": "mrm8488/bert-tiny-5-finetuned-squadv2"
}
]
This route allows a user to delete an existing model on the server such that it is no longer available for inference.
Method and path:
DELETE /models?model=<model name>
Query Parameters:
model(required) - The name of the model to be deleted
Expected Response Format (updated list of available models):
[
{
"name": "distilled-bert",
"tokenizer": "distilbert-base-uncased-distilled-squad",
"model": "distilbert-base-uncased-distilled-squad"
},
{
"name": "deepset-roberta",
"tokenizer": "deepset/roberta-base-squad2",
"model": "deepset/roberta-base-squad2"
}
]
This route uses one of the available models to answer a question, given the context provided in the JSON payload.
Method and path:
POST /answer?model=<model name>
Query Parameters:
model(optional) - The name of the model to be used in answering the question. If no model name is provided use a default model.
Expected Request Body Format:
{
"question": "who did holly matthews play in waterloo rd?",
"context": "She attended the British drama school East 15 in 2005, and left after winning a high-profile role in the BBC drama Waterloo Road, playing the bully Leigh-Ann Galloway.[6] Since that role, Matthews has continued to act in BBC's Doctors, playing Connie Whitfield; in ITV's The Bill playing drug addict Josie Clarke; and she was back in the BBC soap Doctors in 2009, playing Tansy Flack."
}
Expected Response Format:
{
"timestamp": 1621602784,
"model": "deepset-roberta",
"answer": "Leigh-Ann Galloway",
"question": "who did holly matthews play in waterloo rd?",
"context": "She attended the British drama school East 15 in 2005, and left after winning a high-profile role in the BBC drama Waterloo Road, playing the bully Leigh-Ann Galloway.[6] Since that role, Matthews has continued to act in BBC's Doctors, playing Connie Whitfield; in ITV's The Bill playing drug addict Josie Clarke; and she was back in the BBC soap Doctors in 2009, playing Tansy Flack."
}
This route returns recently answered questions.
Method and path:
GET /answer?model=<model name>&start=<start timestamp>&end=<end timestamp>
Query Parameters:
model(optional) - Filter the results by providing a certain model name, such that the results only include answered questions that were answered using the provided model.start(required) - The starting timestamp, such that answers to questions prior to this timestamp won't be returned. This should be a Unix timestamp.end(required) - The ending timestamp, such that answers to questions after this timestamp won't be returned. This should be a Unix timestamp.
Expected Response Format (updated list of available models):
[
{
"timestamp": 1621602784,
"model": "deepset-roberta",
"answer": "Leigh-Ann Galloway",
"question": "who did holly matthews play in waterloo rd?",
"context": "She attended the British drama school East 15 in 2005, and left after winning a high-profile role in the BBC drama Waterloo Road, playing the bully Leigh-Ann Galloway.[6] Since that role, Matthews has continued to act in BBC's Doctors, playing Connie Whitfield; in ITV's The Bill playing drug addict Josie Clarke; and she was back in the BBC soap Doctors in 2009, playing Tansy Flack."
},
{
"timestamp": 1621602930,
"model": "distilled-bert",
"answer": "Travis Pastrana",
"question": "who did the first double backflip on a dirt bike?",
"context": "2006 brought footage of Travis Pastrana completing a double backflip on an uphill/sand setup on his popular /"Nitro Circus/" Freestyle Motocross movies. On August 4, 2006, at X Games 12 in Los Angeles, he became the first rider to land a double backflip in competition. Having landed another trick that many had considered impossible, he vowed never to do it again."
}
]