github.com/jackyzha0/reflect-nlp

reflect-nlp

The backend of reflect which determines intent validity and does stats collection. the main repo.

Anything related to the ingress controller can be found in /ingress. All the NLP stuff can be found in /nlp.

Local Docker build instructions

Note: There is no need to do this when deploying as the Docker images will be rebuilt by CircleCi.

1. Build ingress proxy image: docker build -t jzhao2k19/reflect-nlp-ingress:latest ingress
2. Build NLP model image: docker build -t jzhao2k19/reflect-nlp:latest nlp
3. Push both Docker Images to Docker Hub

Running the K8s deployment locally

First, ensure you have Docker Desktop installed. There are also a few other requirements that you should install as well.

1. VirtualBox You can get this through homebrew by doing brew install virtualbox. VirtualBox allows us to run the VMs.
2. kubectl Pronounced cube-control, kubectl is the command line interface for talking to K8s. Install it by doing brew install kubectl.
3. minikube minikube allows you to run a K8s cluster right on your laptop! Install it by going brew install minikube.
   

Finally, enable some addons for minikube which allow us to configure the horizontal pod autoscalers (HPAs). minikube addons enable heapster minikube addons enable metrics-servce

To spin up the K8s cluster, start minikube then use kubectl to apply our config. minikube start kubectl apply -f k8s_local.yml To see if this is done successfully, run kubectl get pods. It should give you something that looks like the following.

➜ kubectl get pods
NAME                           READY   STATUS    RESTARTS   AGE
ingress-54f9b89fc4-nxl47       2/2     Running   0          5m
nlp-5f77b4946-vs4l2            1/1     Running   0          5m

Note that ingress has 2/2 pods as there is a CloudSQL sidecar. Wait a few minutes for the LoadBalancer to be assigned an external IP address, then run kubectl get svc to list running services.

➜ kubectl get svc
NAME              TYPE           CLUSTER-IP       EXTERNAL-IP   PORT(S)           AGE
ingress-service   LoadBalancer   10.106.149.250   <pending>     80:31274/TCP      7m
kubernetes        ClusterIP      10.96.0.1        <none>        443/TCP           7m
nlp-service       NodePort       10.103.174.142   <none>        30000:32610/TCP   7m

Note that on GKE, the EXTERNAL-IP of ingress-service would be configured for you. However, if you're developing locally on minikube, you can just access the service by doing minikube service ingress-service.

You can check to which HPAs are running by doing kubectl get hpa.

➜ kubectl get hpa
NAME      REFERENCE            TARGETS   MINPODS   MAXPODS   REPLICAS   AGE
ingress   Deployment/ingress   0%/80%    1         3         1          9m
nlp       Deployment/nlp       32%/50%   1         10        1          9m

To test if everything is healthy, run minikube service ingress-service which will open a new browser window. This page should give you a 404 as nothing is listening on /. However, if you visit /healthcheck, you should see a nice JSON like so:

{
	"proxyAlive": true,
	"modelAlive": true
}

Exporting intents

Local logged intents can be exported as a CSV by hitting /export. This endpoint is rate-limited.

Running the NLP Model

This project depends on a bunch of Python libraries. Install them by doing pip install sklearn keras pandas numpy matplotlib

For local development, you can run the server by doing python server.py, which will start a local server on port 5000.

Training the NLP Model

You can train a new version of the neural network on the data/survey.csv data by doing python train.py. This will begin training of a basic 64 cell LSTM model (which is defined in net.py). You can configure the training parameters which are constants at the top of train.py.

TOKENIZER_VOCAB_SIZE = 500 # Vocabulary size of the tokenizer
SEQUENCE_MAX_LENGTH = 75 # Maximum sequence length, all seqs are padded to this
BATCH_SIZE = 128 # number of examples per batch
NUM_EPOCHS = 10 # number of epochs to train for (an epoch is one iteration of the entire dataset)
TRAIN_TEST_SPLIT = 0.1 # percentage of data to use for testing
VALIDATION_SPLIT = 0.1 # percentage of training data to use for validation

Trained models are stored in the models folder. Each model is under its own folder whose folder structure looks as follows:

models
 | - acc%%.%% # where %%.%% represents accuracy on the test set
 |   | - details.yml # stores training details
 |   | - model.json # stores model architectures
 |   | - tokenizer.json # stores tokenizer embeddings
 |   | - weights.h5 # stores weights for neural conenctions
 | ...

Converting models for use in tensorflow.js

Tensorflow.js requires a different format for saved models. We can convert these using the tensorflowjs_converter tool. Run ./convert_to_js.sh <name_of_model> to convert said model into a tensorflow.js usable format. You can find the output in nlp/converted_models.

NLP Model CLI

You can also run the NLP model through the command line (given the model exists) by just providing arguments to serve_model.py. Example usage is as follows,

# e.g.
# serve_model.py -m <nameofmodel> -t <threshold> -i <intent>

python serve_model.py -i "I need to make a marketing post"
# Predicting using model acc81.08 with threshold 0.50 on intent "I need to make a marketing post"
# Output -> True

python serve_model.py -i "I want to browse memes"
# Predicting using model acc81.08 with threshold 0.50 on intent "I want to browse memes"
# Output -> False

Using different NLP models on the server.

Currently, the server is running a default model of the acc85.95 model. This is defined in server.py as follows,

if __name__ == '__main__':
    logging.info("Starting server...")
    m = Model("acc85.95", threshold=0.5)
    app.run()

You may change the model name and threshold however you may see fit.

Data Usage

All data found in data/survey.csv collected from this survey that our team sent out in January of 2020. You may use this data to train your own models.