Simple code and model for making binary predictions on the spam/ham classification task.
Before running code both the *rnn.ht.zip and *tokenizer.zip files need to be unzipped (into same location as respective zip files)
Build docker image by docker-compose in ./docker_clf_project dir: docker-compose build
Run the container using: docker run -it -p 50051:50051 <docker image id>
By default, gRPC connection is via port 50051
Test the gRPC server. In a separate command window run the script ./docker_clf_project/spam_clf_endpoint/spam_clf_client.py
This sends two separate messages to the server. The first should be classified as HAM (score 0.0), and the 2nd as SPAM (score 1.0).
gRPC protobuf config is in the grapevine.proto file. If this configuration is changed, follow instructions in ./docker_clf_project/spam_clf_endpoint/readme.md to re-generate the necessary python protobuf files.
You can use the inference.py script as a command-line tool to save out predictions for your files. By default, a tab-separated file will be written to disk (named preds.csv) where each row is indexed by the filename that predictions were generated for. The script also generates class predictions as well as probability or "confidence value" for predictions by default for each file.
To run the script, clone the repo and cd to keras_spam_predictor directory. There are several command-line arguments that can be passed to the script, but the only required one is --data_dir which points to a directory of plain-text files. The script assumes that each plain-text file within the data_dir directory corresponds to one document that should be classified by the model.
To run a demo using 3 dummy files under the supplied files directory, simple run the following command at the command-line:
python inference.py --data_dir=files
This should produce a CSV preds.csv with 3 rows and corresponding predictions.