TWEET to EMOJI.

This is my take-home test submission for the Data Scientist position at Glia.

The test consisted of creating a multi-class classification NLP model using a publicly avaiable dataset. Then, make it available through a REST API.

I'm using the Kaggle dataset Emotions dataset for NLP which consists of 20k Tweets to 6 different emojis.

Content

1. Dataset exploration.
2. Model training and evaluation.
3. REST API
4. Docker

I. Data Exploration

This section is straight forward. First, make sure to have a python virtual environment and install the requirements.file

From the new environment:

pip install -r requirements.txt

Then, using Jupyter-lab (or notebook) load the Playground.ipynb file and run the cells. This notebook contain the approach taken to explore the dataset and clean it.

II. Model training an evaluation

The proposed model takes a text (tweet) and returns the most fitting emoji. I used pre-trained BERT models to experiment with fine-tuning and transfer learning.

The best model (Fine-tuned TFDistilBertForSequenceClassification) achieved a balanced accuracy of 93% on the validation set:

And good F1-score across all classes on the test set.

Similar to the previos section, I used two notebooks:

• Fine_tune_Bert_GCLB
• Transfer_learning_BERT_GCLB

Both were run on Google Colab using the T4 GPU. If you want to run them locally, make sure you have installed the requirements.txt and use Jupyter-lab (or notebook).

The results for the transfer learning model can be found on the "/screens" directory of on the Transfer_learning_BERT_GCLB.

III. REST API

The resulting model is exposed through a Flask REST API. A GET request is sent to "tweets" end-point with the following format:

GET REQUEST

{
    "text": "Today is a good day!
}

RESPONSE

{
    "confidence": "0.9963295",
    "emoji": "😂"
}

If want to run the API locally you have to first train the model (run the Fine_tune_Bert_GCLB notebook) and save both the tokenizer and model in the "/emoji_api/models/tf_bert_sequence_tok" and "/emoji_api/models/tf_bert_sequence" correspondingly.

MODELS ARE NOT INCLUDED IN THE REPO SINCE THEY ARE BIGGGER THEN 100MB. IS EASIER TO DOWNLOAD AND RUN THE PROVIDED DOCKER CONTAINER WITH THE FINE-TUNED MODELS

Then, install the requirements on "/emoji_api/requirements.txt" (which includes Flask) and then run Flask from within "/emoji_api":

flask run

Once Flask is running, you can send your request to:

http:localhost:<YOUR_PORT>/tweets

Also, you can look at the description of the API thanks to OpenAPI & swagger UI:

http:localhost:<YOUR_PORT>/swagger-ui

III. Docker

Finally the API is containerized on docker container. The image can be downloaded from Docker Hub:

https://hub.docker.com/r/geramantecon/emojiapi_takehome

THIS IS THE BEST WAY TO TRY THE API AS IT CONTAINS THE TRAINED MODEL!

Then run it:

docker run -dp 5005:5000 emojiapi_takehome

If this is not possible, we can build it from within the "emoji_app" (Dockerfile is there) directory using:

docker build -t <YOU_IMAGE_NAME> .

Then run:

docker run -dp 5005:5000 <YOU_IMAGE_NAME>