ML-ZOOMCAMP MIDTERM PROJECT
This is a Midterm project for ML Zoomcamp 2022
In this repo we will be working with the **Wine quality dataset ** dataset from UCI center for Machine Learning.
The goal of this project is to classify the wine quality based on some of their physical and chemical properties.
Data Description
The following table provides an overview of the dataset. The available fields and what they were used for, as well as the type of data.
| Field | Use | Dtype |
|---|---|---|
| type | Feature | str |
| fixed_acidity | Feature | float |
| volatile_acidity | Feature | float |
| citric_acid | Feature | float |
| residual_sugar | Feature | float |
| chlorides | Feature | float |
| free_sulfur_dioxide | Feature | float |
| total_sulfur_dioxide | Feature | float |
| density | Feature | float |
| ph | Feature | float |
| sulphates | Feature | float |
| alcohol | Feature | float |
| Quality | Target | int |
Data Preparation
-
Missing data: The missing values were filled using the median values of each wine type (red or white).
Type fixed_acidity volatile_acidity citric_acid residual_sugar chlorides ph sulphates white 6.856 0.278 0.334 6.393 0.046 3.188 0.490 red 8.322 0.528 0.271 2.539 0.087 3.311 0.658 -
Quality_encoding: Since the objective is classify the quality of the wine, the scores from 0-10 were binned into bad,moderate and good quality wine. Given that good > moderate > bad, we can use ordinal encoding.
Quality Score range code Bad < 5 0 Moderate 5-6 1 Good > 7 2 -
Type Encoding: The type of wine was enconded as red → 0 and white → 1
ML modeling
The models tested were: LogisticRegression, DecisionTreeClassifier, RandomForestClassifier and XGBClassifier
Validation framework
The dataset was splitted as train/test with prpportions of 0.8 and 0.2 respectively . Given that the dataset is imbalanced, the dataset split was done in an stratified fashion using the quality labels.
The train dataset was split in 2 using a 0.25 fraction of it as validation (0.2 fraction of the full dataset)
Pipeline
Before using Pipeline,the datasets are first converted to vectors using DictVectorizer.
The Pipeline object contains to steps:
- Preprocessing: Given the long-tailed distributions of some of the features
RobustScalerwas applied to the feature vectors. - Classifier: Model that is used and evaluated
Evaluation
The model evaluation was performed in using different metrics.
- Area under the RO curve.
- F1 multiclass score:
- F1 macro
- F1 weighted
Additionally, the inspection of the confussion matrix and recall to understand the weaknesses and virtues of the model.
Select Best Model
Recall for the best models are as follows:
| Model | Bad | Moderate | Good |
|---|---|---|---|
| RT_best | 16 | 91 | 63 |
| xgb_best | 16 | 94 | 56 |
| RT_base | 14 | 97 | 50 |
| xgb_base | 14 | 93 | 55 |
We can see that while the metrics are quite similar, some models perform better or worst in some of the categories. In general, we get good predictions for moderate-quality wines. However, Some models are better at predicting either the bad- or good-quality wines.
If we compare the best RandomForest (RT_best) with the best XGB (xgb_best), we can observe that both models have the same low recall for bad-quality wines. RT_best perform better at predicting good-quality while xgb_best is better at predicting moderate quality wines.
If we assume that we care more about prediciting correctly the good quality wine, then we select as the best model RT_best and its metrics are as follows:
- Weighted ACU 0.857
- Macro avg F1-score 0.58
- Weighted avg F1-score 0.82
How to use this project
First things first. Clone this repository and create a virtual environment.
git clone https://github.com/drcoronel/mlzoomcamp_midterm2020.git && cd mlzoomcap_midterm2020
pip install pipenv
pipenv shell Now, you'll have all the dependencies installed in a virtual environment for this project.
There are a few thing that you can try out:
- You can deploy the service using the
bentomlmodel stored inmodelfolder. - You can train a new model and save it using
bentomlandtrain.py
Using the existing model
bentoml models import model/wine_quality_randomforest-ahhbkxtaq26rig2k.bentomodel
bentoml serve service.py:svcThe existing model
name: t file, requirements.txt or pwine_quality_randomforest
version: ahhbkxtaq26rig2k
module: bentoml.sklearn
labels: {}
options: {}
metadata:
auc: 0.857
f1_macro: 0.58
f1_weighted: 0.82
context:
framework_name: sklearn
framework_versions:
scikit-learn: 1.1.3
bentoml_version: 1.0.10
python_version: 3.8.10
signatures:
predict:
batchable: true
batch_dim:
- 0
- 0
predict_proba:
batchable: true
batch_dim:
- 0
- 0
api_version: v1
creation_time: '2022-11-09T23:26:59.052366+00:00' Make your own model
You can change the model or the hyperparameters used during training using train.py.
python train.pyAfter running, in the terminal you'll see the new models tag. Just go and replace it in service.py :p
Deployment using BentoML
bentoml build 
Now, we can build a docker image using the following snippet
bentoml containerize wine_quality_classifier:mcye7jtbds6opr5fAnd we are ready to deploy our server :)
docker run -it --rm -p 3000:3000 wine_quality_classifier:mcye7jtbds6opr5f serve --productionNow that the server is deployed, you can go to http://0.0.0.0:3000 and use the API :)
Try it:
{"type": "white",
"fixed_acidity": 8.8,
"volatile_acidity": 0.23,
"citric_acid": 0.35,
"residual_sugar": 10.7,
"chlorides": 0.04,
"free_sulfur_dioxide": 26.0,
"total_sulfur_dioxide": 183.0,
"density": 0.9984,
"ph": 2.93,
"sulphates": 0.49,
"alcohol": 9.1}
{"type": "red",
"fixed_acidity": 6.9,
"volatile_acidity": 0.52,
"citric_acid": 0.25,
"residual_sugar": 2.6,
"chlorides": 0.081,
"free_sulfur_dioxide": 10.0,
"total_sulfur_dioxide": 37.0,
"density": 0.99685,
"ph": 3.46,
"sulphates": 0.5,
"alcohol": 11.0}You can also use curl to send a request to the server:
curl -X 'POST' \
'http://localhost:3000/classify' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{"type": "white",
"fixed_acidity": 8.8,
"volatile_acidity": 0.23,
"citric_acid": 0.35,
"residual_sugar": 10.7,
"chlorides": 0.04,
"free_sulfur_dioxide": 26.0,
"total_sulfur_dioxide": 183.0,
"density": 0.9984,
"ph": 2.93,
"sulphates": 0.49,
"alcohol": 9.1} 'The response for the first example looks like this:
{
"Probabilities": {
"bad": 0.09,
"moderate": 0.77,
"good": 0.14
},
"Quality": "moderate"
}Cloud Deployment
The model was deployed using AWS Elastic Container Services
First, make sure that aws-cli is installed.
curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
unzip awscliv2.zip
sudo ./aws/install
ln -s /usr/bin/aws aws- We need to create a Elastic Container Regestry and push the docker image to it.
aws ecr get-login-password --region {region} | docker login --username AWS --password-stdin {registry_url}
docker tag {docker_image_tag} {registry_url}/wine_quality_classifier:latest
docker push {registry_url}/wine_quality_classifier:latest-
We create a cluster in EC service. In this case I used an AWS Fargate Linux cluster.
-
We create a task and assigned the URI from the docker image that we pushed to the registry.
-
We run the task
-
Test
Now the service is deployed in the cloud and the API is exposed
WineCloud.mp4
Additonally, we could also send a request using curl
WineCurl.mp4
Reference
P. Cortez, A. Cerdeira, F. Almeida, T. Matos and J. Reis. Modeling wine preferences by data mining from physicochemical properties. In Decision Support Systems, Elsevier, 47(4):547-553, 2009.