MNIST DVC

Project Goals

Classify digits from images using the MNIST dataset

Introduction

This repository uses Data Version Control (DVC) to create a machine learning pipeline and track experiments for the Kaggle competition digit-recognizer. We will use a modified version of the Team Data Science Process as our Data Science Life cycle template. This repository template is based on the cookiecutter data science project template.

In order to start, clone this repository and install DataVersionControl. Follow the instructions below to proceed through the data science life cycle using DVC to manage parameters, scripts, artifacts, and metrics.

1. Domain understanding/problem definition

The first step any data science life cycle is to define the question and to understand the problem domain and prior knowledge. Given a well-formulated question, the team can specify the goal of the machine learning application (e.g., regression, classification, clustering, outlier detection) and how it will be measured, and which data sources will be needed. The scope of the project, key personnel, key milestones, and general project architecture/overview is specified in the Project Charter and iterated throughout the life of the project. A list of data sources which are available or need to be collected is specified in the table of data sources. Finally, the existing data is summarized in a data dictionary that describes the features, number of elements, non-null data, data type (e.g., nominal, ordinal, continuous), data range, as well as a table with key descriptive summary statistics.

Deliverables Step 1:

1. Project charter
2. Table of data sources
3. Data dictionary
4. Summary table of raw dataset

Downloading the dataset

The script download.py will download the dataset from Kaggle. The key artifacts of this DVC stage are the raw training and testing datasets.

In your terminal, use the command-line interface to build the first stage of the pipeline.

dvc run -n download_data -p competition \
    -d src/data/download.py \
    -o data/raw/train.csv \
    -o data/raw/test.csv \
    --desc "Download data from Kaggle"\
    python3 src/data/download.py -tr train.csv -te test.csv -o "./data/raw"

dvc run -n prepare_images \
    -d src/data/prepare_img.py \
    -o data/processed/train/ \
    -o data/processed/test/ \
    --desc "Create images from numpy array"\
    python3 src/data/prepare_img.py -tr train.csv -te test.csv -o "./data/processed/"

2. Data acquisition and understanding

The second step involves acquiring and exploring the data to determine the quality of the data and prepare the data for machine learning models. This step involves exploring and cleaning the data to account for missing data and noise as well as validating that data meet specified validation rules to ensure there were no errors in data collection or data entry (e.g., age and fare cannot be negative). Once the data is cleaned, it is processed to encode categorical string variables as integer classes, continuous features are discretized (optional), and features are normalized (optional). Later stages may iteratively add or create new features from new data or existing features using feature engineering.

Deliverables Step 2:

1. Data quality report
2. Proposed data pipeline/architecture
3. Checkpoint decision

Encoding categorical labels as integer classes

The script encode_labels.py is an intermediate data processing script that accepts the raw training data, and the "dtypes" parameter from the params.yaml file. It encodes the columns with categorical variables as integer values for machine processing and saves the updated dataset and encoding scheme. Importantly, the training and testing data is processed at the same time to ensure the identical label encoding. Key artifacts from this stage include the interim categorized datasets and the label encoding scheme.

dvc run -n encode_labels -p dtypes \
-d src/data/encode_labels.py \
-d data/raw/train.csv \
-d data/raw/test.csv \
-o data/interim/train_categorized.csv \
-o data/interim/test_categorized.csv \
-o data/interim/label_encoding.yaml \
--desc "Convert categorical labels to integer values and save mapping" \
python3 src/data/encode_labels.py -tr data/raw/train.csv -te data/raw/test.csv -o data/interim

Cleaning data

This section involves preparing the data for machine learning. First, missing values are imputed from the training data in replace_nan.py. Next, feature engineering is used to create additional informative representations. Once initial feature engineering is complete, the feature set is explored to identify correlated features and optionally the feature set is reduced using dimensionality reduction techniques. Once a set of feature is identified, the data is optionally normalize_data.py. Key artifacts from this stage include the interim nan-imputed datasets, a Jupyter notebook exploring the dataset and features, interim feature engineering datasets, and the final processed dataset after feature normalization.

Replace missing age values using imputation

dvc run -n impute_nan -p imputation
    -d src/data/replace_nan.py
    -d data/interim/train_categorized.csv
    -d data/interim/test_categorized.csv
    -o data/interim/test_nan_imputed.csv
    -o data/interim/train_nan_imputed.csv
    --desc "Replace missing values for age with imputed values from training dataset."
    python3 src/data/replace_nan.py -tr data/interim/train_categorized.csv -te data/interim/test_categorized.csv -o data/interim

Feature engineering

1. Engineer new features
2. Show feature correlation
3. Identify importance

dvc run -n build_features -p random_seed,feature_eng \
    -d src/features/build_features.py
    -d data/interim/train_nan_imputed.csv
    -d data/interim/test_nan_imputed.csv
    -o data/interim/train_featurized.csv
    -o data/interim/test_featurized.csv
    --desc "Optional feature engineering and dimensionality reduction"
    python3 src/features/build_features.py -tr data/interim/train_nan_imputed.csv -te data/interim/test_nan_imputed.csv -o data/interim/  

Normalize features

dvc run -n normalize_data -p normalize \
    -d src/features/normalize.py \
    -d data/interim/train_featurized.csv \
    -d data/interim/test_featurized.csv \
    -o data/processed/train_processed.csv \
    -o data/processed/test_processed.csv \
    --desc "Optionally normalize features by fitting transforms on the training dataset." \
    python3 src/features/normalize.py -tr data/interim/train_featurized.csv -te data/interim/test_featurized.csv -o data/processed/

3. Modeling

Split data into the train, dev, and test sets

• Data split report

dvc run -n split_train_dev -p random_seed,train_test_split \
    -d src/data/split_train_dev.py \
    -d data/processed/train_processed.csv \
    -o data/processed/split_train_dev.csv \
    --desc "Split training data into the train and dev sets using stratified K-fold cross validation." \
    python3 src/data/split_train_dev.py -tr data/processed/train_processed.csv  -o data/processed/

Model training

dvc run -n train_model -p classifier,model_params,random_seed,train_test_split.target_class \
    -d src/models/train_model.py \
    -d data/processed/train_processed.csv \
    -d data/processed/split_train_dev.csv \
    -o models/estimator.pkl \
    -m results/metrics.json \
    --desc "Train the specified classifier using the pre-allocated stratified K-fold cross validation splits and the current params.yaml settings." \
    python3 src/models/train_model.py -tr data/processed/train_processed.csv -cv data/processed/split_train_dev.csv

Predict output

dvc run -n predict_output -p predict,train_test_split.target_class \
    -d src/models/predict.py \
    -d src/models/metrics.py \
    -d models/estimator.pkl \
    -d data/processed/test_processed.csv \
    -o results/test_predict_proba.csv \
    -o results/test_predict_binary.csv \
    --desc "Predict output on held-out test set for submission to Kaggle." \
    python3 src/models/predict.py -te data/processed/test_processed.csv -rd results/ -md models/

4. Deployment

Status dashboard

• Display system health
• Final modeling report
• Final solution architecture

5. Project conclusion

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience