This is a Deep Learning project for car evaluation using a custom Linear Model implemented with PyTorch. The goal of the project is to classify car evaluations into one of four categories: "unacceptable," "acceptable," "good," and "very good."
Car Evaluation Data Set consists of 7 attributes. Buying, maint, doors, persons, lug_boot, safery are 6 features of cars, and class is the target variable.
The goal is to predict car acceptability (class) of a car using these 6 features.
The table below shows the detailed definition of each attribute.
| Attribute name | Definition |
|---|---|
| buying | Buying price of the car (v-high, high, med, low) |
| maint | Price of the maintenance of car (v-high, high, med, low) |
| doors | Number of doors (2, 3, 4, 5-more) |
| persons | Capacity in terms of persons to carry (2, 4, more) |
| lug_boot | The size of luggage boot (small, med, big) |
| safety | Estimated safety of the car (low, med, high) |
| class | Car acceptability (unacc: unacceptible, acc: acceptible, good: good, v-good: very good) |
Before running the project, make sure you have the following prerequisites:
- Python (>= 3.6)
- PyTorch (>= 1.7.1)
- NumPy
- pandas
- scikit-learn
- matplotlib
- tqdm
- imbalanced-learn
To install the required dependencies, run the following command:
pip install -r requirements.txt
The project consists of several components:
- EDA.ipynb - jupyter notebook with Exploratory Data Analysis
- data_loader.py: contains the custom Dataset class for loading the car evaluation dataset.
- model.py: defines the custom Linear Model.
- engine.py: implements the training and evaluation functions.
- utils.py: contains utility functions for logging, checkpoint saving.
- early_stopping: defines class for early stopping.
- run.py: the main script to run the training and evaluation process.
- made_pred.py - script for making prediction with user's data.
To train the model, run the following command:
python run.py
The progress will be displayed on the console, and the best model checkpoint will be saved in the experiments directory.
To make predictions using the trained model, you can use the make_pred.py script. Pass a CSV file containing the input data as an argument:
python make_pred.py input_data.csv
For now only encoded values are accepted. The predicted labels for each input sample will be printed on the console.
The best validation score achieved during the training process is F1-score
(macro) of 0.992 at epoch 117. The model checkpoint with this score is
saved in the experiments/best directory.
| Num | learning_rate | batch_size | num_epochs | best f1-macro |
|---|---|---|---|---|
| 1 | 0.05 | 32 | 256 | 0.554 |
| 2 | 0.05 | 8 | 256 | 0.8125 |
| 3 | 0.05 | 128 | 256 | 0.844 |
| 4 | 0.02 | 32 | 256 | 0.856 |
| 5 | 0.2 | 32 | 256 | 0.476 |
| 6 | 0.9 | 4096 | 8192 | 0.602 |
| 7 | 0.99 | 4096 | 8192 | 0.546 |
| 8 | 1.0 | 32 | 256 | 0.313 |
| best | 0.01 | 256 | 120 | 0.992 |
Also, there were some experiments with model layers, dropout, regularization, but they were not recorded.
