InternshipCollab

Machine learning and data visualization work from my summer 2025 Python Developer internship at the University of Arkansas at Little Rock.

Project	File	What it does
Parking space classifier	image-classification-python-scikit-learn-master (2).zip	SVM that labels a parking-space image as empty or occupied
Cats vs. dogs CNN (my build)	Untitled4.ipynb	Small Keras CNN trained from scratch in Google Colab
Cats vs. dogs CNN (deeper model)	cats_vs_dogs_image_classification_using_cnn_95 (2).ipynb	4-block CNN with augmentation, callbacks, and full evaluation
Data visualization workshop	DataVisualisationTask.ipynb	Fill-in-the-blank teaching worksheet on bar charts and misleading axes
Write-up	Summary Code (2) (1).pdf	Step-by-step explanation of the parking classifier and its real-world use

---

1. Parking space classifier (scikit-learn)

Detects whether a parking space is empty or occupied, the building block of a parking lot occupancy counter.

How it works

1. Reads images from two folders, empty/ and not_empty/.
2. Resizes each image to 15×15 pixels and flattens it into a 675-value vector (15 × 15 × 3 RGB).
3. Splits the data 80/20 into training and test sets, stratified so both keep the same empty/occupied ratio.
4. Trains a Support Vector Machine (SVC) and uses GridSearchCV to tune C and gamma.
5. Prints test accuracy and saves the best model to model.p with pickle.

The included model.p is the trained model (best parameters: C=10, gamma=0.01).

Run it

unzip "image-classification-python-scikit-learn-master (2).zip"
cd image-classification-python-scikit-learn-master
pip install scikit-learn scikit-image numpy
# Set input_dir in main.py to your dataset folder (containing empty/ and not_empty/)
python main.py

Use the saved model

import pickle
from skimage.io import imread
from skimage.transform import resize

model = pickle.load(open("model.p", "rb"))
img = resize(imread("spot.jpg"), (15, 15)).flatten().reshape(1, -1)
print("occupied" if model.predict(img)[0] == 1 else "empty")

The parking-space image dataset is not included in this repo. The project is based on the Computer Vision Engineer tutorial linked in the inner README.

---

2. Cats vs. dogs CNN: my Colab build (Untitled4.ipynb)

A convolutional neural network I built and trained in Google Colab on the 25,000-image Kaggle Dogs vs. Cats dataset.

• Data: downloaded with the Kaggle API, then randomly split 80/20 into train/ and val/ folders (20,000 / 5,000 images).
• Model: 2 convolution + max-pooling blocks → Dense(512) → Dropout(0.5) → sigmoid output.
• Training: 150×150 images, horizontal-flip augmentation, Adam optimizer, binary cross-entropy, 10 epochs.
• Result: about 97% training accuracy and 80% validation accuracy. The gap shows overfitting, which the deeper model below addresses.
• Plots accuracy and loss curves, saves the model (.h5), and predicts on a single image.

Run it: open in Colab, upload your kaggle.json API key to /content, and run the cells top to bottom.

---

3. Cats vs. dogs CNN: deeper model (cats_vs_dogs_..._95.ipynb)

A more advanced pipeline, adapted from a public Kaggle notebook, that I studied to learn how to fix the overfitting in my first model.

• Split: stratified 80/10/10 train/validation/test.
• Augmentation: rotation, zoom, shear, shifts, and horizontal flip.
• Model: 4 convolution blocks (32→64→128→256 filters), each with BatchNormalization, MaxPooling, and Dropout(0.2), then Dense(512) and a 2-class softmax output.
• Callbacks: ReduceLROnPlateau and EarlyStopping.
• Results (128×128 input): 95.7% validation accuracy, 95.1% test accuracy, with precision and recall of 0.95–0.96 for both classes.
• Includes a classification report, confusion matrix, and Kaggle submission predictions.

Run it: on Kaggle with the Dogs vs. Cats competition data attached (paths expect /kaggle/input/dogs-vs-cats/).

---

4. Data visualization workshop (DataVisualisationTask.ipynb)

A beginner activity I helped run on using charts for business decisions:

• Build a revenue-by-product bar chart.
• Change the y-axis limit (200 vs. 100) to see how scale can make the same data look more or less dramatic.

This is a template. The blanks (____) are meant for participants to fill in, so the notebook raises an error until they do.

---

Tech stack

Python · scikit-learn · scikit-image · NumPy · pandas · TensorFlow / Keras · Matplotlib · Seaborn · Google Colab · Kaggle

Author

Ryan Kim, B.S. Computer Science & Business Administration, Northeastern University