Cat-Dog Image Classification

Requirements

1. Create and activate a virtual environment that uses Python 3.12.
2. Install PyTorch 2.5.1 (see this instruction).
3. Install other dependencies: PIL 10.3.0, numpy 2.0.1, polars 1.21.0.
4. (For server deployment) onnx, onnxscript (install with pip) and OpenVINO Model Server. To run the client code at ovms/client you would need Node.js and NPM.

Inference

Directly using PyTorch model

See this 03_inference.ipynb for more.

Server

Do these steps:

1. Install OpenVINO Model Server into the directory ovms of this repo (see the project's docs for instructions).

2. Start the server:

# Make sure the working directory is ovms
ovms --port 9000 --rest_port 8000 --model_name catdog --model_path models/catdog

3. After the server has been launched, run the client code

# Make sure the working directory is ovms/client
npm start ../../images/may-2.jpg

# Here "../../images/may-2.jpg" is simply the path to the image.
# Replace it with any images you want.

# This client code is a single Node.js script which only depends
# on sharp (see package.json), so it is portable.

Training

How to train from scratch:

1. Download the dataset from Microsoft Download Center (select English).
2. Extract and put the dataset folder under data/. If success, there should be two folders, data/PetImages/Cat/ for cat images and data/PetImages/Dog/ for dog images.
3. Run 01_training.ipynb to train the model on the training set and validates on the validation set.
4. Run 02_evaluation.ipynb to evaluate the model on the testing set.
5. (Optional) Export your PyTorch model to ONNX format (see 04_export_onnx.ipynb).

Images need to be preprocessed before fed to the model (see the image below). See the transform functions in 01_training.ipynb.

This is the plot of the losses and accuracy scores by epoch on the training and validation sets (see the image below).