We were tasked with building an image classifier that would be able to classify images to 5 classes with above 85% accuracy. The dataset consisted of 22,500 images all equally distributed within the following classes: bell, dog, horse, house, and tiger. The dataset was further stratified into train, test, and validate sets. Our final model achieved 90% accuracy on the holdout test set, we placed 2nd in the entire class. In each stage of this project, we iterated upon our best models, making sure that each team member experimented with parallel parameters of the model to ensure that models could be combined at the end of each stage.
You can read more about our project in the Final Report, view our Presentation Slides for a quicker summary of our work, or see the implementation in our Jupyter Notebook.
The following Python libraries/frameworks were used for this project: TensorFlow, Keras, Scikit-Learn, NumPy, Matplotlib
Here is a breakdown of how we iterated each model to increase the accuracy from 56% to 90%. The project was spit into 5 stages, with stages 3, 4, and 5 soley for training the models. Stages 1 and 2 comrpised of retreiving the data from the college server/uploading it to GDrive, and splitting the dataset into test, train, and validate sets.
In the end we ended up training over 20 models using Python's TensorFlow Keras library, the entire process took ~200hours
