Chest X-ray pneumonia classification using transfer learning on a decapitated InceptionV3, built as part of Siraj Raval's Machine Learning Bootcamp, September 2019.
Kaggle chest X-ray pneumonia dataset — binary classification across NORMAL and PNEUMONIA categories, split into train, validation, and test sets.
InceptionV3 pretrained on ImageNet, imported with include_top=False
to discard the classification head. Two Dense layers (1024 → 512 → 2)
appended for binary output. Base model layers frozen up to layer 310,
leaving upper layers trainable. Data pipeline via Keras
ImageDataGenerator. Trained for 10 epochs with Adam optimizer and
categorical crossentropy loss.
The notebook documents the process honestly including a trainability bug that held things up, a shape mismatch on the output layer, and a final test accuracy that was, in the immortal words of Meatloaf, two out of three ain't bad.
Completed for bootcamp submission, September 2019. Not maintained.