Repository navigation
FAQ
Why freeze the VGG16 backbone instead of fine-tuning it? Freezing the convolutional base is the standard "feature extraction" flavor of transfer learning — it reuses the general-purpose visual features VGG16 already learned from ImageNet without risking overfitting the relatively small dogs-vs-cats training set by updating millions of pretrained weights.
Where do I get the dataset?
It's not included in the repository due to size. Use a dogs-vs-cats style dataset (e.g. Kaggle's "Dogs vs. Cats") and arrange it into trainset/{cat,dog}/ and validset/{cat,dog}/ folders as described in Getting Started.
Why does validation accuracy bounce around instead of steadily improving? With only 10 epochs, a fixed learning rate, and heavy data augmentation, some epoch-to-epoch variance in validation accuracy is expected — the model is still learning overall (see the epoch-by-epoch numbers in Methodology), but the smaller dataset size for this exercise doesn't smooth it into a perfectly monotonic curve.
Can this be extended to more than two classes?
Yes — the general approach (frozen backbone + custom head) extends directly to multi-class problems by swapping the final Dense(1, sigmoid) for Dense(n_classes, softmax) and using categorical cross-entropy loss.
Why doesn't CI actually train the model? Training needs the external image dataset (not checked into the repo for size reasons) and would be too slow/resource-intensive for a shared CI runner. CI instead checks that the notebook is valid and that converting it to a script byte-compiles cleanly.
What are the .pptx and .mp4 files for?
They're supplementary material — presentation slides and a video walkthrough summarizing the approach and results, useful for a quick overview without reading the full notebook.