What it does

This is a walkthrough for the classic Kaggle “Dogs vs. Cats” competition. It takes 25,000 labeled training images, shuffles them into class folders via symbolic links (no duplicate disk space), then runs three pre-trained ImageNet models as fixed feature extractors. The resulting 6,144-dimensional vectors get fed into a single dropout layer and a sigmoid output. Validation accuracy lands at 99.6%.

The interesting bit

The author treats heavy CNNs as one-off compilers: extract once, save to HDF5, then train downstream on a CPU in under ten seconds per epoch. It’s a pragmatic hack from the Keras 1.2 era, complete with a clipped LogLoss trick (0.005–0.995) to dodge Kaggle’s penalty for overconfident wrong answers.

Key highlights

• Pre-trains on ResNet50, InceptionV3 and Xception, concatenating their gap (GlobalAveragePooling) outputs
• Symbolic-link preprocessing avoids copying 25,000 images
• Full training of the top classifier runs in ~1 second per epoch after feature extraction
• Kaggle submission score of 0.04141, reportedly ranking 20th out of 1,314 at the time
• Includes a neat Graphviz model diagram showing the three-tower merge

Caveats

• Locked to Keras 1.2.2; newer versions will need syntax tweaks (e.g., nb_sample → samples, nb_epoch → epochs)
• No fine-tuning or data augmentation shown, and the README notes these as obvious next steps
• Feature extraction still needs a GPU or cloud instance (10–20 min per model on AWS p2.xlarge)

Verdict

Worth a skim if you’re teaching transfer learning or need a dead-simple ensemble baseline. Skip it if you want modern PyTorch, end-to-end fine-tuning, or a production pipeline—this is a 2017 tutorial frozen in amber.