What it does

This is the winning solution to Kaggle’s 2017 Web Traffic Time Series Forecasting competition. It predicts future Wikipedia page views using an RNN encoder-decoder with seq2seq architecture, built in TensorFlow and heavily dependent on cuDNN for GPU acceleration.

The interesting bit

The training strategy is deliberately counterintuitive: it runs in “blind mode” with no evaluation of model performance during training, and the README calmly notes that NaN losses are “normal.” The actual trick is extreme ensembling — training three models with different seeds on a single graph, saving ten checkpoints from the final 1000 steps, then averaging 30 model weights at prediction time. It’s less machine learning, more machine stubbornness.

Key highlights

• First-place finish on a high-profile Kaggle forecasting competition
• Single-graph multi-seed training with n_models=3 and checkpoint harvesting from steps 10,500–11,500
• cuDNN-dependent RNN implementation; CPU training explicitly will not work
• Modular pipeline: feature extraction (make_features.py), TF data preprocessing (input_pipe.py), model definition (model.py), and hyperparameter sets (hparams.py)
• Multiple hyperparameter configurations available (s32, definc, inst81, etc.)

Caveats

• Requires GPU; the README is unambiguous that CPU training fails
• Reproduction demands specific Kaggle dataset files (key_2.csv.zip, train_2.csv.zip) that are no longer trivially accessible
• Prediction loads and evaluates 30 model weights, so inference is deliberately slow
• The “blind mode” approach means you get no training feedback until submission time

Verdict

Worth studying if you’re building time-series ensembles or curious about Kaggle-winning heuristics. Skip it if you need a maintainable production pipeline; this is competition code, optimized for leaderboard position over engineering hygiene.