Working through Second Edition François Chollet
The fit_residual idea is to build a model that has direct access to baseline results - in the hope that the model can learn to correct the errors of the baseline with a result that is;
- more accurate than without
fit_residualand/or - achieved with less compute / smaller models.
i.e. this is feature engineering and problem framing engineering to inject useful assumptions into the model.
- 10.2_temperature_forecasting.ipynb
- creates a fully connect model that can do better than the non-machine learning baseline
- 10.2_temperature_forecasting_part2.ipynb
- shows how
fit_residualcan be used to get a single layer LSTM to the same test MAE as stacked GRU - we also use the date features to improve MAE
- shows how
- 10.2_california_housing.ipynb
- shows that the
fit_residualwon't work on all datasets (o:
- shows that the
I use this for trying things out on my CPU-only windows machine.
conda create -n tf python==3.9 -y
conda activate tf
pip install tensorflow
pip install nbdev jupyterlab pandas matplotlib
pip install --upgrade jupyter notebook