Abs loss is de-normalized, and is not used as a loss metric. Other regression losses are normalized.
- Code: regression.py
- DATADIR = twparmbeatmC1_no_nan.csv
- train_size = 0.70
- num_epoch = 100000
- n_hid = [n_in = 151, (16), n_out = 1] (hidden layer only exist in run 11.4)
- run_ID = 11.3, [11.4] (./log/11.4)
- connections = ['fc']
- act_funcs = ['relu']
- loss_funcs = ['square_l2']
- learning_rates = [1e-3]
- beta = 0.01
- plt.plot = True precipitation vs Predicted precipitation
- Mean abs loss = 1.24 (run 11.3), 1.47 (run 11.4), same old collapse-to-zero phenomenon
- All .ipynb files are in colab/.