Train model:
python3 main.py --train=True --model_name='multi-rnn' --device_name='iss GPU' --kernel_initializer='he_normal'
Evaluate model:
python3 main.py --train=False --model_name='multi-rnn' --device_name='iss GPU' --kernel_initializer='he_normal'
Ensemble learning, evaluate and visualization:
python3 ensemble.py
Please see details in presentation slides.
Results of ensemble model with He_normal Glorot_uniform and Glorot_normal initializers:
Architecture | LSTM | GRU |
---|---|---|
Test Accuracy | 91.73% | 94.48% |
Corresponding Hyperparameters for the best GRU-based model:
GRU layers | Dense layers | GRU units | Dense units | Window size | Shift size | Dropout rate | Val accuracy |
---|---|---|---|---|---|---|---|
2 | 3 | 512 | 256 | 250 | 125 | 0.471 | 92.9% |
Visualization of test label and predictions: