HW2 Phoneme Classification

Grading

Info

1. Simple: 0.45797
2. Medium: 0.69747
3. strong: 0.75028
4. Boss: 0.82324

Solution

Model	HyperParameters	Train acc	Valid acc	Public Score	Private Score
MLP	nframes = 21, hidden_layers=3, batchnorm, dropout=0.25	…	…	0.73379	0.73567
LSTM	nframes = 21, hidden_layers=2, fc_dropout=0.30	0.84755	0.73834	0.74466	0.74330
GRU	nframes = 21, hidden_layers=2, fc_dropout=0.30	0.82196	0.74969	0.75266	0.75265
GRU bi	nframes = 21, hidden_layers=2, fc_dropout=0.30	0.82323	0.75220	0.75542	0.75627

Reflection

Note

1. Bigger model makes better performance. I want to try more complecated model but my comuter doesn’t alow. 😟
2. I used some tricks to seperate the traning process. (1) set larger n-frames and make deeper GRUbi model, (2)then seperate training data into 5 pecies and train 5 times using the last trained model params. But it didn’t work well. The score was only about 0.757.
3. I did better than first homework because I made things organized.
4. Keep going.

Code

HW2 strong