Obsessed with my Kaggle score – it’s just a number
September 23, 2016
Posted by on
“a state in which someone thinks about something constantly or frequently especially in a way that is not normal”
Three models grid search on 2 ubuntu boxes with a GTX1070 each
First model I used was a ConvNet (CNN) with multiple Recurrent (RNN) layers on top.
Apparently, this is a common structure NN in speech. The CCN–>RNN tries to exploit temporal relationships within the data. In addition with the 2D structure of the spectrograms.
Imagine for each vertical line in the spectrogram as a timestep, the RNN is an attempt to model the relationships between those timesteps.
The second model, I ripped out the RNN and replaced it with LSTM. The third model, I tried to increase the RNN layers.
The fourth model combines all the electrodes in batches of 64. Fifth model is straight MLP in Tensorflow and Sixth model is straight CNN Tensorflow.
AWS p2.xlarge is a pain, just to try out a pretrained VGG model need 12gb GPU
Bad Data – Leaderboard hacking