Add more changes to the model so that it is consistent with paper

[auroracramer]

• Move ReLU after batch norm
• Use weight decay (assumed L2)
• Set learning rate to 1e-4
• Allow for volume gain less than 0dB, as paper seems like it should allow for that

[hohsiangwu]

I would also recommend that to change all the print in the train function to be logging so it will show up while training using sbatch.

[auroracramer]

If you specify an out file in the sbatch file, it will also save stdout from the script. But it would indeed be more robust to use logging rather than print statements.