Poor Results for IMP and Transformer

[zhanwenchen]

• [x] Match Transformer config
• [x] Redeploy Transformer
• [ ] Match IMP config
• [ ] Redeploy IMP

[zhanwenchen]

Diffs:

1. DTYPE: float16 => DTYPE: float32
2. BASE_LR: 1e-3 => BASE_LR: 1e-4
3. SOLVER.SCHEDULE.TYPE: WarmupReduceLROnPlateau => SOLVER.SCHEDULE.TYPE: WarmupMultiStepLR

[zhanwenchen]

Still only 13.93. Maybe match the model_transformer implementation? Close this ticket and work in #115 instead.