learning_rate: 3.0e-4
learning_rate_schedule: linear
max_steps: 5.0e4
memory_size: 128
normalize: false
num_epoch: 3
num_layers: 2