| 8 | |
| 9 | |
| 10 | class AtariConfig(BaseConfig): |
| 11 | def __init__(self): |
| 12 | super(AtariConfig, self).__init__( |
| 13 | training_steps=100000, |
| 14 | last_steps=20000, |
| 15 | test_interval=10000, |
| 16 | log_interval=1000, |
| 17 | vis_interval=1000, |
| 18 | test_episodes=32, |
| 19 | checkpoint_interval=100, |
| 20 | target_model_interval=200, |
| 21 | save_ckpt_interval=10000, |
| 22 | max_moves=12000, |
| 23 | test_max_moves=12000, |
| 24 | history_length=400, |
| 25 | discount=0.997, |
| 26 | dirichlet_alpha=0.3, |
| 27 | value_delta_max=0.01, |
| 28 | num_simulations=50, |
| 29 | batch_size=256, |
| 30 | td_steps=5, |
| 31 | num_actors=1, |
| 32 | # network initialization/ & normalization |
| 33 | episode_life=True, |
| 34 | init_zero=True, |
| 35 | clip_reward=True, |
| 36 | # storage efficient |
| 37 | cvt_string=True, |
| 38 | image_based=True, |
| 39 | # lr scheduler |
| 40 | lr_warm_up=0.01, |
| 41 | lr_init=0.2, |
| 42 | lr_decay_rate=0.1, |
| 43 | lr_decay_steps=100000, |
| 44 | auto_td_steps_ratio=0.3, |
| 45 | # replay window |
| 46 | start_transitions=8, |
| 47 | total_transitions=100 * 1000, |
| 48 | transition_num=1, |
| 49 | # frame skip & stack observation |
| 50 | frame_skip=4, |
| 51 | stacked_observations=4, |
| 52 | # coefficient |
| 53 | reward_loss_coeff=1, |
| 54 | value_loss_coeff=0.25, |
| 55 | policy_loss_coeff=1, |
| 56 | consistency_coeff=2, |
| 57 | # reward sum |
| 58 | lstm_hidden_size=512, |
| 59 | lstm_horizon_len=5, |
| 60 | # siamese |
| 61 | proj_hid=1024, |
| 62 | proj_out=1024, |
| 63 | pred_hid=512, |
| 64 | pred_out=1024,) |
| 65 | self.discount **= self.frame_skip |
| 66 | self.max_moves //= self.frame_skip |
| 67 | self.test_max_moves //= self.frame_skip |