Views
No views yet
1# install huggingface_ding
2git clone https://github.com/opendilab/huggingface_ding.git
3pip3 install -e ./huggingface_ding/
4# install environment dependencies if needed
5
6pip3 install DI-engine[common_env,video]
7pip3 install LightZero
81# running with trained model
2python3 -u run.py1from lzero.agent import MuZeroAgent
2from ding.config import Config
3from easydict import EasyDict
4import torch
5
6# Pull model from files which are git cloned from huggingface
7policy_state_dict = torch.load("pytorch_model.bin", map_location=torch.device("cpu"))
8cfg = EasyDict(Config.file_to_dict("policy_config.py").cfg_dict)
9# Instantiate the agent
10agent = MuZeroAgent(
11 env_id="PongNoFrameskip-v4", exp_name="PongNoFrameskip-v4-MuZero", cfg=cfg.exp_config, policy_state_dict=policy_state_dict
12)
13# Continue training
14agent.train(step=5000)
15# Render the new agent performance
16agent.deploy(enable_save_replay=True)
171# running with trained model
2python3 -u run.py1from lzero.agent import MuZeroAgent
2from huggingface_ding import pull_model_from_hub
3
4# Pull model from Hugggingface hub
5policy_state_dict, cfg = pull_model_from_hub(repo_id="OpenDILabCommunity/PongNoFrameskip-v4-MuZero")
6# Instantiate the agent
7agent = MuZeroAgent(
8 env_id="PongNoFrameskip-v4", exp_name="PongNoFrameskip-v4-MuZero", cfg=cfg.exp_config, policy_state_dict=policy_state_dict
9)
10# Continue training
11agent.train(step=5000)
12# Render the new agent performance
13agent.deploy(enable_save_replay=True)
141#Training Your Own Agent
2python3 -u train.py1from lzero.agent import MuZeroAgent
2from huggingface_ding import push_model_to_hub
3
4# Instantiate the agent
5agent = MuZeroAgent(env_id="PongNoFrameskip-v4", exp_name="PongNoFrameskip-v4-MuZero")
6# Train the agent
7return_ = agent.train(step=int(500000))
8# Push model to huggingface hub
9push_model_to_hub(
10 agent=agent.best,
11 env_name="OpenAI/Gym/Atari",
12 task_name="PongNoFrameskip-v4",
13 algo_name="MuZero",
14 github_repo_url="https://github.com/opendilab/LightZero",
15 github_doc_model_url=None,
16 github_doc_env_url=None,
17 installation_guide='''
18pip3 install DI-engine[common_env,video]
19pip3 install LightZero
20''',
21 usage_file_by_git_clone="./muzero/pong_muzero_deploy.py",
22 usage_file_by_huggingface_ding="./muzero/pong_muzero_download.py",
23 train_file="./muzero/pong_muzero.py",
24 repo_id="OpenDILabCommunity/PongNoFrameskip-v4-MuZero",
25 platform_info="[LightZero](https://github.com/opendilab/LightZero) and [DI-engine](https://github.com/opendilab/di-engine)",
26 model_description="**LightZero** is an efficient, easy-to-understand open-source toolkit that merges Monte Carlo Tree Search (MCTS) with Deep Reinforcement Learning (RL), simplifying their integration for developers and researchers. More details are in paper [LightZero: A Unified Benchmark for Monte Carlo Tree Search in General Sequential Decision Scenarios](https://huggingface.co/papers/2310.08348).",
27 create_repo=True
28)
291exp_config = {
2 'main_config': {
3 'exp_name': 'PongNoFrameskip-v4-MuZero',
4 'seed': 0,
5 'env': {
6 'stop_value': 1000000,
7 'env_id': 'PongNoFrameskip-v4',
8 'env_name': 'PongNoFrameskip-v4',
9 'obs_shape': [4, 96, 96],
10 'collector_env_num': 8,
11 'evaluator_env_num': 3,
12 'n_evaluator_episode': 3,
13 'manager': {
14 'shared_memory': False
15 }
16 },
17 'policy': {
18 'on_policy': False,
19 'cuda': True,
20 'multi_gpu': False,
21 'bp_update_sync': True,
22 'traj_len_inf': False,
23 'model': {
24 'observation_shape': [4, 96, 96],
25 'frame_stack_num': 4,
26 'action_space_size': 6,
27 'downsample': True,
28 'self_supervised_learning_loss': True,
29 'discrete_action_encoding_type': 'one_hot',
30 'norm_type': 'BN'
31 },
32 'use_rnd_model': False,
33 'sampled_algo': False,
34 'gumbel_algo': False,
35 'mcts_ctree': True,
36 'collector_env_num': 8,
37 'evaluator_env_num': 3,
38 'env_type': 'not_board_games',
39 'action_type': 'fixed_action_space',
40 'battle_mode': 'play_with_bot_mode',
41 'monitor_extra_statistics': True,
42 'game_segment_length': 400,
43 'transform2string': False,
44 'gray_scale': False,
45 'use_augmentation': True,
46 'augmentation': ['shift', 'intensity'],
47 'ignore_done': False,
48 'update_per_collect': 1000,
49 'model_update_ratio': 0.1,
50 'batch_size': 256,
51 'optim_type': 'SGD',
52 'learning_rate': 0.2,
53 'target_update_freq': 100,
54 'target_update_freq_for_intrinsic_reward': 1000,
55 'weight_decay': 0.0001,
56 'momentum': 0.9,
57 'grad_clip_value': 10,
58 'n_episode': 8,
59 'num_simulations': 50,
60 'discount_factor': 0.997,
61 'td_steps': 5,
62 'num_unroll_steps': 5,
63 'reward_loss_weight': 1,
64 'value_loss_weight': 0.25,
65 'policy_loss_weight': 1,
66 'policy_entropy_loss_weight': 0,
67 'ssl_loss_weight': 2,
68 'lr_piecewise_constant_decay': True,
69 'threshold_training_steps_for_final_lr': 50000,
70 'manual_temperature_decay': False,
71 'threshold_training_steps_for_final_temperature': 100000,
72 'fixed_temperature_value': 0.25,
73 'use_ture_chance_label_in_chance_encoder': False,
74 'use_priority': True,
75 'priority_prob_alpha': 0.6,
76 'priority_prob_beta': 0.4,
77 'root_dirichlet_alpha': 0.3,
78 'root_noise_weight': 0.25,
79 'random_collect_episode_num': 0,
80 'eps': {
81 'eps_greedy_exploration_in_collect': False,
82 'type': 'linear',
83 'start': 1.0,
84 'end': 0.05,
85 'decay': 100000
86 },
87 'cfg_type': 'MuZeroPolicyDict',
88 'reanalyze_ratio': 0.0,
89 'eval_freq': 2000,
90 'replay_buffer_size': 1000000
91 },
92 'wandb_logger': {
93 'gradient_logger': False,
94 'video_logger': False,
95 'plot_logger': False,
96 'action_logger': False,
97 'return_logger': False
98 }
99 },
100 'create_config': {
101 'env': {
102 'type': 'atari_lightzero',
103 'import_names': ['zoo.atari.envs.atari_lightzero_env']
104 },
105 'env_manager': {
106 'type': 'subprocess'
107 },
108 'policy': {
109 'type': 'muzero',
110 'import_names': ['lzero.policy.muzero']
111 }
112 }
113}
114