Views
No views yet
HAPPO agent playing smacv2_zerg_10_vs_10.marl-ppo-suite training
code.1# 1. install the codebase (directly from GitHub)
2pip install "marl-ppo-suite @ git+https://github.com/legalaspro/marl-ppo-suite"
3
4# 2. get the weights & config from HF
5wget https://huggingface.co/<repo-id>/resolve/main/final-torch.model
6wget https://huggingface.co/<repo-id>/resolve/main/config.json
7
8# 3-a. Generate a StarCraft II replay file 1 episode in starcraft replay folder
9marl-train --mode render --model final-torch.model --config config.json --render_episodes 1 \
10
11# 3-b. generate additionally video drawn from frames
12marl-train --mode render --model final-torch.model --config config.json --render_episodes 1 --render_mode rgb_array final-torch.model – PyTorch checkpointreplay.mp4 – gameplay of the final policyconfig.json – training configtensorboard/ – full logs1{
2 "clip_param": 0.05,
3 "data_chunk_length": 10,
4 "entropy_coef": 0.01,
5 "fc_layers": 2,
6 "gae_lambda": 0.95,
7 "gamma": 0.99,
8 "hidden_size": 64,
9 "lr": 0.0005,
10 "n_steps": 200,
11 "num_mini_batch": 1,
12 "ppo_epoch": 5,
13 "reward_norm_type": "efficient",
14 "seed": 1,
15 "state_type": "AS",
16 "use_reward_norm": true,
17 "use_rnn": true,
18 "use_value_norm": false,
19 "value_norm_type": "welford"
20}