sample_spec = buffer_conf.memory_conf['spec']
num_hist_steps = buffer_conf.memory_conf['N']
agent = MultiStepActorCriticAgent(agent_conf,
network_conf,
brain_conf,
buffer_conf,
use_attention=use_attention)
config = RunnerConfig(map_name=map_name, reward_func=great_victor_with_kill_bonus,
state_proc_func=process_game_state_to_dgl,
agent=agent,
n_hist_steps=num_hist_steps)
runner_manager = RunnerManager(config, num_runners)
wandb.init(project="sc2rl")
wandb.config.update({'use_attention': use_attention,
'num_runners': num_runners,
'num_samples': num_samples})
wandb.config.update(agent_conf())
wandb.config.update(network_conf())
wandb.config.update(brain_conf())
wandb.config.update(buffer_conf())
iters = 0
while iters < 1000000:
iters += 1
runner_manager.sample(num_samples)
runner_manager.transfer_sample()
raise NotImplementedError(
"Not supported reward function:{}".format(reward_name))
game_state_to_dgl = partial(process_game_state_to_dgl,
use_absolute_pos=use_absolute_pos,
edge_ally_to_enemy=edge_ally_to_enemy)
config = RunnerConfig(map_name=map_name,
reward_func=reward_func,
state_proc_func=game_state_to_dgl,
agent=agent,
n_hist_steps=num_hist_steps,
gamma=gamma,
realtime=False)
runner_manager = RunnerManager(config, num_runners)
wandb.init(project="qmix2", name=exp_name)
wandb.watch(agent)
wandb.config.update({
'use_attention': use_attention,
'num_runners': num_runners,
'num_samples': num_samples,
'use_hierarchical_actor': use_hierarchical_actor,
'map_name': map_name,
'reward': reward_name,
'frame_skip_rate': frame_skip_rate,
'use_absolute_pos': use_absolute_pos,
'victory_coeff': victory_coeff
})
brain_conf = MultiStepActorCriticBrainConfig()
buffer_conf = NstepInputMemoryConfig()
sample_spec = buffer_conf.memory_conf['spec']
num_hist_steps = buffer_conf.memory_conf['N']
agent = MultiStepActorCriticAgent(agent_conf, network_conf, brain_conf,
buffer_conf)
config = RunnerConfig(map_name=map_name,
reward_func=great_victor_with_kill_bonus,
state_proc_func=process_game_state_to_dgl,
agent=agent,
n_hist_steps=num_hist_steps)
runner_manager = RunnerManager(config, 3)
wandb.init(project="sc2rl")
wandb.config.update(agent_conf())
wandb.config.update(network_conf())
wandb.config.update(brain_conf())
wandb.config.update(buffer_conf())
iters = 0
while iters < 10:
iters += 1
runner_manager.sample(10)
runner_manager.transfer_sample()
print("fit at {}".format(iters))
fit_return_dict = agent.fit()
wandb.log(fit_return_dict, step=iters)