def _generate_reset_input(self, training, config: ArenaConfig) -> UnityRLInput:
rl_in = UnityRLInput()
rl_in.is_training = training
rl_in.command = 1
rl_reset = UnityRLResetInput()
if (config is not None):
rl_reset.CopyFrom(config.dict_to_arena_config())
result = UnityInput()
result.rl_input.CopyFrom(rl_in)
result.rl_reset_input.CopyFrom(rl_reset)
return result
def _generate_step_input(self, vector_action, memory, text_action, value) -> UnityRLInput:
rl_in = UnityRLInput()
for b in vector_action:
n_agents = self._n_agents[b]
if n_agents == 0:
continue
_a_s = len(vector_action[b]) // n_agents
_m_s = len(memory[b]) // n_agents
for i in range(n_agents):
action = AgentActionProto(
vector_actions=vector_action[b][i * _a_s: (i + 1) * _a_s],
memories=memory[b][i * _m_s: (i + 1) * _m_s],
text_actions=text_action[b][i],
)
if b in value:
if value[b] is not None:
action.value = float(value[b][i])
rl_in.agent_actions[b].value.extend([action])
rl_in.command = 0
return self.wrap_unity_input(rl_in)