def _run_inference(self) -> MultiObserver:
try:
n_inputs = int(self.widget_text_n_inputs.value)
n_experts = int(self.widget_text_n_experts.value)
n_rollouts = int(self.widget_text_n_rollouts.value)
observer = MultiObserver()
params = Params(**self.state.sacred_reader.config)
params.n_experts = n_experts
params.task_size = n_inputs
self.state.inference_config = dataclasses.asdict(params)
agent = create_agent(params)
self.state.sacred_reader.load_model(agent, 'agent',
self.state.epoch)
rollout_size = n_rollouts # 15 + params.rollout_size
run_inference(params, agent, observer, rollout_size)
return observer
except ValueError as e:
print(f'ValueError: {e}')
def load_agent(p: Params, file: str):
agent = create_agent(p)
agent.load_state_dict(torch.load(file))
return agent
def run_inference(p: Params, agent: SearchAgent, observer: MultiObserver, rollout_size: int) -> float:
with torch.no_grad():
task = create_task(p)
inner_loop = LearningLoop()
agent.init_rollout(p.batch_size, p.n_experts, SearchAgentInitRolloutParams(True))
err = inner_loop.train_fixed_steps(agent, task, rollout_size, 0.0, p.learning_rollout_steps_clip, observer)
return err.item()
def load_agent(p: Params, file: str):
agent = create_agent(p)
agent.load_state_dict(torch.load(file))
return agent
p = Params(random.randint(0, 2 ** 32))
agent = create_agent(p)
task = create_task(p)
inner_loop = LearningLoop()
loss = []
# %%
# for epoch in tqdm(range(1, p.epochs + 1)):
for epoch in range(1, p.epochs + 1):
observer = None
agent.optim.zero_grad()
agent.init_rollout(p.batch_size, p.n_experts, SearchAgentInitRolloutParams(True))
rollout_size = p.rollout_size
err = inner_loop.train_fixed_steps(agent, task, rollout_size, p.learning_exp_decay,
p.learning_rollout_steps_clip, None)