def multi_solve_environment(self):
workers_top20 = []
for arch_epoch in range(self.arch_epochs):
results_queue = Queue()
processes = []
for episode in range(self.episodes):
actions_p, actions_log_p, actions_index = self.controller.sample(
)
actions_p = actions_p.cpu().numpy().tolist()
actions_log_p = actions_log_p.cpu().numpy().tolist()
actions_index = actions_index.cpu().numpy().tolist()
if episode < self.episodes // 3:
worker = Worker(actions_p, actions_log_p, actions_index,
self.args, 'cuda:0')
elif self.episodes // 3 <= episode < 2 * self.episodes // 3:
worker = Worker(actions_p, actions_log_p, actions_index,
self.args, 'cuda:1')
else:
worker = Worker(actions_p, actions_log_p, actions_index,
self.args, 'cuda:3')
process = Process(target=consume, args=(worker, results_queue))
process.start()
processes.append(process)
for process in processes:
process.join()
workers = []
for episode in range(self.episodes):
worker = results_queue.get()
worker.actions_p = torch.Tensor(worker.actions_p).to(
self.device)
worker.actions_index = torch.LongTensor(
worker.actions_index).to(self.device)
workers.append(worker)
for episode, worker in enumerate(workers):
if self.baseline == None:
self.baseline = worker.acc
else:
self.baseline = self.baseline * self.baseline_weight + worker.acc * (
1 - self.baseline_weight)
# sort worker retain top20
workers_total = workers_top20 + workers
workers_total.sort(key=lambda worker: worker.acc, reverse=True)
workers_top20 = workers_total[:20]
top1_acc = workers_top20[0].acc
top5_avg_acc = np.mean(
[worker.acc for worker in workers_top20[:5]])
top20_avg_acc = np.mean([worker.acc for worker in workers_top20])
logging.info(
'arch_epoch {:0>3d} top1_acc {:.4f} top5_avg_acc {:.4f} top20_avg_acc {:.4f} baseline {:.4f} '
.format(arch_epoch, top1_acc, top5_avg_acc, top20_avg_acc,
self.baseline))
for i in range(5):
print(workers_top20[i].genotype)
for ppo_epoch in range(self.ppo_epochs):
loss = 0
for worker in workers:
actions_p, actions_log_p = self.controller.get_p(
worker.actions_index)
loss += self.cal_loss(actions_p, actions_log_p, worker,
self.baseline)
loss /= len(workers)
logging.info('ppo_epoch {:0>3d} loss {:.4f} '.format(
ppo_epoch, loss))
self.adam.zero_grad()
loss.backward()
self.adam.step()
