def lap(self, actors, replay_buffer, num_train, factor, prev_steps):
t = time.time() - self.t
self.total_time += t
num_act = get_num_acts(actors, prev_steps)
act_rate = factor * (num_act - self.num_act) / t
num_buffer = replay_buffer.num_add()
buffer_rate = factor * (num_buffer - self.num_buffer) / t
train_rate = factor * num_train / t
if self.iseval:
print(
"Eval Speed: train: %.1f, act: %.1f, buffer_add: %.1f, buffer_size: %d"
% (train_rate, act_rate, buffer_rate, replay_buffer.size())
)
else:
print(
"Speed: train: %.1f, act: %.1f, buffer_add: %.1f, buffer_size: %d"
% (train_rate, act_rate, buffer_rate, replay_buffer.size())
)
self.num_act = num_act
self.num_buffer = num_buffer
self.num_train += num_train
if self.iseval:
print(
"Eval Total Time: %s, %ds"
% (common_utils.sec2str(self.total_time), self.total_time)
)
else:
print(
"Total Time: %s, %ds"
% (common_utils.sec2str(self.total_time), self.total_time)
)
if self.iseval:
print(
"Eval Total Sample: train: %s, act: %s"
% (
common_utils.num2str(self.num_train),
common_utils.num2str(self.num_act),
)
)
else:
print(
"Total Sample: train: %s, act: %s"
% (
common_utils.num2str(self.num_train),
common_utils.num2str(self.num_act),
)
)
def lap2(self, actors, num_buffer, num_train):
t = time.time() - self.t
self.total_time += t
num_act = get_num_acts(actors)
act_rate = (num_act - self.num_act) / t
# num_buffer = replay_buffer.num_add()
buffer_rate = (num_buffer - self.num_buffer) / t
train_rate = num_train / t
print("Speed: train: %.1f, act: %.1f, buffer_add: %.1f" %
(train_rate, act_rate, buffer_rate))
self.num_act = num_act
self.num_buffer = num_buffer
self.num_train += num_train
print("Total Time: %s, %ds" %
(common_utils.sec2str(self.total_time), self.total_time))
print("Total Sample: train: %s, act: %s" % (common_utils.num2str(
self.num_train), common_utils.num2str(self.num_act)))
if stopwatch is not None:
torch.cuda.synchronize()
stopwatch.time("backprop & update")
replay_buffer.update_priority(priority)
if stopwatch is not None:
stopwatch.time("updating priority")
stat["loss"].feed(loss.detach().item())
stat["grad_norm"].feed(g_norm)
epoch_t = time.time() - t
train_time += epoch_t
print("epoch: %d, time: %.1fs, total time(train): %s" %
(epoch, epoch_t, common_utils.sec2str(train_time)))
tachometer.lap(actors, replay_buffer, args.epoch_len * args.batchsize)
if stopwatch is not None:
stopwatch.summary()
context.pause()
eval_locker.update_model(agent)
score = evaluate(
args.game,
args.num_eval_game,
eval_locker,
actor_cls,
epoch * args.num_eval_game + 1,
args.max_frame,
0,
terminal_on_life_loss=bool(args.one_life),