os.path.join(
model_dir,
'pg-{}-TSP{}-epoch-{}.pt'.format(str(id), args.n_points,
epoch)))
torch.save(
policy,
os.path.join(
model_dir, 'full-model-pg-{}-TSP{}-epoch-{}.pt'.format(
str(id), args.n_points, epoch)))
best_running_reward = val_rwd_log.exp_avg
val_best_dist = val_best_dist_log.val
best_gap = gap
if epoch % args.log_interval == 0:
train_rwd_log.log(log)
train_init_dist_log.log(log)
train_best_dist_log.log(log)
train_policy_loss_log.log(log)
train_entropy_loss_log.log(log)
train_value_loss_log.log(log)
train_loss_log.log(log)
val_rwd_log.log(log)
val_init_dist_log.log(log)
val_best_dist_log.log(log)
print(
'\033[1;32;40m Train - epoch:{} |rwd: {:.2f}'.format(
epoch, train_rwd_log.val),
step_list = [1, 8, 16, 24, 50, 100]
print('training start now: ')
for n_step in step_list:
i = -1
env = EnvironmentState()
state_set = StateDataset(env,
skip_step=n_step,
size=7000000,
random_torque=True,
remove_torque=False)
state_set_loader = DataLoader(state_set, batch_size=1024)
predictor = Predictor(state_set.output_size, 2)
predictor.stepper.recurrent_step = 0
avm = AverageMeter()
with Timer():
for s in state_set_loader:
i += 1
d = s['s1'][:, 12:14] - s['s0'][:, 12:14] # with torques
r, loss = predictor.optimize(s['s0'], d)
avm.log(loss)
if i % 10 == 0:
print(
f'epoch {i}: trn loss {avm.value:.4f} {avm.std:.4f}, rmse {avm.value**0.5:.4f}, '
f'pred to stat error ratio: {torch.mean(torch.abs(r / d)):.4f} '
f'max_d {torch.max(d):.2f} {torch.min(d):.2f}, max_r {torch.max(r):.2f} {torch.min(r):.2f}'
)
predictor.save(f'predictor{n_step}step.pt')
print('training finished.')
