def on_rollout(self, r, at_steps):
state_arrays = r['state_arrays']
if len(state_arrays) == 1:
obs_of_step = state_arrays[0]
elif len(state_arrays) == 2:
obs_of_step = [numpy.concatenate([p,q])
for p,q in zip(state_arrays[0], state_arrays[1])]
else:
raise ValueError("Invalid state_arrays format")
obs_text = format_observations([self.uncook_observation(obs)
for obs in obs_of_step])
reward = numpy.sum(r['rewards']) * lot_scheduling.REWARDSCALE
duration = numpy.sum(r['durations'])
r_rate = reward / (duration / 3600.0) if duration > 0 else 0.0
actions_text = format_actions(r['actions'])
ep_steps = len(r['rewards'])
if self.losses_text == '':
self.losses_text = format_losses(self.losses, self.batch_count)
self.losses = defaultdict(float)
self.batch_count = 0
print("{} episode {} at {} steps: reward={:.3f} rate={:.3f} duration={:.3f} steps={} {} {} {}"
.format(iso8601stamp(),
self.rollout_count, at_steps, reward, r_rate, duration,
ep_steps, actions_text, obs_text, self.losses_text))
ep = self.rollout_count
self.rollout_count += 1
return self.check_best(ep, at_steps, reward, duration)
total_loss_p += loss_p.item()
del inputs, actions, targets
total_loss_i /= nbatches
total_loss_s /= nbatches
total_loss_p /= nbatches
return total_loss_i, total_loss_s, total_loss_p
print('[training]')
for i in range(200):
t0 = time.time()
train_losses = train(opt.epoch_size, opt.npred)
valid_losses = test(int(opt.epoch_size / 2))
n_iter += opt.epoch_size
model.cpu()
torch.save(
{
'model': model,
'optimizer': optimizer.state_dict(),
'n_iter': n_iter
}, opt.model_file + '.model')
if (n_iter / opt.epoch_size) % 10 == 0:
torch.save(model, opt.model_file + f'.step{n_iter}.model')
model.cuda()
log_string = f'step {n_iter} | '
log_string += utils.format_losses(*train_losses, split='train')
log_string += utils.format_losses(*valid_losses, split='valid')
print(log_string)
utils.log(opt.model_file + '.log', log_string)
if opt.test_only == 1:
print('[testing]')
valid_losses = test(10, 200)
else:
print('[training]')
utils.log(opt.model_file + '.log', f'[job name: {opt.model_file}]')
npred = opt.npred if opt.npred != -1 else 16
for i in range(500):
bsize = get_batch_size(npred)
dataloader.opt.batch_size = bsize
train_losses = train(opt.epoch_size, npred)
valid_losses = test(int(opt.epoch_size / 2), npred)
n_iter += opt.epoch_size
model.intype('cpu')
torch.save({'model': model,
'optimizer': optimizer.state_dict(),
'opt': opt,
'npred': npred,
'n_iter': n_iter},
opt.model_file + '.model')
model.intype('gpu')
log_string = f'step {n_iter} | npred {npred} | bsize {bsize} | esize {opt.epoch_size} | '
log_string += utils.format_losses(train_losses[0], train_losses[1], split='train')
log_string += utils.format_losses(valid_losses[0], valid_losses[1], split='valid')
print(log_string)
utils.log(opt.model_file + '.log', log_string)
if i > 0 and(i % opt.curriculum_length == 0) and (opt.npred == -1) and npred < 400:
npred += 8
