# Computing actions by using FDDR
delta = fddr(fragments, running_mean=mean,
running_var=var).double().squeeze(-1)
# Computing reward
pad_delta = F.pad(delta, [1, 0])
delta_diff = (pad_delta[:, 1:] - pad_delta[:, :-1])
reward = torch.sum(delta * returns - c * torch.abs(delta_diff))
test_reward_meter.append(reward.item())
progress_bar.set_description(
'[Epoch %d][Iteration %d][Reward: train = %.4f, test = %.4f]' %
(e, i, train_reward_meter.get_average(-1),
test_reward_meter.get_average(-1)))
if e % save_per_epoch == 0:
torch.save(fddr.state_dict(), os.path.join(log_src, 'fddrl.pkl'))
train_reward_meter.step()
test_reward_meter.step()
# Save the model and reward history
torch.save(fddr.state_dict(), os.path.join(log_src, 'fddrl.pkl'))
np.save(os.path.join(log_src, 'fddrl_reward.npy'),
train_reward_meter.get_average())
# Plot the reward curve
plt.plot(train_reward_meter.get_average())
plt.plot(test_reward_meter.get_average())
plt.show()
# Computing actions by using FDDR
delta = drl(fragments).double().squeeze(-1)
# Computing reward
pad_delta = F.pad(delta, [1, 0])
delta_diff = (pad_delta[:, 1:] - pad_delta[:, :-1])
reward = torch.sum(delta * returns - c * torch.abs(delta_diff))
# Updating FDDR
optimizer.zero_grad()
(-reward).backward()
optimizer.step()
# Recording and showing the information
reward_meter.append(reward.item())
progress_bar.set_description(
'[Epoch %d][Iteration %d][Reward: %.4f]' % (e, i, reward_meter.get_average(-1)))
progress_bar.update()
if e % save_per_epoch == 0:
torch.save(drl.state_dict(), os.path.join(log_src, 'drl.pkl'))
reward_meter.step()
# Save the model and reward history
torch.save(drl.state_dict(), os.path.join(log_src, 'drl.pkl'))
np.save(os.path.join(log_src, 'drl_reward.npy'), reward_meter.get_average())
# Plot the reward curve
plt.plot(reward_meter.get_average())
plt.show()
