def main():
# running_reward = 0
prior_efficiency = 0
continuous_trigger = 0
best_efficiency = -1000
best_epoch = 0
# １エピソードのループ
while(1):
new_node_pos, new_input_nodes, new_input_vectors, new_output_nodes, new_output_vectors, new_frozen_nodes, new_edges_indices, new_edges_thickness = make_continuous_init_graph(origin_nodes_positions, origin_edges_indices, origin_input_nodes, origin_input_vectors,
origin_output_nodes, origin_output_vectors, origin_frozen_nodes, EDGE_THICKNESS)
env = FEMGym(new_node_pos,
new_edges_indices, new_edges_thickness)
env.reset()
if env.confirm_graph_is_connected():
break
nodes_pos, _, _, _ = env.extract_node_edge_info()
first_node_num = nodes_pos.shape[0]
# run inifinitely many episodes
for epoch in tqdm(range(train_num)):
# for epoch in count(1):
# reset environment and episode reward
while(1):
new_node_pos, new_input_nodes, new_input_vectors, new_output_nodes, new_output_vectors, new_frozen_nodes, new_edges_indices, new_edges_thickness = make_continuous_init_graph(origin_nodes_positions, origin_edges_indices, origin_input_nodes, origin_input_vectors,
origin_output_nodes, origin_output_vectors, origin_frozen_nodes, EDGE_THICKNESS)
env = FEMGym(new_node_pos,
new_edges_indices, new_edges_thickness)
env.reset()
if env.confirm_graph_is_connected():
break
state = env.reset()
ep_reward = 0
continuous_trigger = 0
# for each episode, only run 9999 steps so that we don't
# infinite loop while learning
for t in range(max_action):
# select action from policy
action = select_action(first_node_num)
nodes_pos, edges_indices, edges_thickness, adj = env.extract_node_edge_info()
# take the action
state, _, done, info = env.step(action)
if (t == (max_action-1)) and (done is not True): # max_action内にてactionが終わらない時
reward = -final_penalty
elif env.confirm_graph_is_connected():
efficiency = env.calculate_simulation()
if continuous_trigger == 1:
reward = efficiency-prior_efficiency
else:
reward = efficiency+continuous_reward
continuous_trigger = 1
prior_efficiency = efficiency
elif continuous_trigger == 1:
reward = -penalty
else:
reward = 0
GCN.rewards.append(reward)
ep_reward += reward
if done:
steps = t
break
# update cumulative reward
# running_reward = 0.05 * ep_reward + (1 - 0.05) * running_reward
# perform backprop
loss = finish_episode()
# efficiencyの最終結果を求める
if env.confirm_graph_is_connected():
result_efficiency = env.calculate_simulation()
else:
result_efficiency = -1
if best_efficiency < result_efficiency:
best_epoch = epoch
best_efficiency = result_efficiency
save_model(save_name="Good")
env.render(os.path.join(
log_dir, 'render_image/{}.png'.format(epoch+1)))
history['epoch'].append(epoch+1)
history['loss'].append(loss)
history['ep_reward'].append(ep_reward)
history['result_efficiency'].append(result_efficiency)
history['steps'].append(steps+1)
# 学習履歴を保存
with open(os.path.join(log_dir, 'history.pkl'), 'wb') as f:
pickle.dump(history, f)
with open(os.path.join(log_dir, "progress.txt"), mode='a') as f:
f.writelines('epoch %d, loss: %.4f ep_reward: %.4f result_efficiency: %.4f\n' %
(epoch + 1, loss, ep_reward, result_efficiency))
with open(os.path.join(log_dir, "represent_value.txt"), mode='w') as f:
f.writelines('epoch %d, best_efficiency: %.4f\n' %
(best_epoch+1, best_efficiency))
save_model(save_name="Last")
plot_loss_history(history, os.path.join(
log_dir, 'learning_loss_curve.png'))
plot_reward_history(history, os.path.join(
log_dir, 'learning_reward_curve.png'))
plot_efficiency_history(history, os.path.join(
log_dir, 'learning_effi_curve.png'))
plot_steps_history(history, os.path.join(
log_dir, 'learning_steps_curve.png'))
# １エピソードのループ
state = env.reset()
env.confirm_graph_is_connected()
env.render("fem_images/image_first.png")
total_time = 0
total_calc_time = 0
for i in range(500):
# ランダム行動の取得
action = env.random_action()
# １ステップの実行
state, reward, done, info = env.step(action)
if env.confirm_graph_is_connected():
reward = 0
start = time.time()
efficiency = env.calculate_simulation()
elapsed_time = time.time() - start
print("elapsed_time:{0}".format(elapsed_time) + "[sec]")
reward = efficiency
total_time += elapsed_time
total_calc_time += 1
else:
reward = -1
print("一回辺りの計算時間:", total_time/total_calc_time)
# エピソード完了
# if done:
# print('done')
# break
])
origin_frozen_nodes = [1, 3, 5, 7, 9, 11, 13, 15]
# gymに入力する要素を抽出
new_node_pos, new_input_nodes, new_input_vectors, new_output_nodes, new_output_vectors, new_frozen_nodes, new_edges_indices, new_edges_thickness = make_main_node_edge_info(
origin_nodes_positions, origin_edges_indices, origin_input_nodes,
origin_input_vectors, origin_output_nodes, origin_output_vectors,
origin_frozen_nodes, EDGE_THICKNESS)
env = FEMGym(new_node_pos, new_edges_indices, new_edges_thickness)
# １エピソードのループ
state = env.reset()
env.confirm_graph_is_connected()
env.render("fem_images/image_first.png")
for i in range(100):
# ランダム行動の取得
action = env.random_action()
# １ステップの実行
state, reward, done, info = env.step(action)
if env.confirm_graph_is_connected():
reward = env.calculate_simulation()
# env.render("fem_images/graph_connected{}.png".format(i))
print('{}steps reward:{}'.format(i, reward))
# エピソード完了
# if done:
# print('done')
# break