def calculate_simulation(self, mode='displacement'):
nodes_pos, edges_indices, edges_thickness, _ = self.extract_node_edge_info(
)
node_num = nodes_pos.shape[0]
assert node_num >= np.max(
edges_indices), 'edges_indicesに，ノード数以上のindexを示しているものが発生'
input_nodes, output_nodes, frozen_nodes, nodes_pos, edges_indices = remove_node_which_nontouchable_in_edge_indices(
self.input_nodes, self.output_nodes, self.frozen_nodes, nodes_pos,
edges_indices)
displacement = barfem(nodes_pos, edges_indices, edges_thickness,
input_nodes, self.input_vectors, frozen_nodes,
mode)
efficiency = calc_efficiency(input_nodes, self.input_vectors,
output_nodes, self.output_vectors,
displacement)
return efficiency
edges_thickness = np.load(os.path.join(np_save_path, "edges_thickness.npy"))
edges_thickness = edges_thickness * 3
volume = calc_volume(nodes_pos, edges_indices, edges_thickness)
print(volume)
condition_nodes_pos, input_nodes, input_vectors, output_nodes, \
output_vectors, frozen_nodes, condition_edges_indices, condition_edges_thickness\
= make_main_node_edge_info(*condition(), condition_edge_thickness=0.05) # スレンダー比を考慮し，長さ方向に対して1/20の値の幅にした
input_nodes, output_nodes, frozen_nodes, edges_thickness \
= conprocess_seperate_edge_indice_procedure(input_nodes, output_nodes, frozen_nodes, condition_nodes_pos,
condition_edges_indices, condition_edges_thickness,
nodes_pos, edges_indices, edges_thickness)
displacement = barfem(nodes_pos,
edges_indices,
edges_thickness,
input_nodes,
input_vectors,
frozen_nodes,
mode='displacement')
efficiency = calc_efficiency(input_nodes, input_vectors, output_nodes,
output_vectors, displacement)
print(efficiency)
#render_graph(nodes_pos, edges_indices, edges_thickness, os.path.join("GA/体積率に関する調査", "image.png"), display_number=False)
def actor_critic_mean(max_episodes, test_name):
"""Actor-Criticの５回実験したときの平均グラフを作成する関数"""
test_num = 5
log_dir = "confirm/step1/ac_results/{}".format(test_name)
assert not os.path.exists(log_dir), "already folder exists"
os.makedirs(log_dir, exist_ok=True)
history = {}
for i in range(test_num):
history["{}".format(i)] = {}
history["{}".format(i)]['epoch'] = []
history["{}".format(i)]['result_efficiency'] = []
node_pos, input_nodes, input_vectors,\
output_nodes, output_vectors, frozen_nodes,\
edges_indices, edges_thickness, frozen_nodes = easy_dev()
max_steps = 1
lr_actor = 1e-4
lr_critic = 1e-3
weight_decay = 1e-2
gamma = 0.99
device = torch.device('cpu')
actorNet = Edgethick_Actor().to(device)
criticNet = Edgethick_Critic().to(device)
optimizer_actor = optim.Adam(actorNet.parameters(), lr=lr_actor)
optimizer_critic = optim.Adam(criticNet.parameters(),
lr=lr_critic,
weight_decay=weight_decay)
for episode in tqdm(range(max_episodes)):
observation = np.array([0, 1])
for step in range(max_steps):
action = select_action(observation, actorNet, criticNet,
device)
edges_thickness = action['edge_thickness']
displacement = barfem(node_pos,
edges_indices,
edges_thickness,
input_nodes,
input_vectors,
frozen_nodes,
mode="force")
reward = displacement[1 * 3]
criticNet.rewards.append(reward)
loss = finish_episode(criticNet, actorNet, optimizer_critic,
optimizer_actor, gamma)
history["{}".format(i)]['epoch'].append(episode + 1)
history["{}".format(i)]['result_efficiency'].append(reward)
if episode % 1000 == 0:
print("episode:{} total reward:{}".format(episode, reward))
plot_efficiency_history(
history["{}".format(i)],
os.path.join(log_dir, 'learning_effi_curve{}.png'.format(i)))
mean = np.stack([
history["{}".format(i)]['result_efficiency'] for i in range(test_num)
])
std = np.std(mean[:, -1])
print('最終結果の標準偏差：', std)
mean = np.mean(mean, axis=0)
meanhistory = {}
meanhistory['epoch'] = history['0']['epoch']
meanhistory['result_efficiency'] = mean
# 学習履歴を保存
with open(os.path.join(log_dir, 'history.pkl'), 'wb') as f:
pickle.dump(history, f)
plot_efficiency_history(
meanhistory, os.path.join(log_dir, 'mean_learning_effi_curve.png'))
def compare_apdl_barfem(nodes_pos, edges_indices, edges_thickness,
input_nodes, input_vectors, frozen_nodes, tmax=100000, eps=1.0e-11, mode="force"):
# C言語を用いたbarfem
displacement = barfem(nodes_pos, edges_indices, edges_thickness, input_nodes,
input_vectors, frozen_nodes, mode=mode, tmax=tmax, eps=eps)
# APDLの設定
mapdl = launch_mapdl()
mapdl.finish()
mapdl.clear()
mapdl.prep7()
# 材料物性値の設定
mapdl.et(1, 3)
mapdl.mp("ex", 1, 1) # ヤング率
mapdl.mp("prxy", 1, 0.3) # ポアソン比
mapdl.mat(1)
# 節点部分の設定
for i, node_pos in enumerate(nodes_pos):
mapdl.n(i + 1, node_pos[0], node_pos[1], 0)
# エッジ部分の設定
for i, edges_indice in enumerate(edges_indices):
b = 0.2 # 奥行
h = edges_thickness[i]
mapdl.r(i + 1, b * h, b * (h * h**2) / 12, h, 0) # A,I,height=y方向の長さ,SHEARZ
mapdl.real(i + 1)
mapdl.e(edges_indice[0] + 1, edges_indice[1] + 1)
mapdl.finish()
mapdl.run('/solu')
mapdl.antype('static')
# 解析条件設定
# 固定ノード設定
for i in frozen_nodes:
mapdl.d(i + 1, "all", 0)
# 外力設定
for i, input_vector in enumerate(input_vectors):
if mode == "displacement":
mapdl.d(input_nodes[i] + 1, "UX", input_vector[0])
mapdl.d(input_nodes[i] + 1, "UY", input_vector[1])
else:
mapdl.f(input_nodes[i] + 1, "FX", input_vector[0])
mapdl.f(input_nodes[i] + 1, "FY", input_vector[1])
# 解析開始
mapdl.solve()
mapdl.finish()
# 結果出力
x_disp = mapdl.post_processing.nodal_displacement('X')
y_disp = mapdl.post_processing.nodal_displacement('Y')
z_rot = mapdl.post_processing.nodal_rotation('Z')
ansys_disp = np.stack([x_disp, y_disp, z_rot]).T.flatten()
#result = mapdl.result
#nnum, principal_nodal_stress = result.principal_nodal_stress(0)
# von_mises = principal_nodal_stress[:, -1] # von-Mises stress is the right most column
# print(von_mises)
# print(np.max(von_mises))
mapdl.exit() # ポート番号などをリセットするために必要
# 厳密には少し小さい値の部分が異なる為，allclose構文を利用
return np.allclose(ansys_disp, displacement)