def forward_test():
try:
f = forward.forward_gpu()
coef = forward.filter_coefficient()
geo = forward.geoelectric_model()
data = forward.forward_data()
geo_path = [
'../test_data/forward_model1.json',
'../test_data/forward_model2.json',
'../test_data/forward_model3.json'
]
geos = []
responses = []
for i in range(len(geo_path)):
geos.append(forward.geoelectric_model())
responses.append(forward.forward_data())
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
data.generate_time_stamp_by_count(-5, -2, 40)
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.load_time_stamp(data)
for i in range(len(geo_path)):
geos[i].load_from_file(geo_path[i])
f.load_geo_model(geos[i])
f.forward()
responses[i] = f.get_result_magnetic()
responses[i].name = geos[i].name
fig = draw_resistivity(*geos)
fig.show()
fig = draw_forward_result(*responses)
fig.show()
# print(loss(responses[0], responses[1]))
# print(loss(responses[1], responses[2]))
# fig.show()
# fig = draw_forward_result(m)
#
# fig.show()
except Exception as e:
traceback.print_exc()
log.error(repr(e))
finally:
log.debug('forward_test finished')
def forward_test():
try:
f = forward.forward_gpu()
coef = forward.filter_coefficient()
geo = forward.geoelectric_model()
data = forward.forward_data()
m_data = forward.forward_data()
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
geo.load_from_file('../test_data/forward_model1.json')
data.generate_time_stamp_by_count(-5, -2, 20)
f.load_time_stamp(data)
f.load_geo_model(geo)
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.forward()
data = f.get_result_magnetic()
data.name = 'CUDA'
m_data.resize(data.count)
m_data.time = data.time
m_data.idx = data.idx
m_data.name = 'Matlab'
m_data.set_item_s('response', [
1.127565, 0.998018, 0.894319, 0.805530, 0.723312, 0.642532,
0.561468, 0.481032, 0.403363, 0.330769, 0.265176, 0.207871,
0.159425, 0.119744, 0.088197, 0.063800, 0.045399, 0.031824,
0.022005, 0.015023
])
fig = draw_resistivity(geo)
fig.show()
fig = draw_forward_result(data, m_data)
fig.show()
print('loss:')
print(loss(m_data, data))
except Exception as e:
traceback.print_exc()
log.error(repr(e))
finally:
log.debug('forward_test finished')
def forward_test():
try:
f = forward.forward_gpu()
coef = forward.filter_coefficient()
geo = forward.geoelectric_model()
data = forward.forward_data()
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
geo.load_from_file('../test_data/forward_model1.json')
data.generate_time_stamp_by_count(-5, -2, 20)
f.load_time_stamp(data)
f.load_geo_model(geo)
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.forward()
data = f.get_result_magnetic()
data.name = 'Model'
fig = draw_resistivity(geo)
fig.show()
fig = draw_forward_result(data)
fig.show()
except Exception as e:
traceback.print_exc()
log.error(repr(e))
finally:
log.debug('forward_test finished')
def forward_test():
try:
# 正演类
f = forward.forward_gpu()
coef = forward.filter_coefficient()
geo = forward.geoelectric_model()
geo2 = forward.geoelectric_model()
iso = forward.isometric_model()
data = forward.forward_data()
# 各种数据加载
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
geo.load_from_file('../test_data/test_geo_model.json')
geo2.load_from_file('../test_data/test_geo_model2.json')
iso.load_from_file('../test_data/test_iso_model.json')
data.generate_time_stamp_by_count(-5, -2, 20)
# 测试绘制地电模型
fig = draw_resistivity(geo,
geo2,
forward.iso_to_geo(iso),
last_height=300)
fig.show()
# 正演类加载数据
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.load_geo_model(geo)
f.load_time_stamp(data)
# 正演开始
f.forward()
# 获得结果
m = f.get_result_late_m()
e = f.get_result_late_e()
m.name = 'late_m'
e.name = 'late_e'
# 绘制结果
fig = draw_forward_result(m, e)
fig.show()
add_noise(m, 0.1)
fig = draw_forward_result(m)
fig.show()
except Exception as e:
traceback.print_exc()
log.error(repr(e))
finally:
log.debug('forward_test finished')
def forward_test():
try:
f = forward.forward_gpu()
coef = forward.filter_coefficient()
geo = forward.geoelectric_model()
data = forward.forward_data()
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
geo.load_from_file('../test_data/forward_model1.json')
data.generate_time_stamp_by_count(-5, 0, 100)
f.load_time_stamp(data)
f.load_geo_model(geo)
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.forward()
counts = [40, 60, 80, 100]
layers = [5, 6, 8, 10, 12]
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, xlabel='layer count', ylabel='time/s')
for c in counts:
data.generate_time_stamp_by_count(-5, 0, c)
f.load_time_stamp(data)
t = []
for l in layers:
geo.resize(l)
res = npy.ones([l]) * 100
hei = npy.ones([l]) * 50
geo.set_item_s('resistivity', list(res))
geo.set_item_s('height', list(hei))
f.load_geo_model(geo)
s = time.clock()
f.forward()
e = time.clock()
t.append(e - s)
ax.plot(layers, t, label=str(c), marker='+')
ax.legend()
plt.savefig(figure=fig, fname='compare')
except Exception as e:
traceback.print_exc()
log.error(repr(e))
finally:
log.debug('forward_test finished')
def test_nn_class(model, x_test, y_test):
"""
NNを学習させる
Adamと呼ばれるパラメータの最適化手法を使用
@param model: NNの構造モデルオブジェクト
@param x_test: テストデータの特徴量
@param y_test: テストデータの教師信号
@return: 識別率、ロス、識別結果のリスト、ロスのリスト、識別結果の確信度のリスト
"""
# テストサンプル数
test_sample_size = len(x_test)
sum_loss = 0
sum_accuracy = 0
# 実際の出力クラスとその確信度
result_class_list = []
result_loss_list = []
result_class_power_list = []
for i in xrange(0, test_sample_size):
x_batch = x_test[i:i+1]
y_batch = y_test[i:i+1]
# 順伝播させて誤差と精度を算出
loss, acc, output= forward_data(model, x_batch, y_batch, train=False)
# 結果の格納
result_class_list.append(np.argmax(output.data))
result_loss_list.append(float(cuda.to_cpu(loss.data)))
result_class_power_list.append(output.data)
sum_loss += float(cuda.to_cpu(loss.data))
sum_accuracy += float(cuda.to_cpu(acc.data))
loss = sum_loss / test_sample_size
accuracy = sum_accuracy / test_sample_size
return accuracy, loss, result_class_list, result_loss_list, result_class_power_list
def train_nn_class(model,
x_train,
y_train,
batchsize,
epoch_num,
test_flag=False,
x_test=None,
y_test=None,
print_flag=False):
"""
NNを学習させる
Adamと呼ばれるパラメータの最適化手法を使用
@param model: NNの構造モデルオブジェクト
@param x_train: トレーニングデータの特徴量
@param y_train: トレーニングデータの教師信号
@param batchsize: 確率的勾配降下法で学習させる際の1回分のバッチサイズ
@param epoch_num: エポック数（1データセットの学習繰り返し数）
@keyword test_flag: テストデータの識別率を学習と同時並行して出力
"""
#
opts = optimizers.Adam()
opts.setup(model)
#optimizer = optimizers.MomentumSGD(lr=0.01, momentum=0.9)
if print_flag:
pace = 10 # 何epochごとに結果を出力するか
if test_flag:
pace = 10 # 何epochごとに結果を出力するか
test_sample_size = len(x_test)
#データ数
sample_num = len(x_train)
#学習ループ
# 識別率とロス
sum_accuracy = 0
sum_loss = 0
for epoch in xrange(1, epoch_num + 1):
# サンプルの順番をランダムに並び替える
perm = np.random.permutation(
sample_num) # permutationはshuffleと違い、新しいリストを生成する
# 1epochにおける識別率
sum_accuracy_on_epoch = 0
sum_loss_on_epoch = 0
# データをバッチサイズごとに使って学習
# 今回バッチサイズは1なので、サンプルサイズ数分学習
for i in xrange(0, sample_num, batchsize):
x_batch = x_train[perm[i:i + batchsize]]
y_batch = y_train[perm[i:i + batchsize]]
# 勾配(多分、偏微分分のgrad)を初期化
opts.zero_grads()
# 順伝播させて誤差と精度を算出
loss, acc, output = forward_data(model,
x_batch,
y_batch,
train=True)
# 誤差逆伝播で勾配を計算
loss.backward()
# 勾配(gradを使って本物のweight)を更新
opts.update()
# 識別率，ロスの算出
sum_loss_on_epoch += float(cuda.to_cpu(loss.data))
sum_accuracy_on_epoch += float(cuda.to_cpu(acc.data)) # 田村さんに確認
# 識別率とロスの積算
accuracy_on_epoch = sum_accuracy_on_epoch / (sample_num / batchsize)
loss_on_epoch = sum_loss_on_epoch / (sample_num / batchsize)
sum_accuracy += accuracy_on_epoch
sum_loss += loss_on_epoch
if print_flag:
if epoch % pace == 0:
print 'epoch', epoch
print 'train now: accuracy={}, loss={}'.format(
accuracy_on_epoch, loss_on_epoch)
#print 'train mean: loss={}, accuracy={}'.format(sum_loss / epoch, sum_accuracy / epoch)
# テストデータでの誤差と、正解精度を表示。汎化性能を確認。 #########################
if test_flag:
if epoch % pace == 0:
# テストデータでの誤差と、正解精度を表示
accuracy, loss, result_class_list, result_loss_list, result_class_power_list = test_nn_class(
model, x_test, y_test)
print 'test: accuracy={}, loss={}'.format(accuracy, loss)
# test_sum_accuracy = 0
# test_sum_loss = 0
#
# acc_t=np.zeros(((test_sample_size/batchsize), 5))
# for i in xrange(0, test_sample_size, batchsize):
# x_batch = x_test[i:i+batchsize]
# y_batch = y_test[i:i+batchsize]
#
# # 順伝播させて誤差と精度を算出
# loss, acc, output= forward_data(model, x_batch, y_batch, train=False)
# acc_t[i][0]=loss.data
# acc_t[i][1]=acc.data
# acc_t[i][2]=y_batch #教師信号
# acc_t[i][3]=np.max(output.data)#value of max
# acc_t[i][4]=np.argmax(output.data)# 実際の出力結果
#
# test_sum_loss += float(cuda.to_cpu(loss.data))
# test_sum_accuracy += float(cuda.to_cpu(acc.data))
#
# print 'test: accuracy={}, loss={}'.format(test_sum_accuracy / (test_sample_size / batchsize), test_sum_loss / (test_sample_size / batchsize))
return opts
def train_nn_class(model, x_train, y_train, batchsize, epoch_num, test_flag = False, x_test = None, y_test = None, print_flag = False):
"""
NNを学習させる
Adamと呼ばれるパラメータの最適化手法を使用
@param model: NNの構造モデルオブジェクト
@param x_train: トレーニングデータの特徴量
@param y_train: トレーニングデータの教師信号
@param batchsize: 確率的勾配降下法で学習させる際の1回分のバッチサイズ
@param epoch_num: エポック数（1データセットの学習繰り返し数）
@keyword test_flag: テストデータの識別率を学習と同時並行して出力
"""
#
opts = optimizers.Adam()
opts.setup(model)
#optimizer = optimizers.MomentumSGD(lr=0.01, momentum=0.9)
if print_flag:
pace = 10 # 何epochごとに結果を出力するか
if test_flag:
pace = 10 # 何epochごとに結果を出力するか
test_sample_size = len(x_test)
#データ数
sample_num = len(x_train)
#学習ループ
# 識別率とロス
sum_accuracy = 0
sum_loss = 0
for epoch in xrange(1, epoch_num+1):
# サンプルの順番をランダムに並び替える
perm = np.random.permutation(sample_num) # permutationはshuffleと違い、新しいリストを生成する
# 1epochにおける識別率
sum_accuracy_on_epoch = 0
sum_loss_on_epoch = 0
# データをバッチサイズごとに使って学習
# 今回バッチサイズは1なので、サンプルサイズ数分学習
for i in xrange(0, sample_num, batchsize):
x_batch = x_train[perm[i:i+batchsize]]
y_batch = y_train[perm[i:i+batchsize]]
# 勾配(多分、偏微分分のgrad)を初期化
opts.zero_grads()
# 順伝播させて誤差と精度を算出
loss, acc, output = forward_data(model, x_batch, y_batch, train=True)
# 誤差逆伝播で勾配を計算
loss.backward()
# 勾配(gradを使って本物のweight)を更新
opts.update()
# 識別率，ロスの算出
sum_loss_on_epoch += float(cuda.to_cpu(loss.data))
sum_accuracy_on_epoch += float(cuda.to_cpu(acc.data)) # 田村さんに確認
# 識別率とロスの積算
accuracy_on_epoch = sum_accuracy_on_epoch / (sample_num / batchsize)
loss_on_epoch = sum_loss_on_epoch / (sample_num / batchsize)
sum_accuracy += accuracy_on_epoch
sum_loss += loss_on_epoch
if print_flag:
if epoch % pace == 0:
print 'epoch', epoch
print 'train now: accuracy={}, loss={}'.format(accuracy_on_epoch, loss_on_epoch)
#print 'train mean: loss={}, accuracy={}'.format(sum_loss / epoch, sum_accuracy / epoch)
# テストデータでの誤差と、正解精度を表示。汎化性能を確認。 #########################
if test_flag:
if epoch % pace == 0:
# テストデータでの誤差と、正解精度を表示
accuracy, loss, result_class_list, result_loss_list, result_class_power_list = test_nn_class(model, x_test, y_test)
print 'test: accuracy={}, loss={}'.format(accuracy, loss)
# test_sum_accuracy = 0
# test_sum_loss = 0
#
# acc_t=np.zeros(((test_sample_size/batchsize), 5))
# for i in xrange(0, test_sample_size, batchsize):
# x_batch = x_test[i:i+batchsize]
# y_batch = y_test[i:i+batchsize]
#
# # 順伝播させて誤差と精度を算出
# loss, acc, output= forward_data(model, x_batch, y_batch, train=False)
# acc_t[i][0]=loss.data
# acc_t[i][1]=acc.data
# acc_t[i][2]=y_batch #教師信号
# acc_t[i][3]=np.max(output.data)#value of max
# acc_t[i][4]=np.argmax(output.data)# 実際の出力結果
#
# test_sum_loss += float(cuda.to_cpu(loss.data))
# test_sum_accuracy += float(cuda.to_cpu(acc.data))
#
# print 'test: accuracy={}, loss={}'.format(test_sum_accuracy / (test_sample_size / batchsize), test_sum_loss / (test_sample_size / batchsize))
return opts
def inversion():
f = forward.forward_gpu()
real_geo_model = forward.geoelectric_model()
real_geo_model.load_from_file('../test_data/test_geo_model.json')
coef = forward.filter_coefficient()
coef.load_cos_coef('../test_data/cos_xs.txt')
coef.load_hkl_coef('../test_data/hankel1.txt')
time = forward.forward_data()
time.generate_time_stamp_by_count(-5, -2, 20)
f.load_general_params(10, 100, 50)
f.load_filter_coef(coef)
f.load_geo_model(real_geo_model)
f.load_time_stamp(time)
f.forward()
real_response = f.get_result_magnetic()
real_response.name = real_geo_model.name
# helper.add_noise(m, 0.05)
real_response_m = real_response['response']
inversion_geo_model = forward.isometric_model()
inversion_geo_model.resize(20)
inversion_geo_model.height = 10
inversion_geo_model.name = 'inversion'
inv = Inversion(inversion_geo_model, f)
inv.set_initial_geo_model(npy.ones([inversion_geo_model.count]) * 50.0)
inv.set_target_magnetic(real_response_m)
for i in range(500):
log.info('iteration %d ' % i)
inv.rand_grad()
inv.update_model()
inv.update_step()
log.info('iteration %d, loss = %f' % (i, inv.loss()))
if i % 10 == 0:
inversion_geo_model.set_item_s('resistivity',
inv.geo_model.tolist())
inversion_geo_model.name = 'Inversion'
f.load_geo_model(forward.iso_to_geo(inversion_geo_model))
f.forward()
inversion_response = f.get_result_magnetic()
inversion_response.name = 'Inversion'
fig = helper.draw_resistivity(
real_geo_model,
forward.iso_to_geo(inversion_geo_model),
last_height=240)
fig.show()
fig = helper.draw_forward_result(real_response, inversion_response)
fig.show()
print('loss = %f' % helper.loss(real_response, inversion_response))
pass
pass
