def train(model, epoch=1000, learning_rate=1e-3, steps_per_epoch=steps_per_epoch):
# 开始训练
losses = []
loss_min = np.inf
graph_sort = toplogical_sort(model.feed_dict) # 拓扑排序
optim = Nadam(graph_sort)
update_lr = Auto_update_lr(lr=learning_rate, alpha=0.1, patiences=500, print_=True)
for e in range(epoch):
loss = 0
for b in range(steps_per_epoch):
X_batch, y_batch = resample(x_, y_, n_samples=batch_size)
# print(x_.shape,X_batch.shape)
mlp.x.value = X_batch # 在这更新值
mlp.y.value = y_batch[:, None]
# print(X_batch.shape)
run_steps(graph_sort, monitor=False)
optim.update(learning_rate=learning_rate)
Visual_gradient(model)
Grad_Clipping_Disappearance(model, 5)
loss += mlp.MSE_loss.value
update_lr.updata(loss/steps_per_epoch)
print("epoch:{}/{},loss:{:.6f}".format(e,epoch,loss / steps_per_epoch))
losses.append(loss / steps_per_epoch)
if loss / steps_per_epoch < loss_min:
print('loss is {:.6f}, is decreasing!! save moddel'.format(loss / steps_per_epoch))
save_model("model/mlp.xhp", model)
loss_min = loss / steps_per_epoch
print('The min loss:',loss_min)
# print("loss:{}".format(np.mean(losses)))
plt.plot(losses)
plt.savefig("image/many_vectoy.png")
plt.show()
def train(model, train_data, epoch=6, learning_rate=1e-3):
# 开始训练
losses = []
loss_min = np.inf
graph_sort_lstm = toplogical_sort(model.feed_dict) # 拓扑排序
optim = Adam(graph_sort_lstm)
update_lr = Auto_update_lr(lr=learning_rate,
alpha=0.1,
patiences=20,
print_=True)
for e in range(epoch):
for X, Y in train_data:
X, Y = X.numpy(), Y.numpy()
model.x.value = X
model.y.value = Y
run_steps(graph_sort_lstm)
# if model.y_pre.value is not None:
# print(model.y_pre.value.shape,Y.shape)
learning_rate = update_lr.lr
optim.update(learning_rate=learning_rate)
Visual_gradient(model)
Grad_Clipping_Disappearance(model, 5)
loss = model.MSE_loss.value
losses.append(loss)
update_lr.updata(np.mean(np.mean(losses)))
print("epoch:{}/{},loss:{:.6f}".format(e, epoch, np.mean(losses)))
if np.mean(losses) < loss_min:
print('loss is {:.6f}, is decreasing!! save moddel'.format(
np.mean(losses)))
save_model("model/lstm.xhp", model)
loss_min = np.mean(losses)
print('min loss:', loss_min)
plt.plot(losses)
plt.savefig("image/lstm_loss.png")
plt.show()
def predict(x, model):
graph = toplogical_sort(model.feed_dict)
model.x.value = x
run_steps(graph, train=False, valid=False)
y = graph[-2].value
result = np.argmax(y, axis=1)
return result
def train(model, train_data, epoch=4000, learning_rate=0.0128):
#开始训练
accuracies = []
losses = []
losses_valid = []
accuracies_valid = []
loss_min = np.inf
graph_sort_class = toplogical_sort(model.feed_dict) # 拓扑排序
optim = Adam(graph_sort_class)
update_lr = Auto_update_lr(lr=learning_rate,
alpha=0.1,
patiences=200,
print_=True)
for e in range(epoch):
for X, Y in train_data:
X, Y = X.unsqueeze(1).numpy(), Y.numpy()
model.x.value = X
model.y.value = Y
run_steps(graph_sort_class)
learning_rate = update_lr.lr
optim.update(learning_rate=learning_rate)
Visual_gradient(model)
Grad_Clipping_Disappearance(model, 5)
loss = model.cross_loss.value
accuracy = model.cross_loss.accuracy
losses.append(loss)
accuracies.append(accuracy * 100)
for x, y in valid_loader:
x, y = x.unsqueeze(1).numpy(), y.numpy()
model.x.value = x
model.y.value = y
run_steps(graph_sort_class, train=False, valid=True)
loss_valid = model.cross_loss.value
accuracy_valid = model.cross_loss.accuracy
losses_valid.append(loss_valid)
accuracies_valid.append(accuracy_valid * 100)
update_lr.updata(np.mean(losses_valid))
print(
"epoch:{}/{},train loss:{:.8f},train accuracy:{:.6f}%,valid loss:{:.8f},valid accuracy:{:.6f}%"
.format(e, epoch, np.mean(losses), np.mean(accuracies),
np.mean(losses_valid), np.mean(accuracies_valid)))
if np.mean(losses_valid) < loss_min:
print('loss is {:.6f}, is decreasing!! save moddel'.format(
np.mean(losses_valid)))
save_model("model/lstm_class.xhp", model)
loss_min = np.mean(losses_valid)
#save_model("lstm_class.xhp",model)
plt.plot(losses)
plt.savefig("image/lstm_class_loss.png")
plt.show()
def evaluator(test_loader,model):
graph = toplogical_sort(model.feed_dict)
accuracies = []
losses = []
for x, y in test_loader:
x, y = x.numpy(), y.numpy()
model.x.value = x
model.y.value = y
run_steps(graph, train=False, valid=True)
loss_test = model.cross_loss.value
accuracy_test = model.cross_loss.accuracy
losses.append(loss_test)
accuracies.append(accuracy_test)
print("test loss:{},test accuracy:{}".format(np.mean(losses),np.mean(accuracies)))
Visual_gradient(model)
Grad_Clipping_Disappearance(model, 5)
loss += mlp.MSE_loss.value
update_lr.updata(loss/steps_per_epoch)
print("epoch:{}/{},loss:{:.6f}".format(e,epoch,loss / steps_per_epoch))
losses.append(loss / steps_per_epoch)
if loss / steps_per_epoch < loss_min:
print('loss is {:.6f}, is decreasing!! save moddel'.format(loss / steps_per_epoch))
save_model("model/mlp.xhp", model)
loss_min = loss / steps_per_epoch
print('The min loss:',loss_min)
# print("loss:{}".format(np.mean(losses)))
plt.plot(losses)
plt.savefig("image/many_vectoy.png")
plt.show()
train(mlp)
load_model("model/mlp.xhp",mlp)
def predict(x_rm, graph,model):
model.x.value = x_rm
run_steps(graph, monitor=False, train=False,valid=False)
return model.y_pre.value*std_y + mean_y
graph_sort = toplogical_sort(mlp.feed_dict)
print("预测值：",predict(x_[17:50],graph_sort,mlp),"真实值：",y_[17:50]*std_y + mean_y)