def __init__(self, hp, model_name):
self.hp = hp
self.model_name = model_name
self.subfolder = os.getcwd() + "\\" + self.__create_subfolder()
print(self.subfolder)
self.wb1 = WeightsBias(self.hp.num_input, self.hp.num_hidden, self.hp.init_method, self.hp.eta)
self.wb1.InitializeWeights(self.subfolder, False)
self.wb2 = WeightsBias(self.hp.num_hidden, self.hp.num_output, self.hp.init_method, self.hp.eta)
self.wb2.InitializeWeights(self.subfolder, False)
class NeuralNet2(object):
def __init__(self, hp, model_name):
self.hp = hp
self.model_name = model_name
self.subfolder = os.getcwd() + "\\" + self.__create_subfolder()
print(self.subfolder)
self.wb1 = WeightsBias(self.hp.num_input, self.hp.num_hidden,
self.hp.init_method, self.hp.eta)
self.wb1.InitializeWeights(self.subfolder, False)
self.wb2 = WeightsBias(self.hp.num_hidden, self.hp.num_output,
self.hp.init_method, self.hp.eta)
self.wb2.InitializeWeights(self.subfolder, False)
def __create_subfolder(self):
if self.model_name != None:
path = self.model_name.strip()
path = path.rstrip("\\")
isExists = os.path.exists(path)
if not isExists:
os.makedirs(path)
return path
def forward(self, batch_x):
# layer 1
self.Z1 = np.dot(batch_x, self.wb1.W) + self.wb1.B
self.A1 = Sigmoid().forward(self.Z1)
# layer 2
self.Z2 = np.dot(self.A1, self.wb2.W) + self.wb2.B
if self.hp.net_type == NetType.BinaryClassifier:
self.A2 = Logistic().forward(self.Z2)
elif self.hp.net_type == NetType.MultipleClassifier:
self.A2 = Softmax().forward(self.Z2)
else: # NetType.Fitting
self.A2 = self.Z2
#end if
self.output = self.A2
def backward(self, batch_x, batch_y, batch_a):
# 批量下降，需要除以样本数量，否则会造成梯度爆炸
m = batch_x.shape[0]
# 第二层的梯度输入 公式5
dZ2 = self.A2 - batch_y
# 第二层的权重和偏移 公式6
self.wb2.dW = np.dot(self.A1.T, dZ2) / m
# 公式7 对于多样本计算，需要在横轴上做sum，得到平均值
self.wb2.dB = np.sum(dZ2, axis=0, keepdims=True) / m
# 第一层的梯度输入 公式8
d1 = np.dot(dZ2, self.wb2.W.T)
# 第一层的dZ 公式10
dZ1, _ = Sigmoid().backward(None, self.A1, d1)
# 第一层的权重和偏移 公式11
self.wb1.dW = np.dot(batch_x.T, dZ1) / m
# 公式12 对于多样本计算，需要在横轴上做sum，得到平均值
self.wb1.dB = np.sum(dZ1, axis=0, keepdims=True) / m
def update(self):
self.wb1.Update()
self.wb2.Update()
def inference(self, x):
self.forward(x)
return self.output
def train(self, dataReader, checkpoint, need_test):
# calculate loss to decide the stop condition
self.loss_trace = TrainingTrace()
self.loss_func = LossFunction(self.hp.net_type)
loss = 10
if self.hp.batch_size == -1:
self.hp.batch_size = dataReader.num_train
max_iteration = math.ceil(dataReader.num_train / self.hp.batch_size)
checkpoint_iteration = (int)(max_iteration * checkpoint)
need_stop = False
for epoch in range(self.hp.max_epoch):
#print("epoch=%d" %epoch)
dataReader.Shuffle()
for iteration in range(max_iteration):
# get x and y value for one sample
batch_x, batch_y = dataReader.GetBatchTrainSamples(
self.hp.batch_size, iteration)
# get z from x,y
batch_a = self.forward(batch_x)
# calculate gradient of w and b
self.backward(batch_x, batch_y, batch_a)
# update w,b
self.update()
total_iteration = epoch * max_iteration + iteration
if (total_iteration + 1) % checkpoint_iteration == 0:
need_stop = self.CheckErrorAndLoss(dataReader, batch_x,
batch_y, epoch,
total_iteration)
if need_stop:
break
#end if
#end if
# end for
if need_stop:
break
# end for
self.SaveResult()
#self.CheckErrorAndLoss(dataReader, batch_x, batch_y, epoch, total_iteration)
if need_test:
print("testing...")
accuracy = self.Test(dataReader)
print(accuracy)
# end if
def CheckErrorAndLoss(self, dataReader, train_x, train_y, epoch,
total_iteration):
print("epoch=%d, total_iteration=%d" % (epoch, total_iteration))
# calculate train loss
self.forward(train_x)
loss_train = self.loss_func.CheckLoss(self.output, train_y)
accuracy_train = self.__CalAccuracy(self.output, train_y)
print("loss_train=%.6f, accuracy_train=%f" %
(loss_train, accuracy_train))
# calculate validation loss
vld_x, vld_y = dataReader.GetValidationSet()
self.forward(vld_x)
loss_vld = self.loss_func.CheckLoss(self.output, vld_y)
accuracy_vld = self.__CalAccuracy(self.output, vld_y)
print("loss_valid=%.6f, accuracy_valid=%f" % (loss_vld, accuracy_vld))
need_stop = self.loss_trace.Add(epoch, total_iteration, loss_train,
accuracy_train, loss_vld, accuracy_vld)
if loss_vld <= self.hp.eps:
need_stop = True
return need_stop
def Test(self, dataReader):
x, y = dataReader.GetTestSet()
self.forward(x)
correct = self.__CalAccuracy(self.output, y)
print(correct)
def __CalAccuracy(self, a, y):
assert (a.shape == y.shape)
m = a.shape[0]
if self.hp.net_type == NetType.Fitting:
var = np.var(y)
mse = np.sum((a - y)**2) / m
r2 = 1 - mse / var
return r2
elif self.hp.net_type == NetType.BinaryClassifier:
b = np.round(a)
r = (b == y)
correct = r.sum()
return correct / m
elif self.hp.net_type == NetType.MultipleClassifier:
ra = np.argmax(a, axis=1)
ry = np.argmax(y, axis=1)
r = (ra == ry)
correct = r.sum()
def SaveResult(self):
self.wb1.SaveResultValue(self.subfolder, "wb1")
self.wb2.SaveResultValue(self.subfolder, "wb2")
def LoadResult(self):
self.wb1.LoadResultValue(self.subfolder, "wb1")
self.wb2.LoadResultValue(self.subfolder, "wb2")
def ShowTrainingTrace(self):
self.loss_trace.ShowLossHistory(self.hp)
def GetTrainingTrace(self):
return self.loss_trace