class LinearModel(Model):
"""docstring for ClassName"""
def __init__(self,
n_features,
activation=Tanh(),
good_enough_loss=0.045):
Model.__init__(self, good_enough_loss)
self.layer = Layer(n_features, 1, activation=activation)
def init_optimizers_for_layers(self, optimizer):
self.layer.compile(optimizer)
def predict(self, x):
A = self.layer.forward(x)
return A
def step(self, x_batch, y_batch):
A = self.layer.forward(x_batch)
dA = self.loss_func.derivative(y_batch, A)
dA_prev, dW, db = self.layer.backward(dA)
self.layer.update_params(dW, db)
class Network(object):
def __init__(self, input_size, hidden_units, n_class=1):
self.layers = []
self.input_size = input_size
for siz in hidden_units:
hidden_layer = Layer(input_size,
siz,
activation=relu,
derivation=relu_derivation)
self.layers.append(hidden_layer)
input_size = siz
if n_class == 1:
self.output_layer = Layer(input_size,
1,
activation=sigmoid,
derivation=None)
else:
self.output_layer = Layer(input_size,
n_class,
activation=softmax,
derivation=None)
def predict(self, all_x):
preds = []
for j in range(all_x.shape[0]):
x = all_x[j]
pred = self.predict_one_sample(x.reshape(-1, 1))
preds.append(pred)
return np.array(preds).reshape(-1, 10)
def predict_one_sample(self, x):
output = x
for layer in self.layers:
output = layer.forward(output)
output = self.output_layer.forward(output)
return output
def train_one_sample(self, y, pred, learning_rate):
# 交叉熵损失函数
delta = y - pred
self.output_layer.backward(None, delta)
w = self.output_layer.w
for layer in reversed(self.layers):
delta = layer.backward(w, delta)
w = layer.w
self.output_layer.update(learning_rate)
for layer in self.layers:
layer.update(learning_rate)
def train(self, train_x, train_y, epoch, learning_rate):
print("start train ...")
for i in range(epoch):
for j in range(train_x.shape[0]):
x, y = train_x[j], train_y[j]
pred = self.predict_one_sample(x.reshape(-1, 1))
self.train_one_sample(y.reshape(-1, 1), pred, learning_rate)
print("train done")
def loss(self, y_label, y_pred):
ls = y_label * (-np.log(y_pred))
return np.sum(ls)