def __init__(self):
"""
hidden_layer_num: 隐藏层数目
hidden_size: 隐藏层大小
classes_num: 分类数目
"""
super(MLP, self).__init__()
self.LossFunc = FocalLoss()
self.layers = [
Linear(256, 256),
Batchnorm(256),
ReLU(),
Dropout(0.5),
Linear(256, 128),
Batchnorm(128),
ReLU(),
Dropout(0.5),
Linear(128, 128),
Batchnorm(128),
ReLU(),
Dropout(0.5),
Linear(128, 4),
Sigmoid(), self.LossFunc
]
self.parameters = {}
for l in self.layers:
if hasattr(l, 'parameters'):
self.parameters[id(l)] = l.parameters
def __init__(self):
super().__init__(CrossEntropy, learning_rate=0.05)
self.a1 = ReLu(self)
self.a2 = SoftMax(self)
self.l1 = Linear(self, 784, 256)
self.l2 = Linear(self, 256, 120)
self.l3 = Linear(self, 120, 64)
self.l4 = Linear(self, 64, 10)
def layer(name, dim1, dim2, ind):
if "input" == name:
assert False, "input layer must be add-hot pluged to neural net at forward pass per new input!"
if "lin" == name:
return Linear(dim1, dim2, ind)
if "sig" == name:
return Sigmoid(dim1, dim2, ind)
if "relu" == name:
return ReLU(dim1, dim2, ind)
if "ass" == name:
return AntiSaturationSig(dim1, ind)
assert False, "space-layer : <{}> not implemented!".format(name)
return None
def __init__(self):
super().__init__(MSE, learning_rate=0.02)
self.a1 = Passive(self)
self.l1 = Linear(self, 1, 5)
self.l2 = Linear(self, 5, 1)