def load_weights(self, filepath):
"""加载模型参数.
Arguments:
filepath: str, 权重文件加载的路径.
Raises:
KeyError: 模型权重加载失败.
Notes:
模型将不会加载关于优化器的超参数.
"""
# 初始化权重文件.
parameters_gp = io.initialize_weights_file(
filepath=filepath, mode='r', model_name='LogisticRegression')
# 加载模型参数.
try:
compile_ds = parameters_gp['compile']
weights_ds = parameters_gp['weights']
self.optimizer = get_optimizer(compile_ds.attrs['optimizer'])
self.loss = get_loss(compile_ds.attrs['loss'])
self.metric = get_metric(compile_ds.attrs['metric'])
self.beta = weights_ds.attrs['beta']
# 标记加载完成
self.is_loaded = True
except KeyError:
CLASSICML_LOGGER.error('模型权重加载失败, 请检查文件是否损坏')
raise KeyError('模型权重加载失败')
def compile(self,
hidden_units,
optimizer='rbf',
loss='mse',
metric='accuracy'):
"""编译径向基函数网络, 配置训练时使用的超参数.
Arguments:
hidden_units: int, 径向基函数网络的隐含层神经元数量.
optimizer: str, classicML.optimizers.Optimizer 实例,
径向基函数网络使用的优化器.
loss: str, classicML.losses.Loss 实例, default='mse'
径向基函数网络使用的损失函数.
metric: str, classicML.metrics.Metric 实例, default='accuracy'
径向基函数网络使用的评估函数.
Notes:
- 注意RBF只能使用RadialBasisFunctionOptimizer优化器,
之所以开放优化器选项, 只是为了满足用户修改学习率的需求.
- 使用交叉熵作为损失函数有潜在异常的可能性,
除隐含层神经元个数和学习率之外, 建议使用默认参数.
"""
self.hidden_units = hidden_units
self.initializer = get_initializer('rbf_normal', self.seed)
self.optimizer = get_optimizer(optimizer)
self.loss = get_loss(loss)
self.metric = get_metric(metric)
def load_weights(self, filepath):
"""加载模型参数.
Arguments:
filepath: str, 权重文件加载的路径.
Raises:
KeyError: 模型权重加载失败.
Notes:
模型将不会加载关于优化器的超参数.
"""
# 初始化权重文件.
parameters_gp = io.initialize_weights_file(
filepath=filepath,
mode='r',
model_name='RadialBasisFunctionNetwork')
# 加载模型参数.
try:
compile_ds = parameters_gp['compile']
weights_ds = parameters_gp['weights']
self.hidden_units = compile_ds.attrs['hidden_units']
self.optimizer = get_optimizer(compile_ds.attrs['optimizer'])
self.loss = get_loss(compile_ds.attrs['loss'])
self.metric = get_metric(compile_ds.attrs['metric'])
for attr in weights_ds.attrs:
self.parameters[attr] = weights_ds.attrs[attr]
# 标记加载完成
self.is_loaded = True
except KeyError:
CLASSICML_LOGGER.error('模型权重加载失败, 请检查文件是否损坏')
raise KeyError('模型权重加载失败')
def compile(self,
network_structure,
optimizer='sgd',
loss='crossentropy',
metric='accuracy'):
"""编译神经网络, 配置训练时使用的超参数.
Arguments:
network_structure: list,
神经网络的结构, 定义神经网络的隐含层和输出层的神经元个数(输入层目前将自动推理);
例如: [3, 1]是一个隐含层3个神经元和输出层1个神经元的网络,
[5, 5, 2]是一个有两个隐含层每层有5个神经元和输出层2个神经元的网络,
optimizer: str, classicML.optimizers.Optimizer 实例, default='sgd'
神经网络使用的优化器.
loss: str, classicML.losses.Loss 实例, default='crossentropy'
神经网络使用的损失函数.
metric: str, classicML.metrics.Metric 实例, default='accuracy'
神经网络使用的评估函数.
"""
self.network_structure = network_structure
self.initializer = get_initializer(self.initializer, self.seed)
self.optimizer = get_optimizer(optimizer)
self.loss = get_loss(loss)
self.metric = get_metric(metric)
def compile(self,
optimizer='newton',
loss='log_likelihood',
metric='accuracy'):
"""编译模型, 配置训练时使用的超参数.
Arguments:
optimizer: str, classicML.optimizers.Optimizer 实例, default='newton'
模型使用的优化器.
loss: str, classicML.losses.Loss 实例, default='log_likelihood'
模型使用的损失函数.
metric: str, classicML.metrics.Metric 实例, default='accuracy'
模型使用的评估函数.
"""
self.initializer = get_initializer(self.initializer, self.seed)
self.optimizer = get_optimizer(optimizer)
self.loss = get_loss(loss)
self.metric = get_metric(metric)
def __init__(self, seed=None):
"""初始化分类器.
Arguments:
seed: int, default=None,
随机种子.
"""
super(SupportVectorClassifier, self).__init__()
self.seed = np.random.seed(seed)
self.support = None
self.support_vector = None
self.support_alpha = None
self.support_y = None
self.b = 0
self.C = None
self.kernel = None
self.gamma = None
self.tol = None
self.epochs = None
self.optimizer = get_optimizer('SMO')
self.is_trained = False
self.is_loaded = False