Deep Learning 31: 不同版本的keras,对同样的代码,得到不同结果的原因总结

一.疑问

这几天一直纠结于一个问题:

同样的代码,为什么在keras的0.3.3版本中,拟合得比较好,也没有过拟合,验证集准确率一直高于训练准确率. 但是在换到keras的1.2.0版本中的时候,就过拟合了,验证误差一直高于训练误差

二.答案

今天终于发现原因了,原来是这两个版本的keras的optimezer实现不一样,但是它们的默认参数是一样的,因为我代码中用的是adam方法优化,下面就以optimezer中的adam来举例说明:

1.下面是keras==0.3.3时,其中optimezer.py中的adam方法实现:

 class Adam(Optimizer):
     '''Adam optimizer.

     Default parameters follow those provided in the original paper.

     # Arguments
         lr: float >= . Learning rate.
         beta_1/beta_2: floats,  < beta < . Generally close to .
         epsilon: float >= . Fuzz factor.

     # References
         - [Adam - A Method for Stochastic Optimization](http://arxiv.org/abs/1412.6980v8)
     '''
     def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-,
                  *args, **kwargs):
         super(Adam, self).__init__(**kwargs)
         self.__dict__.update(locals())
         self.iterations = K.variable()
         self.lr = K.variable(lr)
         self.beta_1 = K.variable(beta_1)
         self.beta_2 = K.variable(beta_2)

     def get_updates(self, params, constraints, loss):
         grads = self.get_gradients(loss, params)
         self.updates = [(self.iterations, self.iterations+.)]

         t = self.iterations +
         lr_t = self.lr * K.sqrt( - K.pow(self.beta_2, t)) / ( - K.pow(self.beta_1, t))

         for p, g, c in zip(params, grads, constraints):
             # zero init of moment
             m = K.variable(np.zeros(K.get_value(p).shape))
             # zero init of velocity
             v = K.variable(np.zeros(K.get_value(p).shape))

             m_t = (self.beta_1 * m) + ( - self.beta_1) * g
             v_t = (self.beta_2 * v) + ( - self.beta_2) * K.square(g)
             p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

             self.updates.append((m, m_t))
             self.updates.append((v, v_t))
             self.updates.append((p, c(p_t)))  # apply constraints
         return self.updates

     def get_config(self):
         return {"name": self.__class__.__name__,
                 "lr": float(K.get_value(self.lr)),
                 "beta_1": float(K.get_value(self.beta_1)),
                 "beta_2": float(K.get_value(self.beta_2)),
                 "epsilon": self.epsilon}

2.下面是keras==1.2.0时,其中optimezer.py中的adam方法实现:

 class Adam(Optimizer):
     '''Adam optimizer.

     Default parameters follow those provided in the original paper.

     # Arguments
         lr: float >= . Learning rate.
         beta_1/beta_2: floats,  < beta < . Generally close to .
         epsilon: float >= . Fuzz factor.

     # References
         - [Adam - A Method for Stochastic Optimization](http://arxiv.org/abs/1412.6980v8)
     '''
     def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                  epsilon=1e-, decay=., **kwargs):
         super(Adam, self).__init__(**kwargs)
         self.__dict__.update(locals())
         self.iterations = K.variable()
         self.lr = K.variable(lr)
         self.beta_1 = K.variable(beta_1)
         self.beta_2 = K.variable(beta_2)
         self.decay = K.variable(decay)
         self.inital_decay = decay

     def get_updates(self, params, constraints, loss):
         grads = self.get_gradients(loss, params)
         self.updates = [K.update_add(self.iterations, )]

         lr = self.lr
         if self.inital_decay > :
             lr *= (. / (. + self.decay * self.iterations))

         t = self.iterations +
         lr_t = lr * K.sqrt(. - K.pow(self.beta_2, t)) / (. - K.pow(self.beta_1, t))

         shapes = [K.get_variable_shape(p) for p in params]
         ms = [K.zeros(shape) for shape in shapes]
         vs = [K.zeros(shape) for shape in shapes]
         self.weights = [self.iterations] + ms + vs

         for p, g, m, v in zip(params, grads, ms, vs):
             m_t = (self.beta_1 * m) + (. - self.beta_1) * g
             v_t = (self.beta_2 * v) + (. - self.beta_2) * K.square(g)
             p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

             self.updates.append(K.update(m, m_t))
             self.updates.append(K.update(v, v_t))

             new_p = p_t
             # apply constraints
             if p in constraints:
                 c = constraints[p]
                 new_p = c(new_p)
             self.updates.append(K.update(p, new_p))
         return self.updates

     def get_config(self):
         config = {'lr': float(K.get_value(self.lr)),
                   'beta_1': float(K.get_value(self.beta_1)),
                   'beta_2': float(K.get_value(self.beta_2)),
                   'decay': float(K.get_value(self.decay)),
                   'epsilon': self.epsilon}
         base_config = super(Adam, self).get_config()
         return dict(list(base_config.items()) + list(config.items()))

读代码对比,可发现这两者实现方式有不同,而我的代码中一直使用的是adam的默认参数,所以才会结果不一样.

三.解决

要避免这一问题可用以下方法:

1.在自己的代码中,要对优化器的参数给定,不要用默认参数.

adam = optimizers.Adam(lr=1e-)

但是,在keras官方文档中,明确有说明,在用这些优化器的时候,最好使用默认参数,所以也可采用第2种方法.

2.优化函数中的优化方法要给定,也就是在训练的时候,在fit函数中的callbacks参数中的schedule要给定.

比如:

 # Callback that implements learning rate schedule
 schedule = Step([], [1e-, 1e-])

 history = model.fit(X_train, Y_train,
                     batch_size=batch_size, nb_epoch=nb_epoch, validation_data=(X_test,Y_test),
                     callbacks=[
                         schedule,
                         keras.callbacks.ModelCheckpoint(filepath, monitor='val_loss', verbose=,save_best_only=True, mode='auto')# 该回调函数将在每个epoch后保存模型到filepath
                         # ,keras.callbacks.EarlyStopping(monitor='val_loss', patience=, verbose=, mode='auto')# 当监测值不再改善时，该回调函数将中止训练.当early stop被激活（如发现loss相比上一个epoch训练没有下降），则经过patience个epoch后停止训练
                         ],
                     verbose=, shuffle=True)

其中Step函数如下:

 class Step(Callback):

     def __init__(self, steps, learning_rates, verbose=):
         self.steps = steps
         self.lr = learning_rates
         self.verbose = verbose

     def change_lr(self, new_lr):
         old_lr = K.get_value(self.model.optimizer.lr)
         K.set_value(self.model.optimizer.lr, new_lr)
         if self.verbose == :
             print('Learning rate is %g' %new_lr)

     def on_epoch_begin(self, epoch, logs={}):
         for i, step in enumerate(self.steps):
             if epoch < step:
                 self.change_lr(self.lr[i])
                 return
         self.change_lr(self.lr[i+])

     def get_config(self):
         config = {'class': type(self).__name__,
                   'steps': self.steps,
                   'learning_rates': self.lr,
                   'verbose': self.verbose}
         return config

     @classmethod
     def from_config(cls, config):
         offset = config.get('epoch_offset', )
         steps = [step - offset for step in config['steps']]
         return cls(steps, config['learning_rates'],
                    verbose=config.get('verbose', ))