一、梯度下降
梯度下降公式:$W^{[l]} = W^{[l]} - alpha ext{ } dW^{[l]}$,$b^{[l]} = b^{[l]} - alpha ext{ } db^{[l]}$,具体细节和代码实现参考文章神经网络介绍:算法基础
(Batch) Gradient Descent
### 伪代码 X = data_input Y = labels parameters = initialize_parameters(layers_dims) for i in range(0, num_epochs): # Forward propagation a, caches = forward_propagation(X, parameters) # Compute cost. cost = compute_cost(a, Y) # Backward propagation. grads = backward_propagation(a, caches, parameters) # Update parameters. parameters = update_parameters(parameters, grads)
Stochastic Gradient Descent
### 伪代码 X = data_input Y = labels parameters = initialize_parameters(layers_dims) for i in range(0, num_epochs): for j in range(0, m): # Forward propagation a, caches = forward_propagation(X[:,j], parameters) # Compute cost cost = compute_cost(a, Y[:,j]) # Backward propagation grads = backward_propagation(a, caches, parameters) # Update parameters. parameters = update_parameters(parameters, grads)
Mini-Batch Gradient Descent
Mini-Batch Gradient Descent介于(Batch) Gradient Descent和Stochastic Gradient Descent之间,可分为两步进行:
- Build mini-batches from the training set (X, Y)
def random_mini_batches(X, Y, mini_batch_size = 64, seed = 0): """ Creates a list of random minibatches from (X, Y) Arguments: X -- input data, of shape (number of features, number of examples) Y -- true label vector, of shape (1, number of examples) mini_batch_size -- size of the mini-batches, integer Returns: mini_batches -- list of synchronous (mini_batch_X, mini_batch_Y) """ np.random.seed(seed) m = X.shape[1] #number of training examples mini_batches = [] # Step 1: Shuffle (X, Y) permutation = list(np.random.permutation(m)) shuffled_X = X[:, permutation] shuffled_Y = Y[:, permutation].reshape((1,m)) # Step 2: Partition (shuffled_X, shuffled_Y). Minus the end case. num_complete_minibatches = math.floor(m/mini_batch_size) # number of mini batches of size mini_batch_size for k in range(0, num_complete_minibatches): mini_batch_X = shuffled_X[:,k*mini_batch_size:(k+1)*mini_batch_size] mini_batch_Y = shuffled_Y[0,k*mini_batch_size:(k+1)*mini_batch_size].reshape((1,mini_batch_size)) mini_batch = (mini_batch_X, mini_batch_Y) mini_batches.append(mini_batch) # Handling the end case (last mini-batch < mini_batch_size) if m % mini_batch_size != 0: mini_batch_X = shuffled_X[:,num_complete_minibatches*mini_batch_size:] mini_batch_Y = shuffled_Y[0,num_complete_minibatches*mini_batch_size:].reshape((1,m % mini_batch_size)) mini_batch = (mini_batch_X, mini_batch_Y) mini_batches.append(mini_batch) return mini_batches
- Train the network
### 伪代码 X = data_input Y = labels parameters = initialize_parameters(layers_dims) minibatches = random_mini_batches(X, Y) for i in range(0, num_epochs): for minibatch in minibatches: # Select a minibatch (minibatch_X, minibatch_Y) = minibatch # Forward propagation a, caches = forward_propagation(minibatch_X, parameters) # Compute cost cost = compute_cost(a, minibatch_Y) # Backward propagation grads = backward_propagation(a, caches, parameters) # Update parameters. parameters = update_parameters(parameters, grads)
梯度下降、随机梯度下降与mini-batch梯度下降的区别在于进行一次参数更新所使用的训练样本数量。梯度下降一次更新需使用全部样本,计算效率较低;随机梯度下降一次更新仅使用一个样本,计算效率较高,但在过程中会产生较剧烈的震荡,影响最终的优化结果;mini-batch梯度下降则介于梯度下降和随机梯度下降两者之间,若选取合适的mini_batch_size,mini-batch梯度下降常优于梯度下降和随机梯度下降,特别是当训练数据集很大的情况下。
为进一步减少训练过程中不必要的震荡(如上图所示),加快收敛速度,并且尽可能地避免落入局部最优值,有一些改进的算法可供选择:
Momentum
$$ egin{cases}
v_{dW^{[l]}} = eta v_{dW^{[l]}} + (1 - eta) dW^{[l]} \
W^{[l]} = W^{[l]} - alpha v_{dW^{[l]}}
end{cases} ext{ and }egin{cases}
v_{db^{[l]}} = eta v_{db^{[l]}} + (1 - eta) db^{[l]} \
b^{[l]} = b^{[l]} - alpha v_{db^{[l]}}
end{cases}$$
RMSProp
$$egin{cases}
s_{dW^{[l]}} = eta s_{dW^{[l]}} + (1 - eta) {dW^{[l]}}^2 \
W^{[l]} = W^{[l]} - alpha frac{dW^{[l]}}{sqrt{s_{dW^{[l]}}} + varepsilon}
end{cases} ext{ and }egin{cases}
s_{db^{[l]}} = eta s_{db^{[l]}} + (1 - eta) {db^{[l]}}^2 \
b^{[l]} = b^{[l]} - alpha frac{db^{[l]}}{sqrt{s_{db^{[l]}}} + varepsilon}
end{cases}$$
Adam
$$egin{cases}
v_{dW^{[l]}} = eta_1 v_{dW^{[l]}} + (1 - eta_1) dW^{[l]} \
v^{corrected}_{dW^{[l]}} = frac{v_{dW^{[l]}}}{1 - (eta_1)^t} \
s_{dW^{[l]}} = eta_2 s_{dW^{[l]}} + (1 - eta_2) {dW^{[l]}}^2 \
s^{corrected}_{dW^{[l]}} = frac{s_{dW^{[l]}}}{1 - (eta_2)^t} \
W^{[l]} = W^{[l]} - alpha frac{v^{corrected}_{dW^{[l]}}}{sqrt{s^{corrected}_{dW^{[l]}}} + varepsilon}
end{cases} ext{ and }egin{cases}
v_{db^{[l]}} = eta_1 v_{db^{[l]}} + (1 - eta_1) db^{[l]} \
v^{corrected}_{db^{[l]}} = frac{v_{db^{[l]}}}{1 - (eta_1)^t} \
s_{db^{[l]}} = eta_2 s_{db^{[l]}} + (1 - eta_2) {db^{[l]}}^2 \
s^{corrected}_{db^{[l]}} = frac{s_{db^{[l]}}}{1 - (eta_2)^t} \
b^{[l]} = b^{[l]} - alpha frac{v^{corrected}_{db^{[l]}}}{sqrt{s^{corrected}_{db^{[l]}}} + varepsilon}
end{cases}, ext{ }其中t表示迭代次数$$
以上算法中的学习率$alpha$可以是常数,也可以根据训练进度而变化,常用的几种方法有:
Learning Rate Decay
有多种形式,例如$frac{alpha_0}{1+decay\_rate*epoch\_number}$,$alpha_0$为初始学习率
Cyclic Learning Rates
有多种形式,例如下图所示
二、正则化
L2(Weight Decay)
以二分类问题为例,使用L2正则化后的损失函数为:
$$mathcal{J}=underbrace{-frac{1}{m} sumlimits_{i = 1}^{m} [y^{(i)}logleft(a^{[L] (i)} ight) + (1-y^{(i)})logleft(1- a^{[L](i)} ight)]}_ ext{cross-entropy cost} + underbrace{frac{lambda}{2} sumlimits_lsumlimits_ksumlimits_j W_{k,j}^{[l]2} }_ ext{L2 regularization cost}$$
Dropout
网络的训练过程和文章神经网络介绍:算法基础中的基本一致,不同的是Dropout在每次迭代过程中随机关掉一些神经元,
对矩阵$A^{[l]}$,有对应的开关矩阵$D^{[l]}$,生成$D^{[l]}$的代码如下所示:
### 伪代码 Dl = np.random.rand(Al.shape[0], Al.shape[1]) # Step 1: initialize matrix D1 = np.random.rand(..., ...) Dl = (Dl<keep_prob) # Step 2: convert entries of D1 to 0 or 1 (using keep_prob as the threshold)
(1)在前向传播过程中按上述文章中的公式计算出$A^{[l]}$后还需要额外进行处理:$$A^{[l]}=A^{[l]}*D^{[l]} ext{, }A^{[l]}=A^{[l]}/keep\_prob$$
(2)在后向传播过程中使用经过处理后的$A^{[l]}$进行计算,并且按文章中的公式计算出$dA^{[l]}$后还需要额外进行处理:$$dA^{[l]}=dA^{[l]}*D^{[l]} ext{, }dA^{[l]}=dA^{[l]}/keep\_prob$$
(3)需要注意的是在使用网络进行验证或预测时不要使用Dropout(即将keep_prob设为1)