numpy创建神经网络框架

	# 定义网络 define neural network
	class DigitModule(Module):
	    def __init__(self):
	        # 计算顺序就会按照这里定义的顺序进行
	        sequential = Sequential([
	            layers.Linear2(in_dim=ROW_NUM * COLUM_NUM, out_dim=16, coe=2),
	            layers.Relu(16),
	            layers.Linear2(in_dim=16, out_dim=16, coe=2),
	            layers.Relu(16),
	            layers.Linear2(in_dim=16, out_dim=CLASS_NUM, coe=1),
	            layers.Sigmoid(CLASS_NUM)
	        ])
	        super(DigitModule, self).__init__(sequential)
	
	
	module = DigitModule()  # 创建模型 create module
	loss_func = SquareLoss(backward_func=module.backward)  # 定义损失函数 define loss function
	optimizer = SGD(module, lr=learning_rate)  # 定义优化器 define optimizer
	
	
	for i in range(EPOCH_NUM):  # 共训练EPOCH_NUM轮
	    trainning_loss = 0  # 计算一下当前一轮训练的loss值，可以没有
	    for data in train_loader:  # 遍历所有样本，train_loader是可迭代对象，保存了数据集中所有的数据
	        imgs, targets = data  # 将数据拆分成图片和标签
	        outputs = module(imgs)  # 将样本的输入值输入到自己的神经网络中
	        loss = loss_func(outputs, targets, transform=True)  # 计算loss / calculate loss
	        trainning_loss += loss.value
	        loss.backward()  # 通过反向传播计算梯度 / calculate gradiant through back propagation
	        optimizer.step()  # 通过优化器调整模型参数 / adjust the weights of network through optimizer
	    if i % TEST_STEP == 0:  # 每训练TEST_STEP轮就测试一下当前训练的成果
	        show_effect(i, module, loss_func, test_loader, i // TEST_STEP)
	        print("{} turn finished, loss of train set = {}".format(i, trainning_loss))

	def forward(self, x):
	    out = x
	    for layer in self.layers:
	        out = layer(out)
	    return out
	
	def backward(self, output_gradiant):
	    layer_num = len(self.layers)
	    delta = output_gradiant
	    for i in range(layer_num - 1, -1, -1):
	        # 反向遍历各个层, 将期望改变量反向传播
	        delta = self.layers[i].backward(delta)
	
	def step(self, lr):
	    for layer in self.layers:
	        layer.step(lr)

    def forward(self, inputs):
	        """
	        :param inputs: input = (h0, x) h0.shape == (batch, out_dim) x.shape == (seq, batch, in_dim)
	        :return: outputs: outputs.shape == (seq, batch, out_dim)
	        """
	        h = inputs[0]  # 输入的inputs由两部分组成
	        X = inputs[1]
	        if X.shape[2] != self.in_dim or h.shape[1] != self.out_dim:
	            # 检查输入的形状是否有问题
	            raise ShapeNotMatchException(self, "forward: wrong shape: h0 = {}, X = {}".format(h.shape, X.shape))
	
	        self.seq_len = X.shape[0]  # 时间序列的长度
	        self.inputs = X  # 保存输入，之后的反向传播还要用
	        output_list = []  # 保存每个时间点的输出
	        for x in X:
	            # 按时间序列遍历input
	            # x.shape == (batch, in_dim), h.shape == (batch, out_dim)
	            h = self.activation(self.linear(np.c_[h, x]))
	            output_list.append(h)
	        self.outputs = np.stack(output_list, axis=0)  # 将列表转换成一个矩阵保存起来
	        return self.outputs

	def backward(self, output_gradiant):
	    """
	    :param output_gradiant: shape == (seq, batch, out_dim)
	    :return: input_gradiant
	    """
	    if output_gradiant.shape != self.outputs.shape:
	        # 期望得到(seq, batch, out_dim)形状
	        raise ShapeNotMatchException(self, "__backward: expected {}, but we got "
	                                           "{}".format(self.outputs.shape, output_gradiant.shape))
	
	    input_gradients = []
	    # 每个time_step上的虚拟weight_gradient, 最后求平均值就是总的weight_gradient
	    weight_gradients = np.zeros(self.linear.weights_shape())
	    bias_gradients = np.zeros(self.linear.bias_shape())
	    batch_size = output_gradiant.shape[1]
	
	    # total_gradient: 前向传播的时候是将x, h合成为一个矩阵，所以反向传播也先计算这个大矩阵的梯度再拆分为x_grad, h_grad
	    total_gradient = np.zeros((batch_size, self.out_dim + self.in_dim))
	    h_gradient = None
	    
	    # 反向遍历各个时间层，计算该层的梯度值
	    for i in range(self.seq_len - 1, -1, -1):
	        # 前向传播顺序: x, h -> z -> h
	        # 所以反向传播计算顺序：h_grad -> z_grad -> x_grad, h_grad, w_grad, b_grad
	
	        # %%%%%%%%%%%%%%计算平均值的版本%%%%%%%%%%%%%%%%%%%%%%%
	        # h_gradient = (output_gradiant[i] + total_gradient[:, 0:self.out_dim]) / 2
	        # %%%%%%%%%%%%%%不计算平均值的版本%%%%%%%%%%%%%%%%%%%%%%%
	        #  计算h_grad: 这一时间点的h_grad包括输出的grad和之前的时间点计算所得grad两部分
	        h_gradient = output_gradiant[i] + total_gradient[:, 0:self.out_dim]  
	
	        # w_grad和b_grad是在linear.backward()内计算的，不用手动再计算了
	        z_gradient = self.activation.backward(h_gradient)  # 计算z_grad
	        total_gradient = self.linear.backward(z_gradient)  # 计算x_grad和h_grad合成的大矩阵的梯度
	
	        # total_gradient 同时包含了h和x的gradient, shape == (batch, out_dim + in_dim)
	        x_gradient = total_gradient[:, self.out_dim:]
	
	        input_gradients.append(x_gradient)  
	        weight_gradients += self.linear.gradients["w"]
	        bias_gradients += self.linear.gradients["b"]
	
	    # %%%%%%%%%%%%%%%%%%计算平均值的版本%%%%%%%%%%%%%%%%%%%%%%%
	    # self.linear.set_gradients(w=weight_gradients / self.seq_len, b=bias_gradients / self.seq_len)
	    # %%%%%%%%%%%%%%%%%%不计算平均值的版本%%%%%%%%%%%%%%%%%%%%%%%
	    self.linear.set_gradients(w=weight_gradients, b=bias_gradients)  # 设置梯度值
	    
	    list.reverse(input_gradients)  # input_gradients是逆序的，最后输出时需要reverse一下
	    print("sum(weight_gradients) = {}".format(np.sum(weight_gradients)))
	    
	    # np.stack的作用是将列表转变成一个矩阵
	    return np.stack(input_gradients), h_gradient