基于numpy实现的全连接深度学习工具

实现以下功能:

• 激活函数:relu,sigmoid,tanh,softmax
• 损失函数:平均绝对误差(MAE),交叉熵(cross_entropy)
• 前向传播:前向传播(forward),反向传播(backward 使用梯度下降优化器)
• 权重矩阵:权重矩阵形状推导和模型总结(summary)
• 随机函数:随机函数(shuffle)
• 预测函数:输出模型的预测结果(predict)
• 验证损失:使用模型预测并计算损失(verify)
• 绘制曲线:绘制损失变化曲线(plot)

在mnist数据集上的准确率为0.847 使用封装的fit训练函数训练

learning_rate = 1e-4
model = models()
model.add(Dense(input_size=None, hidden_layer_num=128, activation=activations.relu))
model.add(Dense(input_size=None, hidden_layer_num=128, activation=activations.relu))
model.add(Dense(input_size=None, hidden_layer_num=10, activation=activations.softmax))
for i in range(30):
    print('step',i)
    model.fit(train_data=train_x, train_labels=train_y, epochs=1, batch_size=20,
              learning_rate=learning_rate, measure=measure.cross_entropy, shuffle=True)
    model.verify(test_data=test_x, test_labels=test_y, batch_size=20)
model.plot() 

手动进行前向传播和反向传播训练

model = models()
model.add(Dense(input_size=None, hidden_layer_num=32, activation=activations.relu))
model.add(Dense(input_size=None, hidden_layer_num=10, activation=activations.softmax))

model.batch_input = x
model.batch_label = y
model.summary()
model.learning_rate = 0.0003

for i in range(10000):
    data = mnist.train.next_batch(batch_size)
    x = data[0]
    y = data[1]
    model.batch_input = x
    model.batch_label = y
    model.forward()
    model.backward()
model.plot()

demo提供了训练mnist数据的完整示例