【水文】基于 MNN 训练能力解方程

【水文】基于 MNN 训练能力解方程

简介

许久没上,水篇文章,事由同学想由房贷月还款额推测年利率:
月还款额=pow(年利率/12+1,期数)*(年利率/12)*贷款总额/(pow(年利率/12+1,期数)-1)

这个方程直接求解感觉比较麻烦,考虑使用梯度下降法。深度学习框架基本都支持求导和优化,MNN 也支持。所以安装 pymnn ,把方程用 MNN 的仿numpy接口实现一遍,然后写个 loss 训练就好了。

基于 MNN 代码实现

import time
import MNN.numpy as np
import MNN
nn = MNN.nn
F = MNN.expr

# open lazy evaluation for train
F.lazy_eval(True)

# month_pay=pow(rate/12+1, times)*(rate/12)*total/(pow(rate/12+1,times)-1)
# Know month_pa, total, times, solve rate
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        one = np.array([0.001])
        one.fix_as_trainable()
        self.rate = one

    def forward(self, times, total):
        r12 = self.rate / 12.0
        r12_1 = r12 + np.array([1.0])
        total_rate = np.power(r12_1, times)
        p0 = (total_rate * r12 * total) / (total_rate-np.array([1.0]))
        return p0

model = Net()
opt = MNN.optim.SGD(model, 0.0000000001, 0.9, 0.0005)

for iter in range(0, 1000):
    times = np.array([60.0])
    month_pay = np.array([12439.12])
    total = np.array([630000.0])
    month_comp = model.forward(times, total)
    diff = month_pay - month_comp
    loss = diff * diff
    opt.step(loss)

times = np.array([60.0])
month_pay = np.array([12439.12])
total = np.array([630000.0])
month_comp = model.forward(times, total)
print("rate:", model.rate, " ; month_comp: ", month_comp)

运行结果

hw.cpufamily: 458787763 , size = 4
The device support i8sdot:1, support fp16:1, support i8mm: 0
rate: array([0.06879799], dtype=float32) ; month_comp: array([12439.118], dtype=float32)

你可能感兴趣的:(mnn,人工智能,深度学习)