数字识别 Demo，基于 Python 训练出的模型，将 b 和 w 取出，并将 feedForward 函数用 JS 实现。效果并不是太好，但在模拟训练集的笔画后，大体上能识别出来。

  • 效果展现
  • 核心代码
  • 在线 Demo
  • Python 参数获取

效果展现

Demo JS 代码

最核心的代码如下，通过 numjs 用于实现

\[a^{(L)} = \sigma (w^{(L)}a^{(L - 1)} + b^{(L)})\] 
class NetWork {
  constructor(sizes, biases, weights) {
    this.num_layers = sizes.length;
    this.sizes = sizes;
    this.biases = biases;
    this.weights = weights;
  }

  feedForward(input) {
    var a = nj.array(input);
    for (var layer = 1; layer < this.num_layers; layer++) {
      var weights = nj.array(this.weights[layer - 1]);
      var biases = nj.array(this.biases[layer - 1]);

      a = nj.sigmoid(
        nj.add(
          nj.dot(weights, a).reshape(biases.shape),
          biases
        )
      );
    }

    return a;
  }
}

在线Demo

Python 取出参数代码


fo = open('w0.txt', 'w')
np.set_printoptions(threshold = np.prod(self.weights[0].shape))
fo.write(repr(self.weights[0]));
fo.close();

fo = open('w1.txt', 'w')
np.set_printoptions(threshold = np.prod(self.weights[1].shape))
fo.write(repr(self.weights[1]));
fo.close();

fo = open('b0.txt', 'w')
np.set_printoptions(threshold = np.prod(self.biases[0].shape))
fo.write(repr(self.biases[0]));
fo.close();

fo = open('b1.txt', 'w')
np.set_printoptions(threshold = np.prod(self.biases[1].shape))
fo.write(repr(self.biases[1]));
fo.close();

References