AdobeEdu
英文提示词：
import numpy as np import utils # Forward Neural Network (FNN) class FNN: # layers = 3 and shape = [input dim, xx, class num] def __init__(self, shape, activation=’sigmoid’) -> None: self._l = len(shape) – 1 # layer number self._w = [] # weight self._b = [] # bias self._z = [i for i in range(self._l)] # wx+b self._a = [i for i in range(self._l)] # activation(wx+b) self._x = None # input feature self._n = None # batch size self._m = None # mean for normalization self._v = None # variance for normalization self._act = activation # activation function for output # He initialization for ly in range(self._l): self._w.append(np.random.randn(shape[ly], shape[ly + 1]) * np.sqrt(2 / shape[ly])) self._b.append(np.zeros((1, shape[ly + 1]))) assert self._l == len(self._w) == len(self._b) def set_normalize(self, m, v): self._m = m self._v = v def forward(self, _x): if self._m is None: self._x = _x else: self._x = utils.do_normalize(_x, self._m, self._v) self._n = self._x.shape[0] # linear -> relu -> linear -> sigmoid self._z[0] = np.dot(self._x, self._w[0]) + self._b[0] for ly in range(1, self._l): self._a[ly – 1] = utils.relu(self._z[ly – 1]) self._z[ly] = np.dot(self._a[ly – 1], self._w[ly]) + self._b[ly] if self._act == ‘softmax’: self._a[self._l – 1] = utils.softmax(self._z[self._l – 1]) else: self._a[self._l – 1] = utils.sigmoid(self._z[self._l – 1]) return self._a[self._l – 1] # if activation function is ‘softmax’, we use cee # else if activation func is ‘sigmoid’, we use mse def backward(self, _y, lr=0.1): # cee dz = self._a[self._l – 1] – _y # mse if self._act != ‘softmax’: dz = dz * self._a[self._l – 1] * (1 – self._a[self._l – 1]) for ly in reversed(range(1, self._l)): dw = 1. / self._n * np.dot(self._a[ly – 1].T, dz) db = 1. / self._n * np.sum(dz, axis=0) self._w[ly] -= lr * dw self._b[ly] -= lr * db da = np.dot(dz, self._w[ly].T) dz = da.copy() dz[self._a[ly – 1] <= 0] = 0 dw = 1. / self._n * np.dot(self._x.T, dz) db = 1. / self._n * np.sum(dz, axis=0) self._w[0] -= lr * dw self._b[0] -= lr * db # save model weights which in ‘datasetName.txt’ def save(self, _name): file = open(_name + ‘.txt’, ‘w’) for fp in self._w: for i in range(fp.shape[0]): for j in range(fp.shape[1]): file.write(str(fp[i][j])) file.write(‘ ‘) file.write(‘\n’) for fp in self._b: for i in range(fp.shape[0]): for j in range(fp.shape[1]): file.write(str(fp[i][j])) file.write(‘ ‘) file.write(‘\n’) file.close() # load model weights which in ‘datasetName.txt’ def load(self, _path): file = open(_path, ‘r’) s = file.readlines() assert len(s) == len(self._w) * 2 n = len(s) // 2 for i in range(n): s[i] = s[i].strip() s[i] = s[i].split(‘ ‘) s[i] = list(map(float, s[i])) arr = np.array(s[i]) arr = arr.reshape((self._w[i].shape[0], self._w[i].shape[1])) assert arr.shape == self._w[i].shape self._w[i] = arr s[i + n] = s[i + n].strip() s[i + n] = s[i + n].split(‘ ‘) s[i + n] = list(map(float, s[i + n])) arr = np.array(s[i + n]) arr = arr.reshape((self._b[i].shape[0], self._b[i].shape[1])) assert arr.shape == self._b[i].shape self._b[i] = arr
