Python实现的径向基（RBF）神经网络示例

from numpy import array, append, vstack, transpose, reshape, /         dot, true_divide, mean, exp, sqrt, log, /         loadtxt, savetxt, zeros, frombufferfrom numpy.linalg import norm, lstsqfrom multiprocessing import Process, Arrayfrom random import samplefrom time import timefrom sys import stdoutfrom ctypes import c_doublefrom h5py import Filedef metrics(a, b):  return norm(a - b)def gaussian (x, mu, sigma):  return exp(- metrics(mu, x)**2 / (2 * sigma**2))def multiQuadric (x, mu, sigma):  return pow(metrics(mu,x)**2 + sigma**2, 0.5)def invMultiQuadric (x, mu, sigma):  return pow(metrics(mu,x)**2 + sigma**2, -0.5)def plateSpine (x,mu):  r = metrics(mu,x)  return (r**2) * log(r)class Rbf:  def __init__(self, prefix = 'rbf', workers = 4, extra_neurons = 0, from_files = None):    self.prefix = prefix    self.workers = workers    self.extra_neurons = extra_neurons    # Import partial model    if from_files is not None:      w_handle = self.w_handle = File(from_files['w'], 'r')      mu_handle = self.mu_handle = File(from_files['mu'], 'r')      sigma_handle = self.sigma_handle = File(from_files['sigma'], 'r')      self.w = w_handle['w']      self.mu = mu_handle['mu']      self.sigmas = sigma_handle['sigmas']      self.neurons = self.sigmas.shape[0]  def _calculate_error(self, y):    self.error = mean(abs(self.os - y))    self.relative_error = true_divide(self.error, mean(y))  def _generate_mu(self, x):    n = self.n    extra_neurons = self.extra_neurons    # TODO: Make reusable    mu_clusters = loadtxt('clusters100.txt', delimiter='/t')    mu_indices = sample(range(n), extra_neurons)    mu_new = x[mu_indices, :]    mu = vstack((mu_clusters, mu_new))    return mu  def _calculate_sigmas(self):    neurons = self.neurons    mu = self.mu    sigmas = zeros((neurons, ))    for i in xrange(neurons):      dists = [0 for _ in xrange(neurons)]      for j in xrange(neurons):        if i != j:          dists[j] = metrics(mu[i], mu[j])      sigmas[i] = mean(dists)* 2           # max(dists) / sqrt(neurons * 2))    return sigmas  def _calculate_phi(self, x):    C = self.workers    neurons = self.neurons    mu = self.mu    sigmas = self.sigmas    phi = self.phi = None    n = self.n    def heavy_lifting(c, phi):      s = jobs[c][1] - jobs[c][0]      for k, i in enumerate(xrange(jobs[c][0], jobs[c][1])):        for j in xrange(neurons):          # phi[i, j] = metrics(x[i,:], mu[j])**3)          # phi[i, j] = plateSpine(x[i,:], mu[j]))          # phi[i, j] = invMultiQuadric(x[i,:], mu[j], sigmas[j]))          phi[i, j] = multiQuadric(x[i,:], mu[j], sigmas[j])          # phi[i, j] = gaussian(x[i,:], mu[j], sigmas[j]))        if k % 1000 == 0:          percent = true_divide(k, s)*100          print(c, ': {:2.2f}%'.format(percent))      print(c, ': Done')    # distributing the work between 4 workers    shared_array = Array(c_double, n * neurons)    phi = frombuffer(shared_array.get_obj())    phi = phi.reshape((n, neurons))    jobs = []    workers = []    p = n / C    m = n % C    for c in range(C):      jobs.append((c*p, (c+1)*p + (m if c == C-1 else 0)))      worker = Process(target = heavy_lifting, args = (c, phi))      workers.append(worker)      worker.start()    for worker in workers:      worker.join()    return phi  def _do_algebra(self, y):    phi = self.phi    w = lstsq(phi, y)[0]    os = dot(w, transpose(phi))    return w, os    # Saving to HDF5    os_h5 = os_handle.create_dataset('os', data = os)  def train(self, x, y):    self.n = x.shape[0]    ## Initialize HDF5 caches    prefix = self.prefix    postfix = str(self.n) + '-' + str(self.extra_neurons) + '.hdf5'    name_template = prefix + '-{}-' + postfix    phi_handle = self.phi_handle = File(name_template.format('phi'), 'w')    os_handle = self.w_handle = File(name_template.format('os'), 'w')    w_handle = self.w_handle = File(name_template.format('w'), 'w')    mu_handle = self.mu_handle = File(name_template.format('mu'), 'w')    sigma_handle = self.sigma_handle = File(name_template.format('sigma'), 'w')    ## Mu generation    mu = self.mu = self._generate_mu(x)    self.neurons = mu.shape[0]    print('({} neurons)'.format(self.neurons))    # Save to HDF5    mu_h5 = mu_handle.create_dataset('mu', data = mu)    ## Sigma calculation    print('Calculating Sigma...')    sigmas = self.sigmas = self._calculate_sigmas()    # Save to HDF5    sigmas_h5 = sigma_handle.create_dataset('sigmas', data = sigmas)    print('Done')    ## Phi calculation    print('Calculating Phi...')    phi = self.phi = self._calculate_phi(x)    print('Done')    # Saving to HDF5    print('Serializing...')    phi_h5 = phi_handle.create_dataset('phi', data = phi)    del phi    self.phi = phi_h5    print('Done')    ## Algebra    print('Doing final algebra...')    w, os = self.w, _ = self._do_algebra(y)    # Saving to HDF5    w_h5 = w_handle.create_dataset('w', data = w)    os_h5 = os_handle.create_dataset('os', data = os)    ## Calculate error    self._calculate_error(y)    print('Done')  def predict(self, test_data):    mu = self.mu = self.mu.value    sigmas = self.sigmas = self.sigmas.value    w = self.w = self.w.value    print('Calculating phi for test data...')    phi = self._calculate_phi(test_data)    os = dot(w, transpose(phi))    savetxt('iok3834.txt', os, delimiter='/n')    return os  @property  def summary(self):    return '/n'.join( /      ['-----------------',      'Training set size: {}'.format(self.n),      'Hidden layer size: {}'.format(self.neurons),      '-----------------',      'Absolute error  : {:02.2f}'.format(self.error),      'Relative error  : {:02.2f}%'.format(self.relative_error * 100)])def predict(test_data):  mu = File('rbf-mu-212243-2400.hdf5', 'r')['mu'].value  sigmas = File('rbf-sigma-212243-2400.hdf5', 'r')['sigmas'].value  w = File('rbf-w-212243-2400.hdf5', 'r')['w'].value  n = test_data.shape[0]  neur = mu.shape[0]  mu = transpose(mu)  mu.reshape((n, neur))  phi = zeros((n, neur))  for i in range(n):    for j in range(neur):      phi[i, j] = multiQuadric(test_data[i,:], mu[j], sigmas[j])  os = dot(w, transpose(phi))  savetxt('iok3834.txt', os, delimiter='/n')  return os