LSTM_cpd_main.py 文件源码

def generate_noise(D,N):
  """Generate data for the changepoint detection. Data can either be of type 0
  or type 1, but when it's a combination fo both, we define a target label
  Input
  - D,N Dimenstionality arguments D dimensions over N samples
  Output
  - Data in format
  X is a matrix in R^{N x D}
  y is a matrix in R^{N,} not to donfuse with {N,1}"""
  #Check if we have even D, so we can split the array in future
  assert D%2 == 0, 'We need even number of dimensions'
  Dhalf = int(D/2)
  ratioP = 0.5   #balance of targets
  X = np.random.randn(N,D)
  y = np.zeros(N)
  #Generate two filter cofficients
  filters = {}
  filters['b1'],filters['a1'] = signal.butter(4,2.0*cutoff1/fs,btype='lowpass')
  filters['b0'],filters['a0'] = signal.butter(4,2.0*cutoff0/fs,btype='lowpass')
  for i in xrange(N):
    if np.random.rand() > 0.5:  #Half of the samples will have changepoint, other half wont
      signalA = signal.filtfilt(filters['b1'],filters['a1'],X[i])
      signalB = signal.filtfilt(filters['b0'],filters['a0'],X[i])
      X[i] = np.concatenate((signalA[:Dhalf],signalB[:Dhalf]),axis=0)    #Concatenate the two signals
      if True:  #Boolean: do you want to introduce a pattern at the changepoint?
        Dstart = int(Dhalf-pattern_len/2)
        X[i,Dstart:Dstart+pattern_len] = pattern
      y[i] = 1      #The target label
    else:
      mode = int(np.random.rand()>ratioP)
      X[i] = signal.filtfilt(filters['b'+str(mode)],filters['a'+str(mode)],X[i])
      y[i] = 0      #The target label
  return X,y