def check_binary_func(x, y):
     f_add      = lambda x, y: x+y
     f_add_grad = lambda x, y: [nd.ones(x.shape), nd.ones(y.shape)]
     autograd_assert(x, y, func=f_add, grad_func=f_add_grad)
     f_mul      = lambda x, y: x*y
     f_mul_grad = lambda x, y: [y, x]
     autograd_assert(x, y, func=f_mul, grad_func=f_mul_grad)
     f_compose  = lambda x, y: x+x*y
     f_compose_grad = lambda x, y: [nd.ones(x.shape) + y, x]
     autograd_assert(x, y, func=f_compose, grad_func=f_compose_grad)

def test_binary_func():
    x = nd.uniform(shape=(4, 5))
    y = nd.uniform(shape=(4, 5))
    f_add      = lambda x, y: x+y
    f_add_grad = lambda x, y: [nd.ones(x.shape), nd.ones(y.shape)]
    autograd_assert(x, y, func=f_add, grad_func=f_add_grad)
    f_mul      = lambda x, y: x*y
    f_mul_grad = lambda x, y: [y, x]
    autograd_assert(x, y, func=f_mul, grad_func=f_mul_grad)
    f_compose  = lambda x, y: x+x*y
    f_compose_grad = lambda x, y: [nd.ones(x.shape) + y, x]
    autograd_assert(x, y, func=f_compose, grad_func=f_compose_grad)

def test_argnum():
    def f_with_mode(a, b, mode):
        if mode:
            return a+b
        else:
            return a*b

    a = nd.uniform(shape=(3, 2))
    b = nd.uniform(shape=(3, 2))
    f_add_grad = lambda x, y, mode: [nd.ones(x.shape), nd.ones(y.shape)]
    f_mul_grad = lambda x, y, mode: [y, x]
    autograd_assert(a, b, True,
        argnum=[0, 1], func=f_with_mode, grad_func=f_add_grad)
    autograd_assert(a, b, False,
        argnum=[0, 1], func=f_with_mode, grad_func=f_mul_grad)

def test_detach_updated_grad():
    x = nd.ones((2, 2))
    dx = nd.zeros_like(x)
    y = nd.ones_like(x)
    dy = nd.zeros_like(x)
    mark_variables([x, y], [dx, dy])
    assert x._fresh_grad == False
    assert y._fresh_grad == False

    with train_section():
        x2 = x + 2
        y2  = x2 + y
        y2.backward()
    assert (dx.asnumpy() == 1).all()
    assert x._fresh_grad == True
    assert y._fresh_grad == True

    dx[:] = 0
    x._fresh_grad = False
    y._fresh_grad = False
    assert x._fresh_grad == False
    assert y._fresh_grad == False
    with train_section():
        x2 = x + 2
        x2 = x2.detach()
        y2  = x2 + y
        y2.backward()
    assert (dx.asnumpy() == 0).all()
    assert y._fresh_grad == True
    assert x._fresh_grad == False

 def gan_loss(input,target_is_real):
     if target_is_real:
         target = nd.ones(input.shape,ctx=input.context)
     else:
         target = nd.zeros(input.shape, ctx=input.context)
     #mse loss for lsgan
     e = ((input - target) ** 2).mean(axis=0, exclude=True)
     return e

def bilinear(x, W, y, input_size, seq_len, batch_size, num_outputs=1, bias_x=False, bias_y=False):
    """Do xWy

    Parameters
    ----------
    x : NDArray
        (input_size x seq_len) x batch_size
    W : NDArray
        (num_outputs x ny) x nx
    y : NDArray
        (input_size x seq_len) x batch_size
    input_size : int
        input dimension
    seq_len : int
        sequence length
    batch_size : int
        batch size
    num_outputs : int
        number of outputs
    bias_x : bool
        whether concat bias vector to input x
    bias_y : bool
        whether concat bias vector to input y

    Returns
    -------
    output : NDArray
        [seq_len_y x seq_len_x if output_size == 1 else seq_len_y x num_outputs x seq_len_x] x batch_size
    """
    if bias_x:
        x = nd.concat(x, nd.ones((1, seq_len, batch_size)), dim=0)
    if bias_y:
        y = nd.concat(y, nd.ones((1, seq_len, batch_size)), dim=0)

    nx, ny = input_size + bias_x, input_size + bias_y
    # W: (num_outputs x ny) x nx
    lin = nd.dot(W, x)
    if num_outputs > 1:
        lin = reshape_fortran(lin, (ny, num_outputs * seq_len, batch_size))
    y = y.transpose([2, 1, 0])  # May cause performance issues
    lin = lin.transpose([2, 1, 0])
    blin = nd.batch_dot(lin, y, transpose_b=True)
    blin = blin.transpose([2, 1, 0])
    if num_outputs > 1:
        blin = reshape_fortran(blin, (seq_len, num_outputs, seq_len, batch_size))
    return blin

def test_training():
    x = nd.ones((10, 10))
    with train_section():
        y = nd.Dropout(x, p=0.5)
        assert not (y.asnumpy() == x.asnumpy()).all()
        with test_section():
            y = nd.Dropout(x, p=0.5)
            assert (y.asnumpy() == x.asnumpy()).all()

def test_training():
    x = nd.ones((10, 10))
    with record():
        y = nd.Dropout(x, p=0.5)
        assert not (y.asnumpy() == x.asnumpy()).all()
        with pause():
            y = nd.Dropout(x, p=0.5)
            assert (y.asnumpy() == x.asnumpy()).all()

 def forward(self, x):
     if  isinstance(x, np.ndarray):
         x = nd.array(x)
     if self._max_len > x.size:
         pad = nd.ones((self._max_len - x.size,)) * self._fill_value
         x = nd.concat(x, pad, dim=0)
     elif self._max_len < x.size:
         x = x[:self._max_len]
     return x

def dumpR(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
  print('dump verification embedding..')
  data_list = data_set[0]
  issame_list = data_set[1]
  model = mx_model
  embeddings_list = []
  if data_extra is not None:
    _data_extra = nd.array(data_extra)
  time_consumed = 0.0
  if label_shape is None:
    _label = nd.ones( (batch_size,) )
  else:
    _label = nd.ones( label_shape )
  for i in xrange( len(data_list) ):
    data = data_list[i]
    embeddings = None
    ba = 0
    while ba<data.shape[0]:
      bb = min(ba+batch_size, data.shape[0])
      count = bb-ba
      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
      #print(_data.shape, _label.shape)
      time0 = datetime.datetime.now()
      if data_extra is None:
        db = mx.io.DataBatch(data=(_data,), label=(_label,))
      else:
        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
      model.forward(db, is_train=False)
      net_out = model.get_outputs()
      _embeddings = net_out[0].asnumpy()
      time_now = datetime.datetime.now()
      diff = time_now - time0
      time_consumed+=diff.total_seconds()
      if embeddings is None:
        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
      ba = bb
    embeddings_list.append(embeddings)
  embeddings = embeddings_list[0] + embeddings_list[1]
  embeddings = sklearn.preprocessing.normalize(embeddings)
  actual_issame = np.asarray(issame_list)
  outname = os.path.join('temp.bin')
  with open(outname, 'wb') as f:
    pickle.dump((embeddings, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)

 def check_unary_func(x):
     f_exp         = lambda x: nd.exp(x)
     f_exp_grad    = lambda x: [nd.exp(x)]
     autograd_assert(x, func=f_exp, grad_func=f_exp_grad)
     f_half        = lambda x: x/2
     f_half_grad   = lambda x: [nd.ones(x.shape) * 0.5]
     autograd_assert(x, func=f_half, grad_func=f_half_grad)
     f_square      = lambda x: x**2
     f_square_grad = lambda x: [2*x]
     autograd_assert(x, func=f_square, grad_func=f_square_grad)

def test_unary_func():
    x = nd.uniform(shape=(4, 5))
    f_exp         = lambda x: nd.exp(x)
    f_exp_grad    = lambda x: [nd.exp(x)]
    autograd_assert(x, func=f_exp, grad_func=f_exp_grad)
    f_half        = lambda x: x/2
    f_half_grad   = lambda x: [nd.ones(x.shape) * 0.5]
    autograd_assert(x, func=f_half, grad_func=f_half_grad)
    f_square      = lambda x: x**2
    f_square_grad = lambda x: [2*x]
    autograd_assert(x, func=f_square, grad_func=f_square_grad)

def test_module_input_grads():
    a = mx.sym.Variable('a', __layout__='NC')
    b = mx.sym.Variable('b', __layout__='NC')
    c = mx.sym.Variable('c', __layout__='NC')

    c = a + 2 * b + 3 * c
    net = mx.mod.Module(c, data_names=['b', 'c', 'a'], label_names=None,
                        context=[mx.cpu(0), mx.cpu(1)])
    net.bind(data_shapes=[['b', (5, 5)], ['c', (5, 5)], ['a', (5, 5)]],
             label_shapes=None, inputs_need_grad=True)
    net.init_params()

    net.forward(data_batch=mx.io.DataBatch(data=[nd.ones((5, 5)),
                                                 nd.ones((5, 5)),
                                                 nd.ones((5, 5))]))
    net.backward(out_grads=[nd.ones((5, 5))])
    input_grads = net.get_input_grads()
    b_grad = input_grads[0].asnumpy()
    c_grad = input_grads[1].asnumpy()
    a_grad = input_grads[2].asnumpy()
    assert np.all(a_grad == 1), a_grad
    assert np.all(b_grad == 2), b_grad
    assert np.all(c_grad == 3), c_grad

def test_out_grads():
    x = nd.ones((3, 5))
    dx = nd.zeros_like(x)
    mark_variables([x], [dx])
    da = None
    db = nd.array([1,2,3,4,5])
    dc = nd.array([5,4,3,2,1])

    with train_section():
        a, b, c = nd.split(x, axis=0, num_outputs=3, squeeze_axis=True)
        backward([a, b, c], [da, db, dc])

    assert (dx.asnumpy() == np.array(
        [[1,1,1,1,1],
         [1,2,3,4,5],
         [5,4,3,2,1]])).all()

    def lfw_test(nbatch):
      print('testing lfw..')
      embeddings_list = []
      for i in xrange( len(lfw_data_list) ):
        lfw_data = lfw_data_list[i]
        embeddings = None
        ba = 0
        while ba<lfw_data.shape[0]:
          bb = min(ba+args.batch_size, lfw_data.shape[0])
          _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
          _label = nd.ones( (bb-ba,) )
          db = mx.io.DataBatch(data=(_data,), label=(_label,))
          model.forward(db, is_train=False)
          net_out = model.get_outputs()
          _embeddings = net_out[0].asnumpy()
          if embeddings is None:
            embeddings = np.zeros( (lfw_data.shape[0], _embeddings.shape[1]) )
          embeddings[ba:bb,:] = _embeddings
          ba = bb
        embeddings_list.append(embeddings)

      acc_list = []
      embeddings = embeddings_list[0]
      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
      acc_list.append(np.mean(accuracy))
      print('[%d]Accuracy: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
      embeddings = np.concatenate(embeddings_list, axis=1)
      embeddings = sklearn.preprocessing.normalize(embeddings)
      print(embeddings.shape)
      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
      acc_list.append(np.mean(accuracy))
      print('[%d]Accuracy-Flip: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
      pca = PCA(n_components=128)
      embeddings = pca.fit_transform(embeddings)
      embeddings = sklearn.preprocessing.normalize(embeddings)
      print(embeddings.shape)
      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
      acc_list.append(np.mean(accuracy))
      print('[%d]Accuracy-PCA: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
      return max(*acc_list)

#欠拟合：机器学习模型无法得到较低训练误差。
#过拟合：机器学习模型的训练误差远小于其在测试数据集上的误差。

#高维线性回归
# y = 0.05 + sum(0.01*xi) + noise # 这里噪音服从均值0和标准差为0.01的正态分布。

from mxnet import ndarray as nd
from mxnet import autograd
from mxnet import gluon
import mxnet as mx

num_train = 20#训练集大小
num_test = 100#测试集大小
num_inputs = 200#输入神经元个数 xi的个数
#真实模型参数
true_w = nd.ones((num_inputs, 1)) * 0.01# 权重
true_b = 0.05#偏置

#生成 数据集
X = nd.random.normal(shape=(num_train + num_test, num_inputs))#输入
y = nd.dot(X, true_w) + true_b # y = 0.05 + sum(0.01*xi) 
y += .01 * nd.random.normal(shape=y.shape)#噪声 y = 0.05 + sum(0.01*xi) + noise 

X_train, X_test = X[:num_train, :], X[num_train:, :]# 0~19 行  20~99行
y_train, y_test = y[:num_train], y[num_train:]

# 不断读取数据块
import random
batch_size = 1
def data_iter(num_examples):
    idx = list(range(num_examples))

def test(data_set, mx_model, batch_size, nfolds=10, data_extra = None, label_shape = None):
  print('testing verification..')
  data_list = data_set[0]
  issame_list = data_set[1]
  model = mx_model
  embeddings_list = []
  if data_extra is not None:
    _data_extra = nd.array(data_extra)
  time_consumed = 0.0
  if label_shape is None:
    _label = nd.ones( (batch_size,) )
  else:
    _label = nd.ones( label_shape )
  for i in xrange( len(data_list) ):
    data = data_list[i]
    embeddings = None
    ba = 0
    while ba<data.shape[0]:
      bb = min(ba+batch_size, data.shape[0])
      count = bb-ba
      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
      #print(_data.shape, _label.shape)
      time0 = datetime.datetime.now()
      if data_extra is None:
        db = mx.io.DataBatch(data=(_data,), label=(_label,))
      else:
        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
      model.forward(db, is_train=False)
      net_out = model.get_outputs()
      #_arg, _aux = model.get_params()
      #__arg = {}
      #for k,v in _arg.iteritems():
      #  __arg[k] = v.as_in_context(_ctx)
      #_arg = __arg
      #_arg["data"] = _data.as_in_context(_ctx)
      #_arg["softmax_label"] = _label.as_in_context(_ctx)
      #for k,v in _arg.iteritems():
      #  print(k,v.context)
      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
      #exe.forward(is_train=False)
      #net_out = exe.outputs
      _embeddings = net_out[0].asnumpy()
      time_now = datetime.datetime.now()
      diff = time_now - time0
      time_consumed+=diff.total_seconds()
      #print(_embeddings.shape)
      if embeddings is None:
        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
      ba = bb
    embeddings_list.append(embeddings)

  _xnorm = 0.0
  _xnorm_cnt = 0
  for embed in embeddings_list:
    for i in xrange(embed.shape[0]):
      _em = embed[i]
      _norm=np.linalg.norm(_em)
      #print(_em.shape, _norm)
      _xnorm+=_norm
      _xnorm_cnt+=1
  _xnorm /= _xnorm_cnt

  embeddings = embeddings_list[0].copy()
  embeddings = sklearn.preprocessing.normalize(embeddings)
  acc1 = 0.0
  std1 = 0.0
  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
  #acc1, std1 = np.mean(accuracy), np.std(accuracy)

  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
  #embeddings = np.concatenate(embeddings_list, axis=1)
  embeddings = embeddings_list[0] + embeddings_list[1]
  embeddings = sklearn.preprocessing.normalize(embeddings)
  print(embeddings.shape)
  print('infer time', time_consumed)
  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
  acc2, std2 = np.mean(accuracy), np.std(accuracy)
  return acc1, std1, acc2, std2, _xnorm, embeddings_list

def test_badcase(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
  print('testing verification badcase..')
  data_list = data_set[0]
  issame_list = data_set[1]
  model = mx_model
  embeddings_list = []
  if data_extra is not None:
    _data_extra = nd.array(data_extra)
  time_consumed = 0.0
  if label_shape is None:
    _label = nd.ones( (batch_size,) )
  else:
    _label = nd.ones( label_shape )
  for i in xrange( len(data_list) ):
    data = data_list[i]
    embeddings = None
    ba = 0
    while ba<data.shape[0]:
      bb = min(ba+batch_size, data.shape[0])
      count = bb-ba
      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
      #print(_data.shape, _label.shape)
      time0 = datetime.datetime.now()
      if data_extra is None:
        db = mx.io.DataBatch(data=(_data,), label=(_label,))
      else:
        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
      model.forward(db, is_train=False)
      net_out = model.get_outputs()
      _embeddings = net_out[0].asnumpy()
      time_now = datetime.datetime.now()
      diff = time_now - time0
      time_consumed+=diff.total_seconds()
      if embeddings is None:
        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
      ba = bb
    embeddings_list.append(embeddings)
  embeddings = embeddings_list[0] + embeddings_list[1]
  embeddings = sklearn.preprocessing.normalize(embeddings)
  thresholds = np.arange(0, 4, 0.01)
  actual_issame = np.asarray(issame_list)
  nrof_folds = 10
  embeddings1 = embeddings[0::2]
  embeddings2 = embeddings[1::2]
  assert(embeddings1.shape[0] == embeddings2.shape[0])
  assert(embeddings1.shape[1] == embeddings2.shape[1])
  nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
  nrof_thresholds = len(thresholds)
  k_fold = LFold(n_splits=nrof_folds, shuffle=False)
  
  tprs = np.zeros((nrof_folds,nrof_thresholds))
  fprs = np.zeros((nrof_folds,nrof_thresholds))
  accuracy = np.zeros((nrof_folds))
  indices = np.arange(nrof_pairs)
  
  diff = np.subtract(embeddings1, embeddings2)
  dist = np.sum(np.square(diff),1)
  data = data_list[0]

  pouts = []
  nouts = []
  
  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
       
      # Find the best threshold for the fold
      acc_train = np.zeros((nrof_thresholds))
      #print(train_set)
      #print(train_set.__class__)
      for threshold_idx, threshold in enumerate(thresholds):
          p2 = dist[train_set]
          p3 = actual_issame[train_set]
          _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
      best_threshold_index = np.argmax(acc_train)
      for threshold_idx, threshold in enumerate(thresholds):
          tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
      _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
      best_threshold = thresholds[best_threshold_index]
      for iid in test_set:
        ida = iid*2
        idb = ida+1
        asame = actual_issame[iid]
        _dist = dist[iid]
        violate = _dist - best_threshold
        if not asame:
          violate *= -1.0
        if violate>0.0:
          imga = data[ida].asnumpy().transpose( (1,2,0) )[...,::-1] #to bgr
          imgb = data[idb].asnumpy().transpose( (1,2,0) )[...,::-1]
          #print(imga.shape, imgb.shape, violate, asame, _dist)
          if asame:
            pouts.append( (imga, imgb, _dist, best_threshold, ida) )
          else:
            nouts.append( (imga, imgb, _dist, best_threshold, ida) )

        
  tpr = np.mean(tprs,0)
  fpr = np.mean(fprs,0)
  acc = np.mean(accuracy)
  pouts = sorted(pouts, key = lambda x: x[2], reverse=True)
#.........这里部分代码省略.........

    print 'encoder'
    print '=========='
    print o.asnumpy()
    print h[0].asnumpy()
    print '=========='

    attn_decoder = AttnDecoderRNN(2, 5, 1, 10, 0.1)
    attn_decoder.initialize()

    for i in attn_decoder.collect_params().values():
        i.data()[:] = 1.0

    input = F.array([0])
    hidden = attn_decoder.initHidden(ctx=mx.cpu())

    o, h, a = attn_decoder(input, hidden, 0.5*F.ones((10, 2)))

    print 'attn_decoder'
    print '=========='
    print o.asnumpy()
    print h[0].asnumpy()
    print '=========='

    assert False

encoder = EncoderRNN(input_lang.n_words, opt.hidden_size, opt.num_layers)
attn_decoder = AttnDecoderRNN(opt.hidden_size, output_lang.n_words,
                               opt.num_layers, opt.max_length, dropout_p=0.1)

trainIters(encoder, attn_decoder, ctx, opt)

def test(lfw_set, mx_model, batch_size):
  print('testing lfw..')
  lfw_data_list = lfw_set[0]
  issame_list = lfw_set[1]
  model = mx_model
  embeddings_list = []
  for i in xrange( len(lfw_data_list) ):
    lfw_data = lfw_data_list[i]
    embeddings = None
    ba = 0
    while ba<lfw_data.shape[0]:
      bb = min(ba+batch_size, lfw_data.shape[0])
      _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
      _label = nd.ones( (bb-ba,) )
      #print(_data.shape, _label.shape)
      db = mx.io.DataBatch(data=(_data,), label=(_label,))
      model.forward(db, is_train=False)
      net_out = model.get_outputs()
      #_arg, _aux = model.get_params()
      #__arg = {}
      #for k,v in _arg.iteritems():
      #  __arg[k] = v.as_in_context(_ctx)
      #_arg = __arg
      #_arg["data"] = _data.as_in_context(_ctx)
      #_arg["softmax_label"] = _label.as_in_context(_ctx)
      #for k,v in _arg.iteritems():
      #  print(k,v.context)
      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
      #exe.forward(is_train=False)
      #net_out = exe.outputs
      _embeddings = net_out[0].asnumpy()
      #print(_embeddings.shape)
      if embeddings is None:
        embeddings = np.zeros( (lfw_data.shape[0], _embeddings.shape[1]) )
      embeddings[ba:bb,:] = _embeddings
      ba = bb
    embeddings_list.append(embeddings)

  _xnorm = 0.0
  _xnorm_cnt = 0
  for embed in embeddings_list:
    for i in xrange(embed.shape[0]):
      _em = embed[i]
      _norm=np.linalg.norm(_em)
      #print(_em.shape, _norm)
      _xnorm+=_norm
      _xnorm_cnt+=1
  _xnorm /= _xnorm_cnt

  embeddings = embeddings_list[0].copy()
  embeddings = sklearn.preprocessing.normalize(embeddings)
  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
  acc1, std1 = np.mean(accuracy), np.std(accuracy)
  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
  #embeddings = np.concatenate(embeddings_list, axis=1)
  embeddings = embeddings_list[0] + embeddings_list[1]
  embeddings = sklearn.preprocessing.normalize(embeddings)
  print(embeddings.shape)
  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
  acc2, std2 = np.mean(accuracy), np.std(accuracy)
  return acc1, std1, acc2, std2, _xnorm, embeddings_list