Python tensorflow.matrix_band_part方法代码示例

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def ones_matrix_band_part(rows, cols, num_lower, num_upper, out_shape=None):
  """Matrix band part of ones."""
  if all([isinstance(el, int) for el in [rows, cols, num_lower, num_upper]]):
    # Needed info is constant, so we construct in numpy
    if num_lower < 0:
      num_lower = rows - 1
    if num_upper < 0:
      num_upper = cols - 1
    lower_mask = np.tri(cols, rows, num_lower).T
    upper_mask = np.tri(rows, cols, num_upper)
    band = np.ones((rows, cols)) * lower_mask * upper_mask
    if out_shape:
      band = band.reshape(out_shape)
    band = tf.constant(band, tf.float32)
  else:
    band = tf.matrix_band_part(
        tf.ones([rows, cols]), tf.cast(num_lower, tf.int64),
        tf.cast(num_upper, tf.int64))
    if out_shape:
      band = tf.reshape(band, out_shape)

  return band 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def _ones_matrix_band_part(rows, cols, num_lower, num_upper,
    out_shape=None):
    """Matrix band part of ones.
    """
    if all([isinstance(el, int) for el in [rows, cols, num_lower,
        num_upper]]):
    # Needed info is constant, so we construct in numpy
        if num_lower < 0:
            num_lower = rows - 1
        if num_upper < 0:
            num_upper = cols - 1
        lower_mask = np.tri(cols, rows, num_lower).T
        upper_mask = np.tri(rows, cols, num_upper)
        band = np.ones((rows, cols)) * lower_mask * upper_mask
        if out_shape:
            band = band.reshape(out_shape)
        band = tf.constant(band, tf.float32)
    else:
        band = tf.matrix_band_part(tf.ones([rows, cols]),
                                   tf.cast(num_lower, tf.int64),
                                   tf.cast(num_upper, tf.int64))
        if out_shape:
            band = tf.reshape(band, out_shape)
    return band 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def _GetMatrixBandPartTest(dtype_, batch_shape_, shape_):

  def Test(self):
    mat = np.ones(shape_).astype(dtype_)
    batch_mat = np.tile(mat, batch_shape + (1, 1))
    with self.test_session(use_gpu=True):
      for lower in -1, 0, 1, shape_[-2] - 1:
        for upper in -1, 0, 1, shape_[-1] - 1:
          band_np = mat
          if lower >= 0:
            band_np = np.triu(band_np, -lower)
          if upper >= 0:
            band_np = np.tril(band_np, upper)
          if batch_shape is not ():
            band_np = np.tile(band_np, batch_shape + (1, 1))
          band = tf.matrix_band_part(batch_mat, lower, upper)
          self.assertAllEqual(band_np, band.eval())

  return Test 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def _build_predict_train(self):
        Kuf = self._Kuf
        Kuu = [make_Kuu(kern, a, b, self.ms) for kern, a, b, in zip(self.kerns, self.a, self.b)]
        KiKuf = [Kuu_d.solve(Kuf_d) for Kuu_d, Kuf_d in zip(Kuu, Kuf)]
        KfuKi = [tf.transpose(mat) for mat in KiKuf]

        mu = kvs_dot_vec(KfuKi, self.q_mu)
        L = tf.matrix_band_part(self.q_sqrt, -1, 0)
        tmp1 = kvs_dot_mat(KfuKi, L, num_cols=np.prod(self.Ms))

        # Kff:
        var = reduce(tf.multiply, [k.Kdiag(self.X[:, i:i+1]) for i, k in enumerate(self.kerns)])
        # Projected variance Kfu Ki [WWT] Ki Kuf
        # var = var + reduce(tf.multiply, [tf.reduce_sum(tf.square(tmp1_d), 0) for tmp1_d in tmp1])
        var = var + tf.reduce_sum(tf.square(tmp1), 1)
        # Qff
        var = var - reduce(tf.multiply, [tf.reduce_sum(Kuf_d * KiKuf_d, 0) for Kuf_d, KiKuf_d in zip(Kuf, KiKuf)])
        var = tf.reshape(var, (-1, 1))

        return mu, var 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def get_decoder_self_attention_bias(length, dtype=tf.float32):
  """Calculate bias for decoder that maintains model's autoregressive property.

  Creates a tensor that masks out locations that correspond to illegal
  connections, so prediction at position i cannot draw information from future
  positions.

  Args:
    length: int length of sequences in batch.

  Returns:
    float tensor of shape [1, 1, length, length]
  """
  #print("get_decoder_self_attention_bias", dtype)

  with tf.name_scope("decoder_self_attention_bias"):
    #valid_locs = tf.matrix_band_part(tf.ones([length, length], dtype=dtype), -1, 0)
    valid_locs = tf.matrix_band_part(tf.ones([length, length], dtype=tf.float32), -1, 0)
    valid_locs = tf.reshape(valid_locs, [1, 1, length, length])
    neg_inf=_NEG_INF #if (dtype==tf.float32) else _NEG_INF_FP16
    bias = neg_inf * (1.0 - valid_locs)
    #bias=tf.saturate_cast(bias, dtype=dtype)
  return bias 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def _ones_matrix_band_part(rows, cols, num_lower, num_upper,
    out_shape=None):
    """Matrix band part of ones.
    """
    if all([isinstance(el, int) for el in [rows, cols, num_lower,
        num_upper]]):
        # Needed info is constant, so we construct in numpy
        if num_lower < 0:
            num_lower = rows - 1
        if num_upper < 0:
            num_upper = cols - 1
        lower_mask = np.tri(cols, rows, num_lower).T
        upper_mask = np.tri(rows, cols, num_upper)
        band = np.ones((rows, cols)) * lower_mask * upper_mask
        if out_shape:
            band = band.reshape(out_shape)
        band = tf.constant(band, tf.float32)
    else:
        band = tf.matrix_band_part(tf.ones([rows, cols]),
                                   tf.cast(num_lower, tf.int64),
                                   tf.cast(num_upper, tf.int64))
        if out_shape:
            band = tf.reshape(band, out_shape)
    return band 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def get_decoder_self_attention_bias(length):
  """Calculate bias for decoder that maintains model's autoregressive property.

  Creates a tensor that masks out locations that correspond to illegal
  connections, so prediction at position i cannot draw information from future
  positions.

  Args:
    length: int length of sequences in batch.

  Returns:
    float tensor of shape [1, 1, length, length]
  """
  with tf.name_scope("decoder_self_attention_bias"):
    valid_locs = tf.matrix_band_part(tf.ones([length, length]), -1, 0)
    valid_locs = tf.reshape(valid_locs, [1, 1, length, length])
    decoder_bias = _NEG_INF * (1.0 - valid_locs)
  return decoder_bias 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def get_attention_mask(nd, ns, *, dtype):
    """
    this is a TPU compatible version of tf.matrix_band_part(tf.ones([nd, ns]), -1, ns-nd)
    where the lower right triangle contains 1s
    """
    i = tf.range(nd)[:, None]
    j = tf.range(ns)
    m = i >= j - ns + nd
    return tf.cast(m, dtype) 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def build_variational(hps, kernel, z_pos, x, n_particles):
    bn = zs.BayesianNet()
    z_mean = tf.get_variable(
        'z/mean', [hps.n_z], hps.dtype, tf.zeros_initializer())
    z_cov_raw = tf.get_variable(
        'z/cov_raw', initializer=tf.eye(hps.n_z, dtype=hps.dtype))
    z_cov_tril = tf.matrix_set_diag(
        tf.matrix_band_part(z_cov_raw, -1, 0),
        tf.nn.softplus(tf.matrix_diag_part(z_cov_raw)))
    fz = bn.multivariate_normal_cholesky(
        'fz', z_mean, z_cov_tril, n_samples=n_particles)
    bn.stochastic('fx', gp_conditional(z_pos, fz, x, False, kernel))
    return bn 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def cumulative_max(x):
    """Takes the (inclusive) cumulative maximum along the last axis of x. (Not efficient.)"""
    x = tf.convert_to_tensor(x)
    with tf.name_scope('cumulative_max', values=[x]) as scope:
        repeated = tf.tile(
            tf.expand_dims(x, axis=-1),
            tf.concat([tf.ones(x.shape.rank, dtype=tf.int32), tf.shape(x)[-1:]], axis=0))
        trues = tf.ones_like(repeated, dtype=tf.bool)
        upper_triangle = tf.matrix_band_part(trues, 0, -1)
        neg_inf = tf.ones_like(repeated) * tf.dtypes.saturate_cast(-np.inf, dtype=x.dtype)
        prefixes = tf.where(upper_triangle, repeated, neg_inf)
        return tf.math.reduce_max(prefixes, axis=-2, name=scope) 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def attention_mask(nd, ns, *, dtype):
    """1's in the lower triangle, counting from the lower right corner.

    Same as tf.matrix_band_part(tf.ones([nd, ns]), -1, ns-nd), but doesn't produce garbage on TPUs.
    """
    i = tf.range(nd)[:,None]
    j = tf.range(ns)
    m = i >= j - ns + nd
    # to ignore first parts of context (useful for sampling with static shapes)
    # m = tf.math.logical_and(m, tf.math.logical_or(j  >= ignore, i < ignore - ns + nd))
    return tf.cast(m, dtype) 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def test_attention_mask():
    with tf.Session() as sess:
        for nd in 1, 2, 3:
            for ns in range(nd, 4):
                ours = model.attention_mask(nd, ns, dtype=tf.int32)
                theirs = tf.matrix_band_part(tf.ones([nd, ns], dtype=tf.int32), tf.cast(-1, tf.int32), ns-nd)
                ours, theirs = sess.run([ours, theirs])
                print(ours)
                print(theirs)
                assert np.all(ours == theirs) 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def _create_mask(qlen, mlen, dtype=tf.float32, same_length=False):
  """create causal attention mask."""
  attn_mask = tf.ones([qlen, qlen], dtype=dtype)
  mask_u = tf.matrix_band_part(attn_mask, 0, -1)
  mask_dia = tf.matrix_band_part(attn_mask, 0, 0)
  attn_mask_pad = tf.zeros([qlen, mlen], dtype=dtype)
  ret = tf.concat([attn_mask_pad, mask_u - mask_dia], 1)
  if same_length:
    mask_l = tf.matrix_band_part(attn_mask, -1, 0)
    ret = tf.concat([ret[:, :qlen] + mask_l - mask_dia, ret[:, qlen:]], 1)

  return ret 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def ones_matrix_band_part(rows, cols, num_lower, num_upper, out_shape=None):
  """Matrix band part of ones.

  Args:
    rows: int determining number of rows in output
    cols: int
    num_lower: int, maximum distance backward. Negative values indicate
      unlimited.
    num_upper: int, maximum distance forward. Negative values indicate
      unlimited.
    out_shape: shape to reshape output by.

  Returns:
    Tensor of size rows * cols reshaped into shape out_shape.
  """
  if all([isinstance(el, int) for el in [rows, cols, num_lower, num_upper]]):
    # Needed info is constant, so we construct in numpy
    if num_lower < 0:
      num_lower = rows - 1
    if num_upper < 0:
      num_upper = cols - 1
    lower_mask = np.tri(cols, rows, num_lower).T
    upper_mask = np.tri(rows, cols, num_upper)
    band = np.ones((rows, cols)) * lower_mask * upper_mask
    if out_shape:
      band = band.reshape(out_shape)
    band = tf.constant(band, tf.float32)
  else:
    band = tf.matrix_band_part(
        tf.ones([rows, cols]), tf.cast(num_lower, tf.int64),
        tf.cast(num_upper, tf.int64))
    if out_shape:
      band = tf.reshape(band, out_shape)

  return band 

# 需要导入模块: import tensorflow [as 别名]
# 或者: from tensorflow import matrix_band_part [as 别名]
def attention_mask(nd, ns, dtype):
	"""1's in the lower triangle, counting from the lower right corner.

	Same as tf.matrix_band_part(tf.ones([nd, ns]), -1, ns-nd), but doesn't produce garbage on TPUs.
	"""
	i = tf.range(nd)[:,None]
	j = tf.range(ns)
	m = i >= j - ns + nd
	return tf.cast(m, dtype)