def _complexity_cost(weights, mean, sigma, prior_density):
k = -0.5 * math.log(2 * math.pi)
cost = k - gen_math_ops.log(sigma) - 0.5 * gen_math_ops.square(
(weights - mean) / sigma)
cost -= gen_math_ops.log(prior_density(weights))
cost = math_ops.reduce_sum(cost)
return cost
def GraphFn(self, x1, x2):
x = x1
q = math_ops.abs(x)
q = q + 1.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = array_ops.squeeze(q, axis=-2)
q = math_ops.abs(q)
q = q + 2.2
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = array_ops.squeeze(q, axis=3)
q = math_ops.abs(q)
q = q + 3.0
a = gen_math_ops.reciprocal(q)
# this chain of operations has a batch size of 5, which is different from
# the batch size for the other operations.
x = constant_op.constant(np.random.randn(5, 8, 12), dtype=x.dtype)
q = math_ops.abs(x)
q = q + 2.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 2.1
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 4.0
b = gen_math_ops.reciprocal(q)
# TODO(jie): this one will break, broadcasting on batch.
x = x2
q = math_ops.abs(x)
q = q + 5.0
q = gen_math_ops.exp(q)
q = array_ops.squeeze(q, axis=[-1, -2, 3])
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 5.1
q = gen_array_ops.reshape(q, [12, 5, 1, 1, 8, 1, 12])
q = array_ops.squeeze(q, axis=[5, 2, 3])
q = gen_math_ops.sqrt(q)
q = math_ops.abs(q)
q = q + 5.2
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 5.3
c = gen_math_ops.reciprocal(q)
q = a * b
q = q / c
return array_ops.squeeze(q, name="output_0")
def _dkwm_cdf_envelope(n, error_rate, name=None):
"""Computes the CDF envelope that the DKWM inequality licenses.
The [Dvoretzky-Kiefer-Wolfowitz-Massart inequality]
(https://en.wikipedia.org/wiki/CDF-based_nonparametric_confidence_interval)
gives a stochastic bound on the distance between the true cumulative
distribution function (CDF) of any distribution and its empirical
CDF. To wit, for `n` iid samples from any distribution with CDF F,
```none
P(sup_x |F_n(x) - F(x)| > eps) < 2exp(-2n eps^2)
```
This function computes the envelope size `eps` as a function of the
number of samples `n` and the desired limit on the left-hand
probability above.
Args:
n: Tensor of numbers of samples drawn.
error_rate: Floating-point tensor of admissible rates of mistakes.
name: A name for this operation (optional).
Returns:
eps: Tensor of maximum distances the true CDF can be from the
empirical CDF. This scales as `O(sqrt(-log(error_rate)))` and
as `O(1 / sqrt(n))`. The shape is the broadcast of `n` and
`error_rate`.
"""
with ops.name_scope(name, "dkwm_cdf_envelope", [n, error_rate]):
n = math_ops.cast(n, dtype=error_rate.dtype)
return math_ops.sqrt(-gen_math_ops.log(error_rate / 2.) / (2. * n))
def _dkwm_cdf_envelope(n, error_rate, name=None):
"""Computes the CDF envelope that the DKWM inequality licenses.
The [Dvoretzky-Kiefer-Wolfowitz-Massart inequality]
(https://en.wikipedia.org/wiki/CDF-based_nonparametric_confidence_interval)
gives a stochastic bound on the distance between the true cumulative
distribution function (CDF) of any distribution and its empirical
CDF. To wit, for `n` iid samples from any distribution with CDF F,
```none
P(sup_x |F_n(x) - F(x)| > eps) < 2exp(-2n eps^2)
```
This function computes the envelope size `eps` as a function of the
number of samples `n` and the desired limit on the left-hand
probability above.
Args:
n: Tensor of numbers of samples drawn.
error_rate: Floating-point tensor of admissible rates of mistakes.
name: A name for this operation (optional).
Returns:
eps: Tensor of maximum distances the true CDF can be from the
empirical CDF. This scales as `O(sqrt(-log(error_rate)))` and
as `O(1 / sqrt(n))`. The shape is the broadcast of `n` and
`error_rate`.
"""
with ops.name_scope(name, "dkwm_cdf_envelope", [n, error_rate]):
n = math_ops.cast(n, dtype=error_rate.dtype)
return math_ops.sqrt(-gen_math_ops.log(error_rate / 2.) / (2. * n))
def fucking_deep_gaze_logsumexp(input_tensor,axis=None, keepdims=False,
name=None):
"""
Adaptd from
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/math_ops.py.
It is the same as the classic logsumexp instead you substact log(N) where N
in the number of tensor over which compute the logsumexp (if you have 10
readout nets, N=10). I don't know why they do this.
"""
keepdims = False if keepdims is None else keepdims
input_tensor = ops.convert_to_tensor(input_tensor)
with ops.name_scope(name, "ReduceLogSumExp", [input_tensor]) as name:
raw_max = tf.reduce_max(input_tensor, axis=axis, keep_dims=True)
my_max = array_ops.stop_gradient( array_ops.where(
gen_math_ops.is_finite(raw_max), raw_max,
array_ops.zeros_like(raw_max)))
result = gen_math_ops.log(
#reduce_sum( # normal logsumexp
tf.reduce_mean( # fuckimg modif from deep_gaze for the output only
gen_math_ops.exp(tf.subtract(input_tensor, my_max)),
axis, keep_dims=keepdims))
if not keepdims:
my_max = array_ops.reshape(my_max, array_ops.shape(result))
result = gen_math_ops.add(result, my_max)
return result
def signal_to_noise(y_true, y_pred, mode='snr', data_format=None, epsilon=1e-8):
'''Signal-to-noise ratio. (metric)
Calculate the signal-to-noise ratio. It support different modes.
Arguments:
mode: (1) snr: mean [ y_true^2 / (y_pred - y_true)^2 ]
(2) psnr: mean [ max( y_true^2 ) / (y_pred - y_true)^2 ]
data_format: 'channels_first' or 'channels_last'. The default setting is generally
'channels_last' like other tf.keras APIs.
epsilon: used for avoid zero division.
Input:
y_true: label, tensor in any shape.
y_pred: prediction, tensor in any shape.
Output:
scalar, the mean SNR.
'''
get_reduced_axes = get_channels(y_true, data_format)
if mode.casefold() == 'psnr':
signal = math_ops.reduce_max(gen_math_ops.square(y_true), axis=get_reduced_axes)
else:
signal = math_ops.reduce_sum(gen_math_ops.square(y_true), axis=get_reduced_axes)
noise = math_ops.reduce_sum(gen_math_ops.square(y_true - y_pred), axis=get_reduced_axes) + epsilon
coeff = (10.0/2.3025851) # 10/log_e(10)
return coeff*math_ops.reduce_mean(gen_math_ops.log(math_ops.divide(signal, noise)))
def GetParams(self):
"""Test for unary operations in TF-TRT."""
dtype = dtypes.float32
input_name = "input"
input_dims = [12, 5, 8, 1, 1, 12]
output_name = "output"
input2_name = "input_2"
input2_dims = [12, 5, 8, 1, 12, 1, 1]
g = ops.Graph()
with g.as_default():
x = array_ops.placeholder(dtype=dtype, shape=input_dims, name=input_name)
q = math_ops.abs(x)
q = q + 1.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = array_ops.squeeze(q, axis=-2)
q = math_ops.abs(q)
q = q + 2.2
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = array_ops.squeeze(q, axis=3)
q = math_ops.abs(q)
q = q + 3.0
a = gen_math_ops.reciprocal(q)
x = constant_op.constant(np.random.randn(5, 8, 12), dtype=dtype)
q = math_ops.abs(x)
q = q + 2.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 2.1
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 4.0
b = gen_math_ops.reciprocal(q)
# TODO(jie): this one will break, broadcasting on batch.
x = array_ops.placeholder(
dtype=dtype, shape=input2_dims, name=input2_name)
q = math_ops.abs(x)
q = q + 5.0
q = gen_math_ops.exp(q)
q = array_ops.squeeze(q, axis=[-1, -2, 3])
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 5.1
q = gen_array_ops.reshape(q, [12, 5, 1, 1, 8, 1, 12])
q = array_ops.squeeze(q, axis=[5, 2, 3])
q = gen_math_ops.sqrt(q)
q = math_ops.abs(q)
q = q + 5.2
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 5.3
c = gen_math_ops.reciprocal(q)
q = a * b
q = q / c
array_ops.squeeze(q, name=output_name)
return trt_test.TfTrtIntegrationTestParams(
gdef=g.as_graph_def(),
input_names=[input_name, input2_name],
input_dims=[input_dims, input2_dims],
output_names=[output_name],
expected_output_dims=[(12, 5, 8, 12)])
def GetParams(self):
"""Test for unary operations in TF-TRT."""
dtype = dtypes.float32
input_name = "input"
input_dims = [12, 5, 8, 1, 1, 12]
input2_name = "input_2"
input2_dims = [12, 5, 8, 1, 12, 1, 1]
g = ops.Graph()
with g.as_default():
x = array_ops.placeholder(dtype=dtype, shape=input_dims, name=input_name)
q = math_ops.abs(x)
q = q + 1.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = array_ops.squeeze(q, axis=-2)
q = math_ops.abs(q)
q = q + 2.2
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = array_ops.squeeze(q, axis=3)
q = math_ops.abs(q)
q = q + 3.0
a = gen_math_ops.reciprocal(q)
x = constant_op.constant(np.random.randn(5, 8, 12), dtype=dtype)
q = math_ops.abs(x)
q = q + 2.0
q = gen_math_ops.exp(q)
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 2.1
q = gen_math_ops.sqrt(q)
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 4.0
b = gen_math_ops.reciprocal(q)
# TODO(jie): this one will break, broadcasting on batch.
x = array_ops.placeholder(
dtype=dtype, shape=input2_dims, name=input2_name)
q = math_ops.abs(x)
q = q + 5.0
q = gen_math_ops.exp(q)
q = array_ops.squeeze(q, axis=[-1, -2, 3])
q = gen_math_ops.log(q)
q = math_ops.abs(q)
q = q + 5.1
q = gen_array_ops.reshape(q, [12, 5, 1, 1, 8, 1, 12])
q = array_ops.squeeze(q, axis=[5, 2, 3])
q = gen_math_ops.sqrt(q)
q = math_ops.abs(q)
q = q + 5.2
q = gen_math_ops.rsqrt(q)
q = math_ops.negative(q)
q = math_ops.abs(q)
q = q + 5.3
c = gen_math_ops.reciprocal(q)
q = a * b
q = q / c
array_ops.squeeze(q, name=self.output_name)
return trt_test.TfTrtIntegrationTestParams(
gdef=g.as_graph_def(),
input_names=[input_name, input2_name],
input_dims=[input_dims, input2_dims],
num_expected_engines=5,
expected_output_dims=(12, 5, 8, 12),
allclose_atol=1.e-03,
allclose_rtol=1.e-03)
