import warnings
import numpy as np
import orcanet.misc as misc
# for loading via toml
lmods, register = misc.get_register()
class ColumnLabels:
"""
Label of each model output is column with the same name in the h5 file.
This is the default label modifier.
Example
-------
Model has output "energy" --> label is column "energy" from the label
dataset in the h5 file.
Parameters
----------
model : ks.Model
A keras model.
"""
def __init__(self, model):
self.output_names = model.output_names
def __call__(self, info_blob):
ys = {name: info_blob["y_values"][name] for name in self.output_names}
return ys
"""
Some basic sample modifiers to use with orcanet.
Use them by setting .cfg.sample_modifier of the orcanet.core.Organizer.
"""
from abc import abstractmethod
import warnings
import numpy as np
from orcanet.misc import get_register
import tensorflow as tf
# for loading via toml
smods, register = get_register()
class PerInputModifier:
"""
For modifiers that do the same operation on each input.
Apply modify on x_value of each input, and output as dict.
"""
def __call__(self, info_blob):
x_values = info_blob["x_values"]
xs = dict()
for key, x_value in x_values.items():
xs[key] = self.modify(x_value)
return xs
@abstractmethod
def modify(self, x_value):
""" x_value is a batch of input data as a numpy array. """
def test_register(self):
saved, register = misc.get_register()
register(_func)
register(_Cls)
self.assertDictEqual(self.register, saved)
import tensorflow as tf
import tensorflow.keras.backend as K
import tensorflow.keras as ks
import tensorflow.keras.layers as layers
import medgeconv
from orcanet.misc import get_register
# for loading via toml and orcanet custom objects
blocks, register = get_register()
# edge conv blocks
register(medgeconv.DisjointEdgeConvBlock)
@register
class ConvBlock:
"""
1D/2D/3D Convolutional block followed by BatchNorm, Activation,
MaxPooling and/or Dropout.
Parameters
----------
conv_dim : int
Specifies the dimension of the convolutional block, 1D/2D/3D.
filters : int
Number of filters used for the convolutional layer.
strides : int or tuple
The stride length of the convolution.
padding : str or int or list
If str: Padding of the conv block.
If int or list: Padding argument of a ZeroPaddingND layer that
gets added before the convolution.
"""
OrcaNet custom loss functions.
"""
import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp
from orcanet.misc import get_register
# fuzz factor for numerical stability
EPS = tf.constant(1e-7, dtype="float32")
# for loading via toml and orcanet custom objects
loss_functions, _register = get_register()
@_register
def lkl_normal_tfp(y_true, y_pred):
""" Normal distribution using tfp. See lkl_normal. """
mu_true = y_true[:, 0]
mu_pred, sigma_pred = y_pred[:, 0], y_pred[:, 1]
return -1 * tfp.distributions.Normal(
loc=mu_pred,
scale=tf.math.maximum(sigma_pred, EPS),
).log_prob(mu_true)
@_register
def lkl_normal(y_true, y_pred):
"""
Negative normal log-likelihood function for n regression output neurons
with clipping for increased stability.