# Copyright (c) 2020 zfit
import numpy as np
import tensorflow as tf
SWITCH_ON = True
[docs]def is_tensor(x):
return tf.is_tensor(x)
[docs]def allclose_anyaware(x, y, rtol=1e-5, atol=1e-8):
"""Tests if x and y are close by first testing equality (with numpy), then within the limits.
The prepended equality test allow for ANY objects to compare positively if the x and y have the shape (1, n)
with n arbitrary
Args:
x:
y:
rtol:
atol:
Returns:
"""
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.reduce_all(tf.less_equal(tf.abs(x - y), tf.abs(y) * rtol + atol))
else:
x = np.array(x)
y = np.array(y)
if any(ar.dtype == object for ar in (x, y)):
from zfit.core.space import LimitRangeDefinition
equal = []
for x1, y1 in zip(x[0], y[0]):
if isinstance(x1, LimitRangeDefinition) or isinstance(y1, LimitRangeDefinition):
equal.append(x1 < y1 or x1 > y1)
else:
equal.append(np.allclose(x1, y1, rtol=rtol, atol=atol))
allclose = np.array(equal)[None, :]
else:
allclose = np.allclose(x, y, rtol=rtol, atol=atol)
return allclose
[docs]def broadcast_to(input, shape):
if not SWITCH_ON or is_tensor(input):
return tf.broadcast_to(input, shape)
else:
return np.broadcast_to(input, shape)
[docs]def expand_dims(input, axis):
if not SWITCH_ON or is_tensor(input):
return tf.expand_dims(input, axis)
else:
return np.expand_dims(input, axis)
[docs]def reduce_prod(input_tensor, axis=None, keepdims=None):
if not SWITCH_ON or is_tensor(input_tensor):
return tf.reduce_prod(input_tensor, axis, keepdims=keepdims)
else:
if keepdims is None:
return np.prod(input_tensor, axis)
else:
return np.prod(input_tensor, axis, keepdims=keepdims)
[docs]def equal(x, y):
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.equal(x, y)
else:
return np.equal(x, y)
[docs]def reduce_all(input_tensor, axis=None):
if not SWITCH_ON or is_tensor(input_tensor):
return tf.reduce_all(input_tensor, axis)
else:
return np.all(input_tensor, axis)
[docs]def reduce_any(input_tensor, axis=None):
if not SWITCH_ON or is_tensor(input_tensor):
return tf.reduce_any(input_tensor, axis)
else:
return np.any(input_tensor, axis)
[docs]def logical_and(x, y):
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.logical_and(x, y)
else:
return np.logical_and(x, y)
[docs]def logical_or(x, y):
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.logical_or(x, y)
else:
return np.logical_or(x, y)
[docs]def less_equal(x, y):
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.less_equal(x, y)
else:
return np.less_equal(x, y)
[docs]def greater_equal(x, y):
if not SWITCH_ON or is_tensor(x) or is_tensor(y):
return tf.greater_equal(x, y)
else:
return np.greater_equal(x, y)
[docs]def gather(x, indices=None, axis=None):
if not SWITCH_ON or is_tensor(x):
return tf.gather(x, indices=indices, axis=axis)
else:
return np.take(x, indices=indices, axis=axis)
[docs]def concat(values, axis, name=None):
if not SWITCH_ON or is_tensor(values) or any(is_tensor(value) for value in values):
return tf.concat(values=values, axis=axis, name=name)
else:
return np.concatenate(values, axis=axis)
def _try_convert_numpy(tensorlike):
if hasattr(tensorlike, 'numpy'):
tensorlike = tensorlike.numpy()
if not isinstance(tensorlike, np.ndarray):
from zfit.util.exception import CannotConvertToNumpyError
raise CannotConvertToNumpyError(f"Cannot convert {tensorlike} to a Numpy array. This may be because the"
f" object is a Tensor and the function is called in Graph mode (e.g. in"
f"a `z.function` decorated function.\n"
f"If this error appears and is not understandable, it is most likely a bug."
f" Please open an issue on Github.")
return tensorlike
zfit