Skip to content
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Go to file
Cannot retrieve contributors at this time
import numbers
import numpy as np
import theano
import theano.tensor as T
#: Tuple of ``int``-like types for ``isinstance`` checks.
#: Specifically includes long integers and numpy integers.
int_types = (numbers.Integral, np.integer)
def floatX(arr):
"""Converts data to a numpy array of dtype ``theano.config.floatX``.
arr : array_like
The data to be converted.
numpy ndarray
The input array in the ``floatX`` dtype configured for Theano.
If `arr` is an ndarray of correct dtype, it is returned as is.
return np.asarray(arr, dtype=theano.config.floatX)
def shared_empty(dim=2, dtype=None):
"""Creates empty Theano shared variable.
Shortcut to create an empty Theano shared variable with
the specified number of dimensions.
dim : int, optional
The number of dimensions for the empty variable, defaults to 2.
dtype : a numpy data-type, optional
The desired dtype for the variable. Defaults to the Theano
``floatX`` dtype.
Theano shared variable
An empty Theano shared variable of dtype ``dtype`` with
`dim` dimensions.
if dtype is None:
dtype = theano.config.floatX
shp = tuple([1] * dim)
return theano.shared(np.zeros(shp, dtype=dtype))
def as_theano_expression(input):
"""Wrap as Theano expression.
Wraps the given input as a Theano constant if it is not
a valid Theano expression already. Useful to transparently
handle numpy arrays and Python scalars, for example.
input : number, numpy array or Theano expression
Expression to be converted to a Theano constant.
Theano symbolic constant
Theano constant version of `input`.
if isinstance(input, theano.gof.Variable):
return input
return theano.tensor.constant(input)
except Exception as e:
raise TypeError("Input of type %s is not a Theano expression and "
"cannot be wrapped as a Theano constant (original "
"exception: %s)" % (type(input), e))
def collect_shared_vars(expressions):
"""Returns all shared variables the given expression(s) depend on.
expressions : Theano expression or iterable of Theano expressions
The expressions to collect shared variables from.
list of Theano shared variables
All shared variables the given expression(s) depend on, in fixed order
(as found by a left-recursive depth-first search). If some expressions
are shared variables themselves, they are included in the result.
# wrap single expression in list
if isinstance(expressions, theano.Variable):
expressions = [expressions]
# return list of all shared variables
return [v for v in theano.gof.graph.inputs(reversed(expressions))
if isinstance(v, theano.compile.SharedVariable)]
def one_hot(x, m=None):
"""One-hot representation of integer vector.
Given a vector of integers from 0 to m-1, returns a matrix
with a one-hot representation, where each row corresponds
to an element of x.
x : integer vector
The integer vector to convert to a one-hot representation.
m : int, optional
The number of different columns for the one-hot representation. This
needs to be strictly greater than the maximum value of `x`.
Defaults to ``max(x) + 1``.
Theano tensor variable
A Theano tensor variable of shape (``n``, `m`), where ``n`` is the
length of `x`, with the one-hot representation of `x`.
If your integer vector represents target class memberships, and you wish to
compute the cross-entropy between predictions and the target class
memberships, then there is no need to use this function, since the function
:func:`lasagne.objectives.categorical_crossentropy()` can compute the
cross-entropy from the integer vector directly.
if m is None:
m = T.cast(T.max(x) + 1, 'int32')
return T.eye(m)[T.cast(x, 'int32')]
def unique(l):
"""Filters duplicates of iterable.
Create a new list from l with duplicate entries removed,
while preserving the original order.
l : iterable
Input iterable to filter of duplicates.
A list of elements of `l` without duplicates and in the same order.
new_list = []
seen = set()
for el in l:
if el not in seen:
return new_list
def as_tuple(x, N, t=None):
Coerce a value to a tuple of given length (and possibly given type).
x : value or iterable
N : integer
length of the desired tuple
t : type or tuple of type, optional
required type or types for all elements
``tuple(x)`` if `x` is iterable, ``(x,) * N`` otherwise.
if `type` is given and `x` or any of its elements do not match it
if `x` is iterable, but does not have exactly `N` elements
X = tuple(x)
except TypeError:
X = (x,) * N
if (t is not None) and not all(isinstance(v, t) for v in X):
if t == int_types:
expected_type = "int" # easier to understand
elif isinstance(t, tuple):
expected_type = " or ".join(tt.__name__ for tt in t)
expected_type = t.__name__
raise TypeError("expected a single value or an iterable "
"of {0}, got {1} instead".format(expected_type, x))
if len(X) != N:
raise ValueError("expected a single value or an iterable "
"with length {0}, got {1} instead".format(N, x))
return X
def inspect_kwargs(func):
Inspects a callable and returns a list of all optional keyword arguments.
func : callable
The callable to inspect
kwargs : list of str
Names of all arguments of `func` that have a default value, in order
# We try the Python 3.x way first, then fall back to the Python 2.x way
from inspect import signature
except ImportError: # pragma: no cover
from inspect import getargspec
spec = getargspec(func)
return spec.args[-len(spec.defaults):] if spec.defaults else []
else: # pragma: no cover
params = signature(func).parameters
return [ for p in params.values() if p.default is not p.empty]
def compute_norms(array, norm_axes=None):
""" Compute incoming weight vector norms.
array : numpy array or Theano expression
Weight or bias.
norm_axes : sequence (list or tuple)
The axes over which to compute the norm. This overrides the
default norm axes defined for the number of dimensions
in `array`. When this is not specified and `array` is a 2D array,
this is set to `(0,)`. If `array` is a 3D, 4D or 5D array, it is
set to a tuple listing all axes but axis 0. The former default is
useful for working with dense layers, the latter is useful for 1D,
2D and 3D convolutional layers.
Finally, in case `array` is a vector, `norm_axes` is set to an empty
tuple, and this function will simply return the absolute value for
each element. This is useful when the function is applied to all
parameters of the network, including the bias, without distinction.
norms : 1D array or Theano vector (1D)
1D array or Theano vector of incoming weight/bias vector norms.
>>> array = np.random.randn(100, 200)
>>> norms = compute_norms(array)
>>> norms.shape
>>> norms = compute_norms(array, norm_axes=(1,))
>>> norms.shape
# Check if supported type
if not isinstance(array, theano.Variable) and \
not isinstance(array, np.ndarray):
raise RuntimeError(
"Unsupported type {}. "
"Only theano variables and numpy arrays "
"are supported".format(type(array))
# Compute default axes to sum over
ndim = array.ndim
if norm_axes is not None:
sum_over = tuple(norm_axes)
elif ndim == 1: # For Biases that are in 1d (e.g. b of DenseLayer)
sum_over = ()
elif ndim == 2: # DenseLayer
sum_over = (0,)
elif ndim in [3, 4, 5]: # Conv{1,2,3}DLayer
sum_over = tuple(range(1, ndim))
raise ValueError(
"Unsupported tensor dimensionality {}. "
"Must specify `norm_axes`".format(array.ndim)
# Run numpy or Theano norm computation
if isinstance(array, theano.Variable):
# Apply theano version if it is a theano variable
if len(sum_over) == 0:
norms = T.abs_(array) # abs if we have nothing to sum over
norms = T.sqrt(T.sum(array**2, axis=sum_over))
elif isinstance(array, np.ndarray):
# Apply the numpy version if ndarray
if len(sum_over) == 0:
norms = abs(array) # abs if we have nothing to sum over
norms = np.sqrt(np.sum(array**2, axis=sum_over))
return norms
def create_param(spec, shape, name=None):
Helper method to create Theano shared variables for layer parameters
and to initialize them.
spec : scalar number, numpy array, Theano expression, or callable
Either of the following:
* a scalar or a numpy array with the initial parameter values
* a Theano expression or shared variable representing the parameters
* a function or callable that takes the desired shape of
the parameter array as its single argument and returns
a numpy array, a Theano expression, or a shared variable
representing the parameters.
shape : iterable of int
a tuple or other iterable of integers representing the desired
shape of the parameter array.
name : string, optional
The name to give to the parameter variable. Ignored if `spec`
is or returns a Theano expression or shared variable that
already has a name.
Theano shared variable or Theano expression
A Theano shared variable or expression representing layer parameters.
If a scalar or a numpy array was provided, a shared variable is
initialized to contain this array. If a shared variable or expression
was provided, it is simply returned. If a callable was provided, it is
called, and its output is used to initialize a shared variable.
This function is called by :meth:`Layer.add_param()` in the constructor
of most :class:`Layer` subclasses. This enables those layers to
support initialization with scalars, numpy arrays, existing Theano shared
variables or expressions, and callables for generating initial parameter
values, Theano expressions, or shared variables.
import numbers # to check if argument is a number
shape = tuple(shape) # convert to tuple if needed
if any(d <= 0 for d in shape):
raise ValueError((
"Cannot create param with a non-positive shape dimension. "
"Tried to create param with shape=%r, name=%r") % (shape, name))
err_prefix = "cannot initialize parameter %s: " % name
if callable(spec):
spec = spec(shape)
err_prefix += "the %s returned by the provided callable"
err_prefix += "the provided %s"
if isinstance(spec, numbers.Number) or isinstance(spec, np.generic) \
and spec.dtype.kind in 'biufc':
spec = np.asarray(spec)
if isinstance(spec, np.ndarray):
if spec.shape != shape:
raise ValueError("%s has shape %s, should be %s" %
(err_prefix % "numpy array", spec.shape, shape))
# We assume parameter variables do not change shape after creation.
# We can thus fix their broadcast pattern, to allow Theano to infer
# broadcastable dimensions of expressions involving these parameters.
bcast = tuple(s == 1 for s in shape)
spec = theano.shared(spec, broadcastable=bcast)
if isinstance(spec, theano.Variable):
# We cannot check the shape here, Theano expressions (even shared
# variables) do not have a fixed compile-time shape. We can check the
# dimensionality though.
if spec.ndim != len(shape):
raise ValueError("%s has %d dimensions, should be %d" %
(err_prefix % "Theano variable", spec.ndim,
# We only assign a name if the user hasn't done so already.
if not = name
return spec
if "callable" in err_prefix:
raise TypeError("%s is not a numpy array or a Theano expression" %
(err_prefix % "value"))
raise TypeError("%s is not a numpy array, a Theano expression, "
"or a callable" % (err_prefix % "spec"))
def unroll_scan(fn, sequences, outputs_info, non_sequences, n_steps,
Helper function to unroll for loops. Can be used to unroll theano.scan.
The parameter names are identical to theano.scan, please refer to here
for more information.
Note that this function does not support the truncate_gradient
setting from theano.scan.
fn : function
Function that defines calculations at each step.
sequences : TensorVariable or list of TensorVariables
List of TensorVariable with sequence data. The function iterates
over the first dimension of each TensorVariable.
outputs_info : list of TensorVariables
List of tensors specifying the initial values for each recurrent
non_sequences: list of TensorVariables
List of theano.shared variables that are used in the step function.
n_steps: int
Number of steps to unroll.
go_backwards: bool
If true the recursion starts at sequences[-1] and iterates
List of TensorVariables. Each element in the list gives the recurrent
values at each time step.
if not isinstance(sequences, (list, tuple)):
sequences = [sequences]
# When backwards reverse the recursion direction
counter = range(n_steps)
if go_backwards:
counter = counter[::-1]
output = []
prev_vals = outputs_info
for i in counter:
step_input = [s[i] for s in sequences] + prev_vals + non_sequences
out_ = fn(*step_input)
# The returned values from step can be either a TensorVariable,
# a list, or a tuple. Below, we force it to always be a list.
if isinstance(out_, T.TensorVariable):
out_ = [out_]
if isinstance(out_, tuple):
out_ = list(out_)
prev_vals = output[-1]
# iterate over each scan output and convert it to same format as scan:
# [[output11, output12,...output1n],
# [output21, output22,...output2n],...]
output_scan = []
for i in range(len(output[0])):
l = map(lambda x: x[i], output)
return output_scan